fix(pass-through): propagate x-litellm-spend-logs-metadata header

OpenAI-compatible endpoints parse `x-litellm-spend-logs-metadata` in
`LiteLLMProxyRequestSetup._get_spend_logs_metadata_from_request_headers`
and place it at `data["metadata"]["spend_logs_metadata"]`. Pass-through
endpoints (/anthropic/v1/messages, /openai/v1/*, /v2/chat, /vertex_ai/*,
/gemini/*, Cursor) bypass that pipeline, so
`StandardLoggingPayload.metadata.spend_logs_metadata` was always None
for `call_type=pass_through_endpoint`, hiding per-call metadata
(e.g. task / issue) from `/spend/logs`.

This change extracts the header at both `PassthroughStandardLoggingPayload`
construction sites (HTTP and WebSocket), adds an optional
`spend_logs_metadata` field to the TypedDict, and propagates it into
`kwargs["litellm_params"]["metadata"]` at every provider-handler
payload builder (base, Anthropic, OpenAI, Cohere, Gemini, Vertex) — both
the non-streaming and streaming (collected-chunks) paths. Two small
shared helpers on `BasePassthroughLoggingHandler` keep the propagation
consistent across handlers:

- `_apply_spend_logs_metadata(kwargs, payload)` — writes to the same
  key path OpenAI-compatible flows already use, so
  `get_standard_logging_object_payload` surfaces the value identically.
- `_seed_streaming_kwargs_from_logging_obj(logging_obj)` — seeds the
  fresh kwargs used by streaming logging paths with the payload that
  `success_handler.pass_through_async_success_handler` stashes on
  `model_call_details` before any provider handler runs.

Tests in `tests/test_litellm/proxy/pass_through_endpoints/test_spend_logs_metadata_propagation.py`
cover: header extraction (valid / missing / malformed JSON), helper
writes to the matching key path, preservation of existing metadata
keys, no-op when payload or metadata is absent, streaming-kwargs seed
behavior.

litellm/proxy/pass_through_endpoints/architecture.md

@@ -59,6 +59,7 @@ sequenceDiagram
 |-----------|-------------|
 | `x-pass-*` headers | Strip prefix and forward (e.g., `x-pass-anthropic-beta` → `anthropic-beta`) |
 | `x-litellm-tags` header | Extract tags and add to request metadata for logging |
+| `x-litellm-spend-logs-metadata` header | Parse as JSON and propagate into `StandardLoggingPayload.metadata.spend_logs_metadata` (matches OpenAI-compatible endpoints) |
 | Streaming chunk collection | Collect chunks async for logging after stream completes |
 | Multipart form handling | Reconstruct multipart/form-data requests for file uploads |
 | Guardrails (opt-in) | Run content filtering when explicitly configured |

litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py

@@ -19,6 +19,8 @@ from litellm.types.passthrough_endpoints.pass_through_endpoints import (
 )
 from litellm.types.utils import LiteLLMBatch, ModelResponse, TextCompletionResponse
 
+from .base_passthrough_logging_handler import BasePassthroughLoggingHandler
+
 if TYPE_CHECKING:
     from litellm.types.passthrough_endpoints.pass_through_endpoints import EndpointType
 
@@ -157,6 +159,10 @@ class AnthropicPassthroughLoggingHandler:
                         {"proxy_server_request": {"body": {"user": user}}}
                     )
 
+                BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+                    kwargs, passthrough_logging_payload
+                )
+
             # pretty print standard logging object
             verbose_proxy_logger.debug(
                 "kwargs= %s",
@@ -224,7 +230,9 @@ class AnthropicPassthroughLoggingHandler:
         kwargs = AnthropicPassthroughLoggingHandler._create_anthropic_response_logging_payload(
             litellm_model_response=complete_streaming_response,
             model=model,
-            kwargs={},
+            kwargs=BasePassthroughLoggingHandler._seed_streaming_kwargs_from_logging_obj(
+                litellm_logging_obj
+            ),
             start_time=start_time,
             end_time=end_time,
             logging_obj=litellm_logging_obj,

litellm/proxy/pass_through_endpoints/llm_provider_handlers/base_passthrough_logging_handler.py

@@ -93,6 +93,49 @@ class BasePassthroughLoggingHandler(ABC):
             return get_end_user_id_from_request_body(request_body)
         return None
 
