Merge remote-tracking branch 'origin/litellm_internal_staging' into litellm_lit_7174_stream_tool_call_rewrites

.github/workflows/_test-unit-base.yml

@@ -113,7 +113,7 @@ jobs:
         if: steps.changes.outputs.decision != 'skip'
         timeout-minutes: 8
         run: |
-          .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra mongodb
+          .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml
           uv run --no-sync python -c 'import os, sys; print(sys.version); assert f"{sys.version_info.major}.{sys.version_info.minor}" == os.environ["UV_PYTHON"]'
 
       - name: Cache Prisma binaries

.github/workflows/test-litellm-ui-unit.yml

@@ -66,7 +66,7 @@ jobs:
           BASE_SHA: ${{ github.event.pull_request.base.sha }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
-          full_suite() { npm run test -- --run --pool forks --poolOptions.forks.maxForks=14; }
+          full_suite() { npm run test -- --run --pool forks --maxWorkers=14; }
 
           if [ -z "$BASE_SHA" ]; then
             echo "Push to $GITHUB_REF_NAME: running the full suite"
@@ -95,4 +95,4 @@ jobs:
 
           echo "Pull request: running tests related to ${#changed_files[@]} changed UI files"
           npm run test -- related "${changed_files[@]}" --run --passWithNoTests \
-            --pool forks --poolOptions.forks.maxForks=14
+            --pool forks --maxWorkers=14

Dockerfile

@@ -67,7 +67,6 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
     --extra semantic-router \
     --extra saml \
     --extra bedrock-realtime \
-    --extra mongodb \
     --python python3.13
 
 # Copy full source tree
@@ -90,7 +89,6 @@ RUN uv sync --frozen --no-default-groups --no-editable \
     --extra semantic-router \
     --extra saml \
     --extra bedrock-realtime \
-    --extra mongodb \
     --python python3.13
 
 RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

docker/Dockerfile.database

@@ -65,7 +65,6 @@ RUN uv sync --frozen --no-install-project --no-install-workspace --no-default-gr
     --extra semantic-router \
     --extra saml \
     --extra bedrock-realtime \
-    --extra mongodb \
     --python python3.13
 
 # Copy full source tree
@@ -88,7 +87,6 @@ RUN uv sync --frozen --no-default-groups --no-editable \
     --extra semantic-router \
     --extra saml \
     --extra bedrock-realtime \
-    --extra mongodb \
     --python python3.13
 
 RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

docker/Dockerfile.non_root

@@ -71,7 +71,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
     --extra semantic-router \
     --extra saml \
     --extra bedrock-realtime \
-    --extra mongodb \
     --python python3.13
 
 # Copy full source tree
@@ -100,7 +99,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
         --extra semantic-router \
         --extra saml \
         --extra bedrock-realtime \
-        --extra mongodb \
         --python python3.13 \
         --no-sources-package litellm-proxy-extras; \
     else \
@@ -111,7 +109,6 @@ RUN --mount=type=cache,target=/app/.cache/uv,id=litellm-uv-cache \
         --extra semantic-router \
         --extra saml \
         --extra bedrock-realtime \
-        --extra mongodb \
         --python python3.13; \
     fi
 

gateway/Dockerfile

@@ -47,7 +47,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra extra_proxy \
         --extra semantic-router \
         --extra bedrock-realtime \
-        --extra mongodb \
         --python python3.13
 
 # Stage 2 — copy source and install the project + workspace members.
@@ -60,7 +59,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra extra_proxy \
         --extra semantic-router \
         --extra bedrock-realtime \
-        --extra mongodb \
         --python python3.13
 
 RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \

litellm/__init__.py

@@ -326,6 +326,9 @@ ssl_certificate: Optional[str] = None
 user_url_validation: bool = True
 user_url_allowed_hosts: List[str] = []
 provider_url_destination_allowed_hosts: List[str] = []
+#: "override" (default) or "additive": whether a key or team destination replaces
+#: the operator's exporter for that backend or exports alongside it.
+otel_tenant_destination_mode: str | None = None
 ssl_ecdh_curve: Optional[str] = None  # Set to 'X25519' to disable PQC and improve performance
 disable_streaming_logging: bool = False
 disable_token_counter: bool = False

litellm/constants.py

@@ -1374,6 +1374,7 @@ bedrock_embedding_models: Final[set] = set(
         "cohere.embed-multilingual-v3",
         "cohere.embed-v4:0",
         "twelvelabs.marengo-embed-2-7-v1:0",
+        "twelvelabs.marengo-embed-3-0-v1:0",
     ]
 )
 

litellm/integrations/callback_configs.json

@@ -367,6 +367,13 @@
         "ui_name": "Headers",
         "description": "Headers for OTEL exporter (e.g., x-honeycomb-team=YOUR_API_KEY)",
         "required": false
+      },
+      "otel_exporter_otlp_protocol": {
+        "type": "select",
+        "ui_name": "Export Protocol",
+        "description": "OTLP wire format for trace exports. Use http/json for collectors that cannot decode protobuf",
+        "options": ["http/protobuf", "http/json"],
+        "required": false
       }
     },
     "description": "OpenTelemetry Logging Integration"

litellm/integrations/mlflow.py

@@ -133,17 +133,17 @@ class MlflowLogger(CustomLogger):
         if final_response:
             end_time_ns: Final = int(end_time.timestamp() * 1e9)
 
-            self._extract_and_set_chat_attributes(span, kwargs, final_response)
-            self._end_span_or_trace(
-                span=span,
-                outputs=final_response,
-                status=SpanStatusCode.OK,
-                end_time_ns=end_time_ns,
-            )
-
-            # Remove the stream_id from the map
-            with self._lock:
-                self._stream_id_to_span.pop(litellm_call_id)
+            try:
+                self._extract_and_set_chat_attributes(span, kwargs, final_response)
+                self._end_span_or_trace(
+                    span=span,
+                    outputs=final_response,
+                    status=SpanStatusCode.OK,
+                    end_time_ns=end_time_ns,
+                )
+            finally:
+                with self._lock:
+                    self._stream_id_to_span.pop(litellm_call_id, None)
 
