fix(bedrock): drop client_metadata for ARNs that hide the model family

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

litellm/llms/bedrock/chat/converse_transformation.py

@@ -86,6 +86,7 @@ from litellm.utils import (
 from ..common_utils import (
     BedrockError,
     BedrockModelInfo,
+    bedrock_arn_hides_model_family,
     bedrock_converse_supports_parallel_tool_use_config,
     get_anthropic_beta_from_headers,
     get_bedrock_tool_name,
@@ -1325,7 +1326,8 @@ class AmazonConverseConfig(BaseConfig):
 
         additional_request_params.pop("parallel_tool_calls", None)
 
-        if base_model.startswith("anthropic") and additional_request_params.pop("client_metadata", None) is not None:
+        drops_client_metadata: Final = base_model.startswith("anthropic") or bedrock_arn_hides_model_family(model)
+        if drops_client_metadata and additional_request_params.pop("client_metadata", None) is not None:
             litellm.verbose_logger.debug(
                 "Bedrock Converse: dropping `client_metadata` for model=%s, Anthropic rejects it with "
                 "'client_metadata: Extra inputs are not permitted'",

litellm/llms/bedrock/common_utils.py

@@ -720,6 +720,15 @@ def get_bedrock_base_model(model: str) -> str:
     return model
 
 
+def bedrock_arn_hides_model_family(model: str) -> bool:
+    """
+    True for an ARN-addressed model whose base name carries no ``provider.model``
+    id, such as an application inference profile or a provisioned throughput ARN.
+    Callers that gate behavior on the model family cannot resolve one here.
+    """
+    return "arn:" in model.lower() and "." not in get_bedrock_base_model(model)
+
+
 def bedrock_converse_supports_parallel_tool_use_config(model: str) -> bool:
     return any(
         (litellm.model_cost.get(candidate) or {}).get("supports_parallel_tool_use_config") is True

tests/test_litellm/llms/bedrock/chat/test_converse_transformation.py

@@ -1021,6 +1021,49 @@ def test_client_metadata_kept_for_non_anthropic_converse_request():
     assert data["additionalModelRequestFields"]["client_metadata"] == {"originator": "codex_cli_rs"}
 
 
+@pytest.mark.parametrize(
+    "model",
+    [
+        "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abcdef123456",
+        "arn:aws:bedrock:us-east-1:123456789012:provisioned-model/abcdef123456",
+    ],
+)
+def test_client_metadata_stripped_for_arn_models_converse(model):
+    """An ARN hides which family serves the request, and pointing one at Claude is how
+    teams route codex traffic, so the field has to go there too or the 400 comes back."""
+    config = AmazonConverseConfig()
+
+    data = config._transform_request_helper(
+        model=model,
+        system_content_blocks=[],
+        optional_params={
+            "maxTokens": 16,
+            "client_metadata": {"originator": "codex_cli_rs"},
+        },
+        messages=None,
+    )
+
+    assert "client_metadata" not in data.get("additionalModelRequestFields", {})
+
+
+def test_client_metadata_kept_for_arn_naming_a_non_anthropic_family():
+    """An inference profile ARN that still spells out the family is resolvable, so a
+    non-Anthropic one keeps its passthrough."""
+    config = AmazonConverseConfig()
+
+    data = config._transform_request_helper(
+        model="arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.amazon.nova-pro-v1:0",
+        system_content_blocks=[],
+        optional_params={
+            "maxTokens": 16,
+            "client_metadata": {"originator": "codex_cli_rs"},
+        },
+        messages=None,
+    )
+
+    assert data["additionalModelRequestFields"]["client_metadata"] == {"originator": "codex_cli_rs"}
+
+
 def test_parallel_tool_calls_config_kept_for_sonnet_5(monkeypatch):
     old_env = os.environ.get("LITELLM_LOCAL_MODEL_COST_MAP")
     old_cost = litellm.model_cost