+    @staticmethod
+    def _apply_spend_logs_metadata(
+        kwargs: dict,
+        passthrough_logging_payload: Optional["PassthroughStandardLoggingPayload"],
+    ) -> None:
+        """
+        Propagate `spend_logs_metadata` from a pass-through logging payload into
+        `kwargs["litellm_params"]["metadata"]`.
+
+        OpenAI-compatible endpoints write the value at
+        `data["metadata"]["spend_logs_metadata"]` before the call; this helper
+        matches that key path at logging time so downstream
+        `get_standard_logging_object_payload` surfaces the same field for
+        pass-through traffic.
+        """
+        if not passthrough_logging_payload:
+            return
+        metadata = passthrough_logging_payload.get("spend_logs_metadata")
+        if not metadata:
+            return
+        kwargs.setdefault("litellm_params", {})
+        kwargs["litellm_params"].setdefault("metadata", {})
+        kwargs["litellm_params"]["metadata"]["spend_logs_metadata"] = metadata
+
+    @staticmethod
+    def _seed_streaming_kwargs_from_logging_obj(
+        litellm_logging_obj: LiteLLMLoggingObj,
+    ) -> dict:
+        """
+        Build the initial `kwargs` dict for streaming logging paths so the
+        pass-through payload (populated by `success_handler.pass_through_async_success_handler`
+        before any provider handler runs) survives into the response logging
+        builder. Without this seed, streaming paths would lose the
+        `spend_logs_metadata` that non-streaming traffic carries via kwargs.
+        """
+        initial: dict = {}
+        passthrough_payload = litellm_logging_obj.model_call_details.get(
+            "passthrough_logging_payload"
+        )
+        if passthrough_payload is not None:
+            initial["passthrough_logging_payload"] = passthrough_payload
+        return initial
+
     def _create_response_logging_payload(
         self,
         litellm_model_response: Union[ModelResponse, TextCompletionResponse],
@@ -129,6 +172,8 @@ class BasePassthroughLoggingHandler(ABC):
                         {"proxy_server_request": {"body": {"user": user}}}
                     )
 
+                self._apply_spend_logs_metadata(kwargs, passthrough_logging_payload)
+
             # Make standard logging object for Anthropic
             standard_logging_object = get_standard_logging_object_payload(
                 kwargs=kwargs,
@@ -209,7 +254,7 @@ class BasePassthroughLoggingHandler(ABC):
         kwargs = self._create_response_logging_payload(
             litellm_model_response=complete_streaming_response,
             model=model,
-            kwargs={},
+            kwargs=self._seed_streaming_kwargs_from_logging_obj(litellm_logging_obj),
             start_time=start_time,
             end_time=end_time,
             logging_obj=litellm_logging_obj,

litellm/proxy/pass_through_endpoints/llm_provider_handlers/cohere_passthrough_logging_handler.py

@@ -142,6 +142,10 @@ class CoherePassthroughLoggingHandler(BasePassthroughLoggingHandler):
                             {"proxy_server_request": {"body": {"user": user}}}
                         )
 
+                    BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+                        kwargs, passthrough_logging_payload
+                    )
+
                 # Create standard logging object
                 if litellm_model_response is not None:
                     get_standard_logging_object_payload(

litellm/proxy/pass_through_endpoints/llm_provider_handlers/gemini_passthrough_logging_handler.py

@@ -17,6 +17,8 @@ from litellm.types.utils import (
     TextCompletionResponse,
 )
 
+from .base_passthrough_logging_handler import BasePassthroughLoggingHandler
+
 if TYPE_CHECKING:
     from litellm.types.passthrough_endpoints.pass_through_endpoints import EndpointType
 
@@ -139,7 +141,11 @@ class GeminiPassthroughLoggingHandler:
         - Creates standard logging object
         - Logs in litellm callbacks
         """
-        kwargs: Dict[str, Any] = {}
+        kwargs: Dict[str, Any] = (
+            BasePassthroughLoggingHandler._seed_streaming_kwargs_from_logging_obj(
+                litellm_logging_obj
+            )
+        )
         model = model or GeminiPassthroughLoggingHandler.extract_model_from_url(
             url_route
         )
@@ -242,6 +248,10 @@ class GeminiPassthroughLoggingHandler:
         kwargs["model"] = model
         kwargs["custom_llm_provider"] = custom_llm_provider
 
+        BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+            kwargs, kwargs.get("passthrough_logging_payload")
+        )
+
         # pretty print standard logging object
         verbose_proxy_logger.debug("kwargs= %s", kwargs)
 

litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_passthrough_logging_handler.py

@@ -379,6 +379,10 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                         "proxy_server_request", {}
                     ).setdefault("body", {})["user"] = user
 