     def _add_chunk_events(self, span, response_obj):
         from mlflow.entities import SpanEvent
@@ -282,15 +282,15 @@ class MlflowLogger(CustomLogger):
         """End an MLflow span or a trace."""
         if span.parent_id is None:
             self._client.end_trace(
-                trace_id=span.request_id,
+                span.request_id,
                 outputs=outputs,
                 status=status,
                 end_time_ns=end_time_ns,
             )
         else:
             self._client.end_span(
-                trace_id=span.request_id,
-                span_id=span.span_id,
+                span.request_id,
+                span.span_id,
                 outputs=outputs,
                 status=status,
                 end_time_ns=end_time_ns,

litellm/integrations/otel/logger.py

@@ -16,9 +16,11 @@ from opentelemetry.trace import (
     Span,
     Tracer,
     get_current_span,
+    get_tracer_provider,
     set_span_in_context,
     use_span,
 )
+from opentelemetry.trace import TracerProvider as ApiTracerProvider
 
 import litellm
 from litellm._logging import verbose_logger
@@ -63,6 +65,7 @@ from litellm.integrations.otel.plumbing.metrics import (
     create_genai_metrics,
 )
 from litellm.integrations.otel.plumbing.providers import (
+    attach_tenant_fan_out,
     build_tracer_provider,
     get_event_logger,
     get_meter,
@@ -85,6 +88,7 @@ if TYPE_CHECKING:
     )
 
 LITELLM_TRACER_NAME: Final = "litellm"
+_published_v2_provider: ApiTracerProvider | None = None
 
 
 def _span_error_from_exception(
@@ -180,7 +184,9 @@ class OpenTelemetryV2(CustomLogger):
         self.config: OpenTelemetryV2Config = config or OpenTelemetryV2Config(**kwargs)
         self.callback_name = callback_name
         self._tracer_provider: TracerProvider = (
-            tracer_provider if tracer_provider is not None else build_tracer_provider(self.config)
+            tracer_provider
+            if tracer_provider is not None
+            else build_tracer_provider(self.config, tenant_overrides=True)
         )
         self.tracer: Tracer = get_tracer(self._tracer_provider, LITELLM_TRACER_NAME)
         self._metrics_recorder = self._init_metrics(meter_provider)
@@ -195,6 +201,11 @@ class OpenTelemetryV2(CustomLogger):
         self._open_llm_calls: OrderedDict[str, _LLMCallSpan] = OrderedDict()
         self._init_otel_logger_on_litellm_proxy()
 
+    @property
+    def tracer_provider(self) -> TracerProvider:
+        """The provider this logger emits through, read-only to its callers."""
+        return self._tracer_provider
+
     def _init_metrics(self, meter_provider: "MeterProvider | None") -> "GenAIMetricRecorder | None":
         """Create the six GenAI histograms when metrics are enabled, else ``None``.
 
@@ -863,12 +874,33 @@ def publish_global_otel_v2_provider(
     ``opentelemetry.trace.set_tracer_provider``) are injected so the publish step is
     unit-testable without reading or mutating real global OTel state. Returns the
     logger whose provider was published.
+
+    The published provider is also the one that fans spans out to key/team
+    destinations, because it is the only provider the whole request tree passes
+    through; see :func:`attach_tenant_fan_out`. It is remembered for
+    :func:`fan_out_provider` because neither the OTel global (``set_tracer_provider``
+    keeps the first provider it was ever handed) nor
+    ``proxy_server.open_telemetry_logger`` (a legacy v1 logger can hold that slot)
+    reliably leads back to it.
     """
+    global _published_v2_provider
     logger: Final = select_global_otel_v2_logger(in_memory_loggers, registered=registered)
-    set_global_provider(logger._tracer_provider)
+    attach_tenant_fan_out(logger.tracer_provider, *_v2_configs(in_memory_loggers, logger))
+    set_global_provider(logger.tracer_provider)
+    _published_v2_provider = logger.tracer_provider  # rebind-ok: startup records the one provider carrying the fan-out
     return logger
 
 
+def _v2_configs(in_memory_loggers: Sequence[object], logger: "OpenTelemetryV2") -> tuple[OpenTelemetryV2Config, ...]:
+    """Every v2 logger's config, the published logger's first.
+
+    Each preset keeps its own provider and exporters, so the accounts the operator
+    writes to are spread over all of them, not held by the published logger alone.
+    """
+    others: Final = tuple(cb.config for cb in in_memory_loggers if isinstance(cb, OpenTelemetryV2) and cb is not logger)
+    return (logger.config, *others)
+
+
 def _registered_v2_logger() -> "OpenTelemetryV2 | None":
     try:
         from litellm.proxy import proxy_server
@@ -904,6 +936,25 @@ def seed_request_identity(user_api_key_dict: object, model: str | None = None) -
         logger.seed_request_identity(user_api_key_dict, model=model)
 
 
+def fan_out_provider() -> ApiTracerProvider:
+    """The provider :func:`publish_global_otel_v2_provider` gave the tenant fan-out.
+
+    Read off the publish itself, not the OTel global and not the registered logger:
+    the global keeps whichever provider claimed it first (auto-instrumentation, a
+    legacy logger), and the registered slot can hold a v1 logger while the publish
+    picked a v2 one from ``_in_memory_loggers``. Either detour lands on a provider
+    with no fan-out and drops every destination at auth.
+    """
+    published: Final = _published_v2_provider
+    if published is not None:
+        return published
+    logger: Final = _registered_v2_logger()
+    if logger is not None:
+        attach_tenant_fan_out(logger.tracer_provider, logger.config)
+        return logger.tracer_provider
+    return get_tracer_provider()
+
+
 @contextmanager
 def phase_span(name: str) -> "Iterator[Span | None]":
     logger: Final = _registered_v2_logger()

litellm/integrations/otel/mappers/legacy.py

@@ -23,6 +23,7 @@ from litellm.integrations.otel.model.payloads import (
     ServiceSpanData,
     ToolDefinition,
 )
+from litellm.integrations.otel.model.semconv import Error
 
