fix(spend): preserve deployment identity in spend logs when litellm_metadata is present

Spend logging reads litellm_metadata when a request carries one, but the router
writes the selected deployment's identity (model_info.id, model_group, deployment
name) into metadata. add_missing_spend_metadata_to_litellm_metadata only copied
user_api_key* keys, so openai/ passthrough deployments logged a blank model_id and
a provider-prefix-stripped model for requests that included litellm_metadata.

Fixes #35472

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

litellm/litellm_core_utils/core_helpers.py

@@ -164,17 +164,33 @@ def remove_items_at_indices(items: Optional[List[Any]], indices: Iterable[int])
             items.pop(index)
 
 
+SPEND_TRACKING_DEPLOYMENT_IDENTITY_KEYS = (
+    "model_info",
+    "model_group",
+    "deployment",
+    "deployment_model_name",
+)
+
+
 def add_missing_spend_metadata_to_litellm_metadata(litellm_metadata: dict, metadata: dict) -> dict:
     """
     Helper to get litellm metadata for spend tracking
 
     PATCH for issue where both `litellm_metadata` and `metadata` are present in the kwargs
-    and user_api_key values are in 'metadata'.
+    and spend-tracking values are in 'metadata'.
+
+    The router always writes the selected deployment's identity (model_info.id,
+    model_group, deployment name) into `metadata`, but spend logging reads from
+    `litellm_metadata` whenever a request carries one (e.g. caller-supplied tags).
+    Without copying these over, the same deployment logs a blank `model_id` and a
+    provider-prefix-stripped `model` for requests that include `litellm_metadata`.
     """
-    potential_spend_tracking_metadata_substring = "user_api_key"
     for key, value in metadata.items():
-        if potential_spend_tracking_metadata_substring in key:
+        if "user_api_key" in key:
             litellm_metadata[key] = value
+    for key in SPEND_TRACKING_DEPLOYMENT_IDENTITY_KEYS:
+        if key not in litellm_metadata and key in metadata:
+            litellm_metadata[key] = metadata[key]
     return litellm_metadata
 
 

tests/test_litellm/litellm_core_utils/test_core_helpers.py

@@ -4,6 +4,7 @@ import pytest
 
 from litellm.litellm_core_utils.core_helpers import (
     _FINISH_REASON_MAP,
+    get_litellm_metadata_from_kwargs,
     get_or_create_metadata_bucket,
     map_finish_reason,
     reconstruct_model_name,
@@ -11,6 +12,73 @@ from litellm.litellm_core_utils.core_helpers import (
 )
 
 
+def _spend_log_identity(kwargs: dict) -> dict:
+    """Mirror how spend logging derives model / model_id from request kwargs."""
+    metadata = get_litellm_metadata_from_kwargs(kwargs)
+    return {
+        "model": reconstruct_model_name(
+            kwargs.get("model") or "", kwargs.get("custom_llm_provider"), metadata or {}
+        ),
+        "model_id": metadata.get("model_info", {}).get("id", ""),
+        "model_group": metadata.get("model_group", ""),
+    }
+
+
+def test_spend_log_deployment_identity_consistent_with_litellm_metadata():
+    """Regression for #35472: an openai/ passthrough deployment must log the same
+    model / model_id whether or not the request also carries litellm_metadata.
+
+    The router writes the deployment identity into `metadata`, but spend logging
+    reads `litellm_metadata` when present, so those keys must be carried over."""
+    router_metadata = {
+        "deployment": "openai/anthropic/claude-sonnet-5",
+        "model_info": {"id": "9da5dfc9-2223-4f77-b3c9-f9100d9cb2a0"},
+        "model_group": "my-model",
+        "user_api_key_hash": "abc",
+    }
+
+    without_litellm_metadata = _spend_log_identity(
+        {
+            "model": "anthropic/claude-sonnet-5",
+            "custom_llm_provider": "openai",
+            "litellm_params": {"metadata": dict(router_metadata)},
+        }
+    )
+    with_litellm_metadata = _spend_log_identity(
+        {
+            "model": "anthropic/claude-sonnet-5",
+            "custom_llm_provider": "openai",
+            "litellm_params": {
+                "metadata": dict(router_metadata),
+                "litellm_metadata": {"tags": ["t1"]},
+            },
+        }
+    )
+
+    expected = {
+        "model": "openai/anthropic/claude-sonnet-5",
+        "model_id": "9da5dfc9-2223-4f77-b3c9-f9100d9cb2a0",
+        "model_group": "my-model",
+    }
+    assert without_litellm_metadata == expected
+    assert with_litellm_metadata == expected
+
+
+def test_get_litellm_metadata_from_kwargs_does_not_overwrite_existing_identity():
+    """Caller-supplied identity keys in litellm_metadata must win over metadata."""
+    kwargs = {
+        "litellm_params": {
+            "metadata": {"model_group": "router-group", "deployment": "router-dep"},
+            "litellm_metadata": {"model_group": "caller-group"},
+        }
+    }
+
+    metadata = get_litellm_metadata_from_kwargs(kwargs)
+
+    assert metadata["model_group"] == "caller-group"
+    assert metadata["deployment"] == "router-dep"
+
+
 class TestGetOrCreateMetadataBucket:
     """The single owner every guardrail writer and reader shares, so the response
     header and the spend log can never disagree about which dict a record lives in."""