+                BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+                    kwargs, passthrough_logging_payload
+                )
+
             # Create standard logging object
             if litellm_model_response is not None:
                 get_standard_logging_object_payload(
@@ -572,6 +576,10 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                         "proxy_server_request", {}
                     ).setdefault("body", {})["user"] = user
 
+                BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+                    kwargs, passthrough_logging_payload
+                )
+
             # Create standard logging object
             get_standard_logging_object_payload(
                 kwargs=kwargs,

litellm/proxy/pass_through_endpoints/llm_provider_handlers/vertex_passthrough_logging_handler.py

@@ -16,6 +16,9 @@ from litellm.llms.vertex_ai.vector_stores.search_api.transformation import (
 )
 from litellm.llms.vertex_ai.videos.transformation import VertexAIVideoConfig
 from litellm.proxy._types import PassThroughEndpointLoggingTypedDict
+from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+    BasePassthroughLoggingHandler,
+)
 from litellm.types.utils import (
     Choices,
     EmbeddingResponse,
@@ -341,7 +344,11 @@ class VertexPassthroughLoggingHandler:
         - Creates standard logging object
         - Logs in litellm callbacks
         """
-        kwargs: Dict[str, Any] = {}
+        kwargs: Dict[str, Any] = (
+            BasePassthroughLoggingHandler._seed_streaming_kwargs_from_logging_obj(
+                litellm_logging_obj
+            )
+        )
         model = model or VertexPassthroughLoggingHandler.extract_model_from_url(
             url_route
         )
@@ -547,6 +554,10 @@ class VertexPassthroughLoggingHandler:
         kwargs["response_cost"] = response_cost
         kwargs["model"] = model
 
+        BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+            kwargs, kwargs.get("passthrough_logging_payload")
+        )
+
         # pretty print standard logging object
         verbose_proxy_logger.debug("kwargs= %s", kwargs)
 

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -752,11 +752,17 @@ async def pass_through_request(  # noqa: PLR0915
             params={"timeout": 600},
         )
         async_client = async_client_obj.client
+        spend_logs_metadata = (
+            LiteLLMProxyRequestSetup._get_spend_logs_metadata_from_request_headers(
+                _safe_get_request_headers(request)
+            )
+        )
         passthrough_logging_payload = PassthroughStandardLoggingPayload(
             url=str(url),
             request_body=_parsed_body,
             request_method=getattr(request, "method", None),
             cost_per_request=cost_per_request,
+            spend_logs_metadata=spend_logs_metadata,
         )
         kwargs = HttpPassThroughEndpointHelpers._init_kwargs_for_pass_through_endpoint(
             user_api_key_dict=user_api_key_dict,
@@ -1386,11 +1392,18 @@ async def websocket_passthrough_request(  # noqa: PLR0915
     )
 
     # Create passthrough logging payload
+    websocket_headers = dict(websocket.headers) if hasattr(websocket, "headers") else {}
+    spend_logs_metadata = (
+        LiteLLMProxyRequestSetup._get_spend_logs_metadata_from_request_headers(
+            websocket_headers
+        )
+    )
     passthrough_logging_payload = PassthroughStandardLoggingPayload(
         url=target,
         request_body={},  # WebSocket doesn't have a traditional request body
         request_method="WEBSOCKET",
         cost_per_request=cost_per_request,
+        spend_logs_metadata=spend_logs_metadata,
     )
 
     # Create a dummy request object for WebSocket connections to maintain compatibility

litellm/types/passthrough_endpoints/pass_through_endpoints.py

@@ -46,3 +46,12 @@ class PassthroughStandardLoggingPayload(TypedDict, total=False):
 
     Optional field, we use this for cost tracking only if it's set.
     """
+
+    spend_logs_metadata: Optional[dict]
+    """
+    Parsed `x-litellm-spend-logs-metadata` request header, if present.
+
+    Propagated into `StandardLoggingPayload.metadata.spend_logs_metadata` so
+    pass-through traffic surfaces the same per-call metadata (e.g. `task`,
+    `issue`) that OpenAI-compatible endpoints already emit.
+    """

tests/test_litellm/proxy/pass_through_endpoints/test_spend_logs_metadata_propagation.py