 # Attribute keys in the semconv-ai / Traceloop vocabulary.
 _LEGACY_SYSTEM: Final = "gen_ai.system"
@@ -36,7 +37,7 @@ _LEGACY_PRESENCE_PENALTY: Final = "llm.presence_penalty"
 _LEGACY_STOP_SEQUENCES: Final = "llm.chat.stop_sequences"
 _LEGACY_SERVICE: Final = "service"
 _LEGACY_CALL_TYPE: Final = "call_type"
-_LEGACY_ERROR: Final = "error"
+_LEGACY_ERROR: Final = Error.MESSAGE_LEGACY
 
 
 class LegacyMapper:

litellm/integrations/otel/model/config.py

@@ -69,7 +69,7 @@ class ExporterSpec(BaseModel):
 
     kind: str = Field(
         default="console",
-        description="console | in_memory | otlp_http | otlp_grpc | <factory kind>",
+        description="console | in_memory | otlp_http | http/json | otlp_grpc | <factory kind>",
     )
     endpoint: str | None = None
     traces_endpoint: str | None = Field(
@@ -269,7 +269,9 @@ class OpenTelemetryV2Config(BaseSettings):
         if (self.endpoint or self.traces_endpoint) and self.exporter == "console":
             self.exporter = "otlp_http"
         # When no explicit destinations are given, fold the single-destination
-        # shorthand into one spec so the provider always has a destination.
+        # shorthand into one spec so the provider always has a destination. A spec
+        # with no fields set is how the presets tell "nothing configured" from an
+        # operator who asked for the console by name.
         if not self.exporters:
             self.exporters = [
                 ExporterSpec(
@@ -278,6 +280,8 @@ class OpenTelemetryV2Config(BaseSettings):
                     traces_endpoint=self.traces_endpoint,
                     headers=self.headers,
                 )
+                if not self.model_fields_set.isdisjoint(("exporter", "endpoint", "headers"))
+                else ExporterSpec()
             ]
         # Ensure ``genai`` is always present and first.
         names = list(self.mapper_names)

litellm/integrations/otel/model/destination.py

@@ -0,0 +1,49 @@
+"""The resolved OTLP destination a request's traces export to.
+
+Backend-agnostic on purpose: every OTEL backend reduces to an endpoint plus auth
+headers. The per-backend field mapping lives in ``presets.destinations``.
+"""
+
+from collections.abc import Mapping
+from typing import Final
+from urllib.parse import quote
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class OtelDestination(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    endpoint: str
+    headers: Mapping[str, str] = Field(default_factory=dict)
+    resource_attributes: Mapping[str, str] = Field(default_factory=dict)
+    callback_name: str | None = None
+    protocol: str | None = Field(
+        default=None,
+        description=(
+            "OTLP transport, defaulting to the backend's own. Not derivable from the "
+            "scheme: Arize's ``https://otlp.arize.com/v1`` is gRPC."
+        ),
+    )
+
+    def header_string(self) -> str:
+        """Render headers as the ``k=v,k2=v2`` form an ``ExporterSpec`` expects.
+
+        Values are percent-encoded because ``providers.parse_headers`` decodes them
+        with the SDK's W3C-Baggage parser: a value carrying a ``,`` or ``=`` (a
+        Langfuse project name, a base64 Authorization payload ending in ``==``)
+        would otherwise be split into bogus pairs on the way back out.
+        """
+        return ",".join(f"{key}={quote(value, safe='')}" for key, value in self.headers.items())
+
+    def cache_key(self) -> tuple[str, tuple[tuple[str, str], ...], tuple[tuple[str, str], ...], str | None]:
+        """Identity for processor reuse, so one destination means one exporter."""
+        return (
+            self.endpoint,
+            tuple(sorted(self.headers.items())),
+            tuple(sorted(self.resource_attributes.items())),
+            self.protocol,
+        )
+
+
+NO_DESTINATIONS: Final[tuple[OtelDestination, ...]] = ()

litellm/integrations/otel/model/semconv.py

@@ -204,6 +204,9 @@ class Error:
 
     TYPE: Final = "error.type"
     MESSAGE: Final = "error.message"
+    # The same text under the bare key the semconv-ai / Traceloop vocabulary uses
+    # (see ``LegacyMapper``), so anything reading or redacting error text covers both.
+    MESSAGE_LEGACY: Final = "error"
 
 
 class LiteLLMError:

litellm/integrations/otel/plumbing/context.py

@@ -1,8 +1,9 @@
 """Trace-context + Baggage helpers."""
 
+import os
 from collections.abc import Mapping
 from contextvars import ContextVar, Token
-from typing import Final
+from typing import TYPE_CHECKING, Final
 
 from opentelemetry import baggage
 from opentelemetry.context import Context, get_current
@@ -21,6 +22,9 @@ from opentelemetry.trace.propagation.tracecontext import (
 
 from litellm.integrations.otel.model.semconv import HTTP
 
+if TYPE_CHECKING:
+    from litellm.integrations.otel.model.destination import OtelDestination
+
 _PROPAGATOR: Final = TraceContextTextMapPropagator()
 
 # The request's root span — the FastAPI-owned SERVER span — captured ONCE when the
@@ -304,3 +308,65 @@ def extract_traceparent(headers: Mapping[str, str]) -> Context | None:
         return None
     carrier: Final = {str(key).lower(): value for key, value in headers.items()}
     return _PROPAGATOR.extract(carrier)
+
+
+# The OTLP destinations this request's key or team pointed its traces at, resolved
+# once during auth. A ``ContextVar`` for the same reason the root span above is one:
+# it rides the request task's context into the ``asyncio.create_task`` children that
+# close the LLM span, and it is visible to every ``SpanProcessor.on_end`` that fires
+# on the request task. Stateful MCP handlers set and reset it per message; the
+# request-task value otherwise dies with that task.
+_request_destinations: Final['ContextVar[tuple["OtelDestination", ...]]'] = ContextVar(
+    "litellm_otel_request_destinations", default=()
+)
+
+
+def set_request_destinations(destinations: 'tuple["OtelDestination", ...]') -> "Token[tuple[OtelDestination, ...]]":
+    """Anchor the destinations this request exports to and return a reset token."""
+    return _request_destinations.set(destinations)
+
+
+def reset_request_destinations(token: "Token[tuple[OtelDestination, ...]]") -> None:
+    _request_destinations.reset(token)
+
+
+def request_destinations() -> 'tuple["OtelDestination", ...]':
+    """The destinations resolved for this request, empty outside a proxy request."""
+    return _request_destinations.get()
+
+
+#: ``litellm_settings: otel_tenant_destination_mode`` and its env equivalent.
+ADDITIVE_DESTINATION_MODE: Final = "additive"
+OTEL_TENANT_DESTINATION_MODE_ENV: Final = "LITELLM_OTEL_TENANT_DESTINATION_MODE"
+
+
+def tenant_destinations_are_additive() -> bool:
+    """Whether a tenant destination exports alongside the operator's own exporter.
+
+    Override is the default: the tenant's traffic reaches the tenant's account and
+    nowhere else. Operators running one org-wide backend across every team set this
+    to ``additive`` so the same trace lands in both places.
+    """
+    import litellm
+
+    configured: Final = litellm.otel_tenant_destination_mode or os.environ.get(OTEL_TENANT_DESTINATION_MODE_ENV)
+    return isinstance(configured, str) and configured.strip().lower() == ADDITIVE_DESTINATION_MODE
+
+
+def destination_backends() -> frozenset[str]:
+    """Backends this request resolved a tenant destination for.
+
+    The fan-out already carries the whole trace to those destinations, so the
+    per-request tracer route must never send a second copy, in either mode.
+    """
+    return frozenset(d.callback_name for d in _request_destinations.get() if d.callback_name)
+
+
+def suppressed_backends() -> frozenset[str]:
+    """Backends whose operator-level exporters this request must NOT reach.
+
+    Empty under ``additive``, where the operator keeps its copy of every span.
+    """
+    if tenant_destinations_are_additive():
+        return frozenset()
+    return destination_backends()

