fix(router): fix ContentPolicyViolationError policy shadowing and partial-policy zero-threshold

Two bugs from Greptile review on PR #34416:

- ContentPolicyViolationError subclasses BadRequestError, so listing
  BadRequestError first in _EXCEPTION_POLICY_FIELDS made the isinstance
  check always match BadRequestError for content-policy errors, using the
  wrong allowed_fails threshold. Reordered so the subclass is checked first.

- A deployment with a partial allowed_fails_policy and no deployment-wide
  allowed_fails forced allowed_fails_override=0 for any exception type its
  policy didn't cover, cooling the deployment down on the first unrelated
  failure. Now defers to router-level behavior for uncovered exception
  types instead of forcing an immediate cooldown.

litellm/router_utils/cooldown_handlers.py

@@ -38,11 +38,12 @@ else:
 
 
 _EXCEPTION_POLICY_FIELDS: tuple[tuple[type, str], ...] = (
+    # ContentPolicyViolationError subclasses BadRequestError, so it must be checked first.
+    (litellm.ContentPolicyViolationError, "ContentPolicyViolationErrorAllowedFails"),
     (litellm.BadRequestError, "BadRequestErrorAllowedFails"),
     (litellm.AuthenticationError, "AuthenticationErrorAllowedFails"),
     (litellm.Timeout, "TimeoutErrorAllowedFails"),
     (litellm.RateLimitError, "RateLimitErrorAllowedFails"),
-    (litellm.ContentPolicyViolationError, "ContentPolicyViolationErrorAllowedFails"),
     (litellm.InternalServerError, "InternalServerErrorAllowedFails"),
     (litellm.ServiceUnavailableError, "ServiceUnavailableErrorAllowedFails"),
     (litellm.BadGatewayError, "BadGatewayErrorAllowedFails"),
@@ -103,14 +104,22 @@ def _should_cooldown_based_on_deployment_policy(
     dep_policy: dict[str, int] | None,
     dep_allowed_fails: int | None,
 ) -> bool:
-    """Resolve deployment-level allowed-fails and delegate to the shared counting logic."""
+    """Resolve deployment-level allowed-fails and delegate to the shared counting logic.
+
+    When the deployment's policy doesn't cover *original_exception*'s type and no
+    deployment-wide `allowed_fails` is set either, defer to router-level behavior
+    instead of forcing an immediate cooldown.
+    """
     allowed_fails_from_policy = _resolve_allowed_fails_from_policy(dep_policy, original_exception)
     if allowed_fails_from_policy is not None:
-        allowed_fails_override: int = allowed_fails_from_policy
+        allowed_fails_override: int | None = allowed_fails_from_policy
-        cache_key_suffix: str = type(original_exception).__name__
+        cache_key_suffix: str | None = type(original_exception).__name__
-    else:
+    elif dep_allowed_fails is not None:
-        allowed_fails_override = dep_allowed_fails if dep_allowed_fails is not None else 0
+        allowed_fails_override = dep_allowed_fails
         cache_key_suffix = "generic"
+    else:
+        allowed_fails_override = None
+        cache_key_suffix = None
 
     dep = litellm_router_instance.get_model_info(id=deployment)
     cooldown_time_override: float | None = None

tests/test_litellm/router_utils/test_cooldown_handlers.py

@@ -135,7 +135,11 @@ class TestShouldCooldownBasedOnDeploymentPolicy:
         assert call_kwargs["allowed_fails_override"] == 3
         assert call_kwargs["cache_key_suffix"] == "generic"
 
-    def test_no_policy_and_no_dep_allowed_fails_defaults_to_zero(self):
+    def test_no_policy_and_no_dep_allowed_fails_defers_to_router_level(self):
+        """When neither a deployment policy nor a deployment-wide allowed_fails covers
+        this exception, defer to router-level behavior instead of forcing an
+        immediate cooldown (allowed_fails_override=0 would trip on the first failure
+        of any exception type the deployment's config doesn't mention)."""
         exc = litellm.InternalServerError("500", "openai", "gpt-4")
         router = self._make_router({"litellm_params": {}, "model_info": {}})
 
@@ -146,7 +150,26 @@ class TestShouldCooldownBasedOnDeploymentPolicy:
             _should_cooldown_based_on_deployment_policy(router, "dep-1", exc, None, None)
 
         call_kwargs = mock_sc.call_args[1]
-        assert call_kwargs["allowed_fails_override"] == 0
+        assert call_kwargs["allowed_fails_override"] is None
+        assert call_kwargs["cache_key_suffix"] is None
+
+    def test_partial_policy_without_dep_allowed_fails_defers_for_uncovered_exception(self):
+        """A deployment that only sets RateLimitErrorAllowedFails must not force a
+        zero-fail threshold on an unrelated TimeoutError; it should defer to
+        router-level behavior for exception types its policy doesn't mention."""
+        policy = {"RateLimitErrorAllowedFails": 0}
+        exc = litellm.Timeout("timed out", "openai", "gpt-4")
+        router = self._make_router({"litellm_params": {}, "model_info": {}})
+
+        with patch(
+            "litellm.router_utils.cooldown_handlers.should_cooldown_based_on_allowed_fails_policy"
+        ) as mock_sc:
+            mock_sc.return_value = False
+            _should_cooldown_based_on_deployment_policy(router, "dep-1", exc, policy, dep_allowed_fails=None)
+
+        call_kwargs = mock_sc.call_args[1]
+        assert call_kwargs["allowed_fails_override"] is None
+        assert call_kwargs["cache_key_suffix"] is None
 
     def test_cooldown_time_from_litellm_params_passed_through(self):
         exc = litellm.RateLimitError("429", "openai", "gpt-4")