@@ -0,0 +1,135 @@
+"""Tests for `x-litellm-spend-logs-metadata` propagation through pass-through endpoints.
+
+OpenAI-compatible endpoints parse this header in
+`LiteLLMProxyRequestSetup._get_spend_logs_metadata_from_request_headers` and
+stuff it into `data["metadata"]["spend_logs_metadata"]`. Pass-through endpoints
+used to skip this extraction, so `/spend/logs` entries for
+`call_type=pass_through_endpoint` always had
+`metadata.spend_logs_metadata = None`.
+
+These tests pin the fix:
+
+1. Header extraction handles present/missing/malformed input.
+2. `BasePassthroughLoggingHandler._apply_spend_logs_metadata` writes the
+   metadata to the same key path the OpenAI-compatible flow uses.
+3. `_seed_streaming_kwargs_from_logging_obj` threads the pass-through payload
+   from `logging_obj.model_call_details` into fresh streaming kwargs, so
+   streaming responses retain the metadata.
+"""
+
+import json
+from unittest.mock import MagicMock
+
+import pytest
+
+
+@pytest.mark.parametrize(
+    "headers,expected",
+    [
+        ({"x-litellm-spend-logs-metadata": json.dumps({"task": "t1"})}, {"task": "t1"}),
+        ({}, None),
+        ({"x-litellm-spend-logs-metadata": "not-json"}, None),
+    ],
+)
+def test_header_extraction(headers, expected):
+    """Helper handles valid JSON, missing header, and malformed JSON."""
+    from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
+
+    assert (
+        LiteLLMProxyRequestSetup._get_spend_logs_metadata_from_request_headers(headers)
+        == expected
+    )
+
+
+def test_apply_spend_logs_metadata_writes_matching_key_path():
+    """The helper must write to `kwargs["litellm_params"]["metadata"]["spend_logs_metadata"]`
+    — the same key path OpenAI-compatible flows use, so
+    `get_standard_logging_object_payload` picks it up identically.
+    """
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+        BasePassthroughLoggingHandler,
+    )
+
+    kwargs: dict = {}
+    metadata = {"task": "cost-report", "issue": "#42"}
+
+    BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+        kwargs,
+        {"url": "x", "spend_logs_metadata": metadata},
+    )
+
+    assert kwargs["litellm_params"]["metadata"]["spend_logs_metadata"] == metadata
+
+
+def test_apply_spend_logs_metadata_preserves_existing_metadata():
+    """Existing metadata keys must not be clobbered when attaching
+    `spend_logs_metadata`.
+    """
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+        BasePassthroughLoggingHandler,
+    )
+
+    kwargs: dict = {"litellm_params": {"metadata": {"tags": ["prod"]}}}
+
+    BasePassthroughLoggingHandler._apply_spend_logs_metadata(
+        kwargs,
+        {"url": "x", "spend_logs_metadata": {"task": "t1"}},
+    )
+
+    assert kwargs["litellm_params"]["metadata"] == {
+        "tags": ["prod"],
+        "spend_logs_metadata": {"task": "t1"},
+    }
+
+
+@pytest.mark.parametrize(
+    "payload",
+    [None, {}, {"url": "x"}, {"url": "x", "spend_logs_metadata": None}],
+)
+def test_apply_spend_logs_metadata_is_noop_when_missing(payload):
+    """When the payload is missing or carries no metadata, the helper must
+    leave kwargs untouched — this is the path for existing callers that don't
+    send the header, and it must remain backward-compatible.
+    """
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+        BasePassthroughLoggingHandler,
+    )
+
+    kwargs: dict = {}
+    BasePassthroughLoggingHandler._apply_spend_logs_metadata(kwargs, payload)
+    assert kwargs == {}
+
+
+def test_seed_streaming_kwargs_copies_payload_from_logging_obj():
+    """Streaming logging paths build fresh kwargs; the seed helper must lift
+    the pass-through payload stashed by `success_handler` onto that fresh dict
+    so `_create_*_response_logging_payload` can still see it.
+    """
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+        BasePassthroughLoggingHandler,
+    )
+
+    payload = {"url": "x", "spend_logs_metadata": {"task": "t1"}}
+    logging_obj = MagicMock(model_call_details={"passthrough_logging_payload": payload})
+
+    seeded = BasePassthroughLoggingHandler._seed_streaming_kwargs_from_logging_obj(
+        logging_obj
+    )
+
+    assert seeded == {"passthrough_logging_payload": payload}
+
+
+def test_seed_streaming_kwargs_empty_when_no_payload():
+    """No pass-through payload on `logging_obj` → empty dict (prior behavior)."""
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.base_passthrough_logging_handler import (
+        BasePassthroughLoggingHandler,
+    )
+
+    logging_obj = MagicMock(model_call_details={})
+
+    assert (
+        BasePassthroughLoggingHandler._seed_streaming_kwargs_from_logging_obj(
+            logging_obj
+        )
+        == {}
+    )