litellm/integrations/otel/plumbing/otlp_json.py

@@ -0,0 +1,70 @@
+"""OTLP/HTTP span exporter that sends the OTLP/JSON encoding instead of protobuf.
+
+The SDK only ships a protobuf OTLP/HTTP exporter; this reuses its transport and
+retry loop and swaps the payload for OTLP/JSON (enums as integers, ids as hex).
+"""
+
+import base64
+import json
+from collections.abc import Mapping, Sequence
+from types import MappingProxyType
+from typing import Final, TypeAlias
+
+from google.protobuf.json_format import MessageToDict
+from opentelemetry.exporter.otlp.proto.common.trace_encoder import encode_spans
+from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.trace import ReadableSpan
+
+JSON_CONTENT_TYPE: Final = "application/json"
+_HEX_ID_KEYS: Final = frozenset({"traceId", "spanId", "parentSpanId"})
+
+_JsonValue: TypeAlias = "Mapping[str, _JsonValue] | Sequence[_JsonValue] | str | int | float | bool | None"
+_JsonObject: TypeAlias = Mapping[str, "_JsonValue"]
+
+
+def _objects(node: _JsonObject, key: str) -> tuple[_JsonObject, ...]:
+    items: Final = node.get(key)
+    if isinstance(items, str) or not isinstance(items, Sequence):
+        return ()
+    return tuple(item for item in items if isinstance(item, Mapping))
+
+
+def _hex_ids(node: _JsonObject) -> _JsonObject:
+    return MappingProxyType(
+        {
+            key: base64.b64decode(item).hex() if key in _HEX_ID_KEYS and isinstance(item, str) else item
+            for key, item in node.items()
+        }
+    )
+
+
+def _hex_span(span: _JsonObject) -> _JsonObject:
+    links: Final = _objects(span, "links")
+    if not links:
+        return _hex_ids(span)
+    return MappingProxyType({**_hex_ids(span), "links": tuple(_hex_ids(link) for link in links)})
+
+
+def _hex_scope_spans(scope: _JsonObject) -> _JsonObject:
+    return MappingProxyType({**scope, "spans": tuple(_hex_span(span) for span in _objects(scope, "spans"))})
+
+
+def _hex_resource_spans(resource: _JsonObject) -> _JsonObject:
+    scope_spans: Final = tuple(_hex_scope_spans(scope) for scope in _objects(resource, "scopeSpans"))
+    return MappingProxyType({**resource, "scopeSpans": scope_spans})
+
+
+def encode_spans_json(spans: Sequence[ReadableSpan]) -> bytes:
+    payload: Final[_JsonObject] = MessageToDict(encode_spans(spans), use_integers_for_enums=True)
+    resource_spans: Final = tuple(_hex_resource_spans(resource) for resource in _objects(payload, "resourceSpans"))
+    hexed: Final[_JsonObject] = MappingProxyType({**payload, "resourceSpans": resource_spans})
+    return json.dumps(hexed, default=dict, separators=(",", ":")).encode()
+
+
+class OTLPJsonSpanExporter(OTLPSpanExporter):
+    def __init__(self, endpoint: str | None, headers: dict[str, str]) -> None:  # mutable-ok: SDK __init__ takes Dict
+        super().__init__(endpoint=endpoint, headers=headers)
+        self._session.headers["Content-Type"] = JSON_CONTENT_TYPE
+
+    def _serialize_spans(self, spans: Sequence[ReadableSpan]) -> bytes:
+        return encode_spans_json(spans)

litellm/integrations/otel/plumbing/providers.py

@@ -1,9 +1,14 @@
 """Provider / exporter factory + the Baggage span processor."""
 
-from collections.abc import Callable, Iterable
+import queue
+import threading
+import time
+from collections import OrderedDict
+from collections.abc import Callable, Iterable, Mapping, Sequence
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal
 
-from opentelemetry import _logs, baggage, metrics
+from opentelemetry import _logs, baggage, metrics, trace
 from opentelemetry._events import EventLogger
 from opentelemetry._logs import LoggerProvider, NoOpLoggerProvider
 from opentelemetry.context import Context
@@ -19,7 +24,8 @@ from opentelemetry.sdk._logs.export import (
 )
 from opentelemetry.sdk.metrics import MeterProvider as SDKMeterProvider
 from opentelemetry.sdk.resources import Resource
-from opentelemetry.sdk.trace import ReadableSpan, SpanProcessor, TracerProvider
+from opentelemetry.sdk.trace import Event, ReadableSpan, SpanProcessor, TracerProvider
+from opentelemetry.sdk.trace import Span as SDKSpan
 from opentelemetry.sdk.trace.export import (
     BatchSpanProcessor,
     ConsoleSpanExporter,
@@ -29,18 +35,35 @@ from opentelemetry.sdk.trace.export import (
 from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
     InMemorySpanExporter,
 )
-from opentelemetry.trace import Span, SpanKind, Tracer
+from opentelemetry.trace import Span, SpanKind, Status, Tracer
 from opentelemetry.util.re import parse_env_headers
+from opentelemetry.util.types import Attributes, AttributeValue
 
+from litellm._logging import verbose_logger
 from litellm._version import version as litellm_version
 from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
-from litellm.integrations.otel.model.semconv import LiteLLM
+from litellm.integrations.otel.model.semconv import (
+    DB,
+    MCP,
+    Error,
+    ExceptionEvent,
+    GenAI,
+    LiteLLM,
+    LiteLLMError,
+    Server,
+)
 from litellm.integrations.otel.model.spans import LiteLLMSpanKind
+from litellm.integrations.otel.plumbing.context import (
+    request_destinations,
+    suppressed_backends,
+)
 
 if TYPE_CHECKING:
     from opentelemetry.metrics import Meter
     from opentelemetry.sdk.metrics.export import MetricReader
 
+    from litellm.integrations.otel.model.destination import OtelDestination
+
 _SPAN_KIND_BY_ROLE_KIND: Final[dict[LiteLLMSpanKind, SpanKind]] = {
     LiteLLMSpanKind.SERVER: SpanKind.SERVER,
     LiteLLMSpanKind.CLIENT: SpanKind.CLIENT,
@@ -136,7 +159,8 @@ def parse_headers(raw: str | None) -> dict[str, str]:
 
 
 _IN_MEMORY_KINDS: Final = ("in_memory", "inmemory", "memory")
-_OTLP_HTTP_KINDS: Final = ("otlp_http", "http", "http/protobuf", "http/json")
+_OTLP_HTTP_JSON_KINDS: Final = ("http/json",)
+_OTLP_HTTP_KINDS: Final = ("otlp_http", "http", "http/protobuf", *_OTLP_HTTP_JSON_KINDS)
 _OTLP_GRPC_KINDS: Final = ("otlp_grpc", "grpc")
 
 
@@ -164,6 +188,13 @@ def _exporter_from_spec(spec: ExporterSpec) -> SpanExporter:
         return factory(spec)
     if kind in _IN_MEMORY_KINDS:
         return InMemorySpanExporter()
+    if kind in _OTLP_HTTP_JSON_KINDS:
+        from litellm.integrations.otel.plumbing.otlp_json import OTLPJsonSpanExporter
+
+        return OTLPJsonSpanExporter(
+            endpoint=spec.traces_endpoint or _otlp_traces_endpoint(spec.endpoint),
+            headers=parse_headers(spec.headers),
+        )
     if kind in _OTLP_HTTP_KINDS:
         from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
             OTLPSpanExporter as HTTPExporter,
@@ -194,6 +225,555 @@ def _processor_for(exporter: SpanExporter, use_simple: bool | None) -> SpanProce
     return SimpleSpanProcessor(exporter) if use_simple else BatchSpanProcessor(exporter)
 
 
+#: Distinct tenant destinations whose exporters stay alive. Each holds a connection
+#: pool and a batch thread, so the cache is bounded and evicts least-recently-used.
+_MAX_CACHED_DESTINATION_PROCESSORS: Final = 32
+
+#: Workers closing shed destination processors, bounding the threads a tenant can
+#: create by cycling its destination config.
+_DRAIN_WORKERS: Final = 2
+
+#: Shed processors waiting to be closed before the fan-out stops building new ones.
+#: Each still owns a batch thread until its close returns, and a collector that never
+#: answers makes every close take the exporter's full timeout, so past this many the
+#: operator's exporter keeps the span instead (see ``deliverable``).
+_MAX_PENDING_DRAINS: Final = 64
+
+#: How long ``shutdown`` waits for spans already being forwarded, so teardown closes
+#: no processor under one. Bounded: an exporter that never returns must not hold the
+#: proxy open.
+_SHUTDOWN_DRAIN_SECONDS: Final = 5.0
+
+#: An exporter's account: its normalized endpoint and the credentials it presents.
+_SinkKey = tuple[str, tuple[tuple[str, str], ...]]
+
+#: Header names that spell one credential two ways. Arize's operator exporter sends
+#: ``space_id`` where a tenant destination sends ``arize-space-id``.
+_CREDENTIAL_ALIASES: Final = MappingProxyType({"arize_space_id": "space_id"})
+
+
+class _DrainPool:
+    """Closes shed destination processors off the span-export path.
+
+    ``shutdown`` flushes over the network and is reached from ``on_end``, so closing
+    one inline would let a single unreachable tenant collector stall every other
+    tenant's spans behind it. A fixed set of workers rather than a thread per
+    processor means a tenant cycling its destination config cannot spawn threads as
+    fast as it can send requests; slow shutdowns queue behind each other.
+
+    The workers are daemons and belong to the fan-out that sheds the processors, so
+    neither an unreachable collector nor a lazily built process-wide singleton can
+    hold the proxy open on the way down.
+    """
+
+    def __init__(
+        self,
+        workers: int = _DRAIN_WORKERS,
+        pending: "queue.Queue[SpanProcessor | None] | None" = None,
+        capacity: int = _MAX_PENDING_DRAINS,
+    ) -> None:
+        self._workers: Final = workers
+        self._capacity: Final = capacity
+        self._lock: Final = threading.Lock()
+        self._closed = False
+        self._backlog = 0  # guarded by ``_lock``: submitted processors whose close has not returned
+        self._pending: Final[queue.Queue[SpanProcessor | None]] = pending if pending is not None else queue.Queue()
+        self._threads: Final = tuple(
+            threading.Thread(target=self._drain_until_closed, daemon=True, name="litellm-otel-destination-drain")
+            for _ in range(workers)
+        )
+        for worker in self._threads:
+            worker.start()
+
+    def submit(self, processor: SpanProcessor) -> None:
+        """Queue ``processor`` for closing, or hand it off once the pool is retired.
+
+        The check and the put share one lock. Reading a closed flag on its own leaves
+        room for :meth:`close` to run in between, and the processor would land behind
+        the sentinels every worker has already exited on.
+
+        Past close there is no worker left to take it, and the caller is whichever
+        thread just ended a span, so closing it inline would park that thread on a
+        network flush the shutdown deadline has already stopped waiting for. The extra
+        thread is bounded by the same close: the fan-out stops handing processors out
+        at that point, so only the ones already exporting when it happened arrive here.
+        """
+        with self._lock:
+            if not self._closed:
+                self._backlog += 1
+                self._pending.put(processor)
+                return
+        threading.Thread(
+            target=_shutdown_quietly,
+            args=(processor,),
+            daemon=True,
+            name="litellm-otel-destination-drain-straggler",
+        ).start()
+
+    def saturated(self) -> bool:
+        """Whether enough closes are outstanding that building another processor must wait.
+
+        The workers close in order and each close blocks for as long as its exporter
+        does, so a collector that stopped answering would otherwise turn every new
+        destination into one more batch thread parked behind them, for as long as the
+        tenants keep rotating. Holding the count here rather than reading the queue
+        keeps the two processors a worker is mid-close on in the total.
+        """
+        with self._lock:
+            return self._backlog >= self._capacity
+
+    def close(self, timeout: float | None = None) -> None:
+        """Retire the workers once they have closed everything already queued.
+
+        A proxy that rebuilds its telemetry builds another fan-out, so workers that
+        outlive the one that started them are two more threads per reload, forever.
+
+        ``timeout`` bounds how long the caller waits for that draining to finish. The
+        workers are daemons, so whatever is still flushing when it expires is dropped
+        by the interpreter rather than holding it open.
+        """
+        with self._lock:
+            if self._closed:
+                return
+            self._closed = True
+            for _ in range(self._workers):
+                self._pending.put(None)
+        if timeout is None:
+            return
+        deadline: Final = time.monotonic() + timeout
+        for worker in self._threads:
+            worker.join(timeout=max(0.0, deadline - time.monotonic()))
+
+    def _drain_until_closed(self) -> None:
+        while True:
+            processor: SpanProcessor | None = self._pending.get()  # rebind-ok: loop variable
+            if processor is None:
+                return
+            _shutdown_quietly(processor)
+            with self._lock:
+                self._backlog -= 1
+
+
+_NO_ATTRIBUTES: Final[Mapping[str, AttributeValue]] = MappingProxyType({})
+_DB_SYSTEM_KEYS: Final = frozenset({DB.SYSTEM_NAME, DB.SYSTEM_LEGACY})
+# Keys on a database span that describe the proxy's own datastore: its host, its
+# port, and its schema.
+_DATASTORE_ENDPOINT_KEYS: Final = frozenset({Server.ADDRESS, Server.PORT, DB.NAMESPACE})
+# A span carrying one of these describes the tenant's own call (the model call, the
+# MCP call, the guardrail), so its error text is theirs to see. Every other span is
+# the proxy's own work, whose error text names the operator's infrastructure.
+_TENANT_OWNED_KEYS: Final = frozenset({GenAI.OPERATION_NAME, MCP.METHOD_NAME, LiteLLM.GUARDRAIL_NAME})
+_PROXY_ERROR_TEXT_KEYS: Final = frozenset({Error.MESSAGE, Error.MESSAGE_LEGACY})
+# A guardrail that never answered carries the exception it raised as its response,
+# which names the operator's guardrail endpoint. The second spelling is the legacy
+# status the request-level logger still maps.
+_GUARDRAIL_UNREACHABLE_STATUSES: Final = frozenset({"guardrail_failed_to_respond", "failure"})
+# Attribute prefixes the FastAPI instrumentor uses for headers the operator opted to
+# capture (``OTEL_INSTRUMENTATION_HTTP_CAPTURE_HEADERS_SERVER_*``). The request
+# side carries the caller's bearer token verbatim.
+_CAPTURED_HEADER_PREFIXES: Final = ("http.request.header.", "http.response.header.")
+# The instrumentor stamps the request URL on the server span with its query string,
+# under the old convention and the new one, and litellm accepts a virtual key as a
+# ``?key=`` query parameter.
+_URL_KEYS: Final = frozenset({"http.url", "http.target", "url.full"})
+_URL_QUERY_KEY: Final = "url.query"
+
+
+class _TenantSpanView(ReadableSpan):
+    """A ``ReadableSpan`` view for one destination, leaving the operator's own span alone."""
+
+    def __init__(
+        self,
+        inner: ReadableSpan,
+        resource: Resource,
+        attributes: Attributes,
+        events: Sequence[Event],
+        status: Status,
+    ) -> None:
+        super().__init__(
+            name=inner.name,
+            context=inner.context,
+            parent=inner.parent,
+            resource=resource,
+            attributes=attributes,
+            events=events,
+            links=inner.links,
+            kind=inner.kind,
+            status=status,
+            start_time=inner.start_time,
+            end_time=inner.end_time,
+            instrumentation_scope=inner.instrumentation_scope,
+        )
+
+
+def _is_database_span(attributes: Mapping[str, AttributeValue]) -> bool:
+    return any(key in attributes for key in _DB_SYSTEM_KEYS)
+
+
+def _is_tenant_owned_span(attributes: Mapping[str, AttributeValue]) -> bool:
+    return any(key in attributes for key in _TENANT_OWNED_KEYS)
+
+
+def _guardrail_unreachable(attributes: Mapping[str, AttributeValue]) -> bool:
+    return attributes.get(LiteLLM.GUARDRAIL_STATUS) in _GUARDRAIL_UNREACHABLE_STATUSES
+
+
+def _tenant_visible(key: str, database: bool, owned: bool, unreachable_guardrail: bool) -> bool:
+    if key.startswith(_CAPTURED_HEADER_PREFIXES) or key in (LiteLLMError.STACK_TRACE, _URL_QUERY_KEY):
+        return False
+    if database and key in _DATASTORE_ENDPOINT_KEYS:
+        return False
+    if unreachable_guardrail and key == LiteLLM.GUARDRAIL_RESPONSE:
+        return False
+    return owned or key not in _PROXY_ERROR_TEXT_KEYS
+
+
+def _without_query(key: str, value: AttributeValue) -> AttributeValue:
+    if key not in _URL_KEYS or not isinstance(value, str):
+        return value
+    return value.partition("?")[0]
+
+
+def _same_attributes(kept: Mapping[str, AttributeValue], attributes: Mapping[str, AttributeValue]) -> bool:
+    return len(kept) == len(attributes) and all(kept[key] is value for key, value in attributes.items())
+
+
+def _without_stack_trace(event: Event) -> Event:
+    attributes: Final = event.attributes or _NO_ATTRIBUTES
+    if ExceptionEvent.STACKTRACE not in attributes:
+        return event
+    return Event(
+        name=event.name,
+        attributes=MappingProxyType(
+            {key: value for key, value in attributes.items() if key != ExceptionEvent.STACKTRACE}
+        ),
+        timestamp=event.timestamp,
+    )
+
+
+def _for_destination(span: ReadableSpan, destination: "OtelDestination") -> ReadableSpan:
+    """The view of ``span`` a tenant destination receives.
+
+    A span the tenant's own call produced keeps its error text. Every other span is
+    the proxy's own work (the request root, auth, the database), and its error text,
+    its events and its status description come off, since a Prisma failure there
+    spells out the operator's Postgres endpoint. A database span loses that endpoint
+    too, and a guardrail that failed to respond loses its response text, which is the
+    exception it raised and names the operator's guardrail endpoint. Stack traces walk
+    the operator's install and come off every span, as do the headers the operator
+    captures on the server span, whose request side holds the caller's bearer token,
+    and the query string of the request URL, which can hold the same key. The span
+    itself stays, so the tenant still gets the whole trace tree.
+    """
+    extra: Final = destination.resource_attributes
+    attributes: Final = span.attributes or _NO_ATTRIBUTES
+    database: Final = _is_database_span(attributes)
+    owned: Final = _is_tenant_owned_span(attributes)
+    unreachable: Final = _guardrail_unreachable(attributes)
+    kept: Final = MappingProxyType(
+        {
+            key: _without_query(key, value)
+            for key, value in attributes.items()
+            if _tenant_visible(key, database, owned, unreachable)
+        }
+    )
+    recorded: Final = span.events
+    events: Final = tuple(_without_stack_trace(event) for event in recorded) if owned else ()
+    unchanged: Final = owned and _same_attributes(kept, attributes) and all(a is b for a, b in zip(events, recorded))
+    if not extra and unchanged:
+        return span
+    resource: Final = span.resource.merge(Resource(extra)) if extra else span.resource
+    status: Final = span.status if owned else Status(span.status.status_code)
+    return _TenantSpanView(span, resource, kept, events, status)
+
+
+class TenantFanOutSpanProcessor(SpanProcessor):
+    """Export every finished span to each destination this request resolved.
+
+    Destinations ride a request-scoped ``ContextVar`` set during auth, so concurrent
+    requests stay isolated. The forwarded view keeps the original trace and parent
+    ids, so the tenant gets the same tree the operator would have received.
+
+    Exactly one provider carries this processor, the one published as the OTel global
+    (see :func:`attach_tenant_fan_out`). That provider is the only one every span
+    passes through: the FastAPI server span, the auth span and the post-call database
+    spans are emitted on the global, while a second v2 logger's provider sees only
+    that logger's own gen-AI span. Attaching the fan-out per logger would hand a
+    tenant a one-span trace whenever its backend is not the global one, and two
+    copies of the model call whenever it is.
+    """
+
+    def __init__(
+        self,
+        processor_factory: 'Callable[["OtelDestination"], SpanProcessor | None] | None' = None,
+        shutdown_drain_seconds: float = _SHUTDOWN_DRAIN_SECONDS,
+        operator_sinks: frozenset[_SinkKey] = frozenset(),
+        pending_drains: int = _MAX_PENDING_DRAINS,
+        drain_pool: _DrainPool | None = None,
+    ) -> None:
+        self._operator_sinks: Final = operator_sinks
+        self._drain_seconds: Final = shutdown_drain_seconds
+        self._lock: Final = threading.Condition()
+        self._closed = False  # guarded by ``_lock``: an unlocked read races the teardown it gates
+        self._build: Final = processor_factory if processor_factory is not None else _destination_processor
+        self._processors: OrderedDict[object, SpanProcessor] = OrderedDict()  # mutable-ok: bounded LRU
+        self._retired: OrderedDict[int, SpanProcessor] = OrderedDict()  # mutable-ok: drains as exports finish
+        self._exporting: dict[int, int] = {}  # mutable-ok: per-processor in-flight export count
+        self._drain: Final = drain_pool if drain_pool is not None else _DrainPool(capacity=pending_drains)
+
+    def on_start(self, span: SDKSpan, parent_context: Context | None = None) -> None:
+        return None
+
+    def on_end(self, span: ReadableSpan) -> None:
+        suppressed: Final = suppressed_backends()
+        for destination in request_destinations():
+            if self._operator_already_writes(destination, suppressed):
+                continue
+            processor = self._acquire(destination)  # rebind-ok: loop variable; pyright forbids Final in a loop
+            if processor is None:
+                continue
+            try:
+                processor.on_end(_for_destination(span, destination))
+            except Exception as exc:  # noqa: BLE001  # one destination's failure must not cost the others their span
+                verbose_logger.debug("OTel V2 fan-out: forwarding to %s failed: %s", destination.endpoint, exc)
+            finally:
+                self._release(processor)
+
+    def _operator_already_writes(self, destination: "OtelDestination", suppressed: frozenset[str]) -> bool:
+        """Whether the operator's own exporter is sending this span to the same account.
+
+        Only reachable under ``additive``, where nothing is suppressed: a team that
+        names the operator's own project would otherwise have every span written
+        there twice, once by the operator's exporter and once by the fan-out.
+        """
+        return (
+            destination.callback_name not in suppressed
+            and _sink_key(destination.endpoint, destination.headers) in self._operator_sinks
+        )
+
+    def shutdown(self) -> None:
+        """Close every destination processor, once the spans in flight have landed.
+
+        ``on_end`` runs on whichever thread ends a span and can reach this fan-out
+        while the SDK is tearing the provider down, so closing blind would drop a
+        trace mid-forward and would hand the next caller a fresh exporter nothing
+        will ever close. Refusing new work and then waiting out the in-flight ones
+        keeps both from happening. A straggler past the bound is retired instead of
+        closed: the thread still exporting it closes it through the drain as soon as
+        its export returns, so no span is dropped mid-forward.
+
+        Every close then goes to the drain rather than running here. Closing a
+        destination processor flushes it over the network and the SDK joins its own
+        worker with no timeout of its own, so one tenant collector that answers but
+        never finishes a response would otherwise hold process teardown open for as
+        long as it likes. The drain's workers are daemons, and the whole teardown
+        shares one deadline.
+        """
+        deadline: Final = time.monotonic() + self._drain_seconds
+        with self._lock:
+            self._closed = True
+            self._lock.wait_for(lambda: not self._exporting, timeout=self._drain_seconds)
+            live: Final = tuple((id(p), p) for p in (*self._processors.values(), *self._retired.values()))
+            closing: Final = tuple(p for ident, p in live if ident not in self._exporting)
+            self._processors.clear()
+            self._retired = OrderedDict(  # mutable-ok: the same bounded map, keeping only what is still exporting
+                (ident, p) for ident, p in live if ident in self._exporting
+            )
+        for processor in closing:
+            self._drain.submit(processor)
+        self._drain.close(timeout=max(0.0, deadline - time.monotonic()))
+
+    def force_flush(self, timeout_millis: int = 30000) -> bool:
+        results: Final = tuple(self._flush_one(processor, timeout_millis) for processor in self._snapshot())
+        return all(results)
+
+    def _snapshot(self) -> tuple[SpanProcessor, ...]:
+        with self._lock:
+            return (*self._processors.values(), *self._retired.values())
+
+    @staticmethod
+    def _flush_one(processor: SpanProcessor, timeout_millis: int) -> bool:
+        try:
+            return processor.force_flush(timeout_millis)
+        except Exception:  # noqa: BLE001  # one exporter's flush failure must not fail the whole flush
+            return False
+
+    def deliverable(self, destinations: Iterable["OtelDestination"]) -> tuple["OtelDestination", ...]:
+        """The subset of ``destinations`` this fan-out can actually export to.
+
+        A destination whose exporter will not build (a protocol whose package is not
+        installed, a malformed endpoint) has to be dropped before the request anchors
+        it, not when its first span ends. By then the operator's own exporter has been
+        told to hold that backend's spans back for this request, so dropping there
+        loses the span outright instead of leaving it where it would have gone with no
+        override at all.
+        """
+        return tuple(destination for destination in destinations if self._buildable(destination))
+
+    def _buildable(self, destination: "OtelDestination") -> bool:
+        """Whether a processor for ``destination`` exists or can be built right now."""
+        with self._lock:
+            if self._closed:
+                return False
+            built: Final = self._cached_or_built_locked(destination, anchored=False)
+            drained: Final = self._drainable_locked()
+            for shed in drained:
+                self._drain.submit(shed)
+            return built is not None
+
+    def _acquire(self, destination: "OtelDestination") -> SpanProcessor | None:
+        """The processor for ``destination``, marked busy until ``_release``.
+
+        The build happens under the same lock that reads the cache, so a cold cache
+        met by a burst of concurrent requests yields one exporter rather than one per
+        thread with all but the winner shed. Building an exporter opens no connection,
+        so the cost of holding the lock is a constructor, once per destination.
+        """
+        with self._lock:
+            if self._closed:
+                return None
+            processor: Final = self._cached_or_built_locked(destination, anchored=True)
+            if processor is None:
+                return None
+            self._exporting[id(processor)] = self._exporting.get(id(processor), 0) + 1
+            drained: Final = self._drainable_locked()
+            for shed in drained:
+                self._drain.submit(shed)
+            return processor
+
+    def _cached_or_built_locked(self, destination: "OtelDestination", *, anchored: bool) -> SpanProcessor | None:
+        """The cached processor for ``destination``, or a new one if the drain can take it.
+
+        Every build past the cache cap sheds one processor into the drain, so while the
+        shed ones are stuck closing against a collector that stopped answering, a
+        destination that is not yet anchored is refused rather than parked behind them:
+        ``deliverable`` then leaves its spans with the operator's exporter until the
+        drain catches up. One the request already anchored is rebuilt regardless. The
+        operator's exporter has stood down for it, so refusing here would drop the span,
+        and other tenants' auths can evict it in the meantime, with that eviction being
+        what tips the drain over. Eviction holds while the drain is saturated, so such a
+        rebuild costs the cache one entry rather than shedding another processor, and
+        the total stays at one per destination in flight.
+        """
+        key: Final = destination.cache_key()
+        if (cached := self._processors.get(key)) is not None:
+            self._processors.move_to_end(key)
+            self._retire_overflow_locked()
+            return cached
+        if not anchored and self._drain.saturated():
+            verbose_logger.debug("OTel V2 fan-out: drain saturated, not building for %s", destination.endpoint)
+            return None
+        return self._build_locked(destination, key)
+
+    def _build_locked(self, destination: "OtelDestination", key: object) -> SpanProcessor | None:
+        built: Final = self._build(destination)
+        if built is None:
+            return None
+        self._processors[key] = built
+        self._retire_overflow_locked()
+        return built
+
+    def _release(self, processor: SpanProcessor) -> None:
+        with self._lock:
+            remaining: Final = self._exporting.get(id(processor), 1) - 1
+            if remaining > 0:
+                self._exporting[id(processor)] = remaining
+            else:
+                self._exporting.pop(id(processor), None)
+                if not self._exporting:
+                    self._lock.notify_all()
+            drained: Final = self._drainable_locked()
+            for retired in drained:
+                self._drain.submit(retired)
+
+    def _retire_overflow_locked(self) -> None:
+        """Move the LRU processor out of the cache once it is past the cap, drain permitting.
+
+        Eviction is what feeds the drain, and a destination a request already anchored
+        is rebuilt on its next span, which would shed another one. While the shed ones
+        are stuck closing against a collector that stopped answering, evicting would
+        churn the cache at one more processor, and one more batch thread, per span.
+        Holding above the cap instead keeps the total at one processor per destination
+        in flight, since ``deliverable`` anchors no new destination while the drain is
+        saturated. Once it has room again, every hit and build trims one entry.
+        """
+        if len(self._processors) <= _MAX_CACHED_DESTINATION_PROCESSORS or self._drain.saturated():
+            return
+        _, evicted = self._processors.popitem(last=False)
+        self._retired[id(evicted)] = evicted
+
+    def _drainable_locked(self) -> tuple[SpanProcessor, ...]:
+        """Retired processors no thread is exporting through, removed from the list.
+
+        ``on_end`` holds a processor across an export, so closing an evicted one there
+        drops the span it is holding. A retiree is out of the cache and can never be
+        handed out again, so once its export count reaches zero it stays there.
+        """
+        idle: Final = tuple(key for key in self._retired if self._exporting.get(key, 0) == 0)
+        return tuple(self._retired.pop(key) for key in idle)
+
+
+def _destination_processor(destination: "OtelDestination") -> SpanProcessor | None:
+    """A batching OTLP processor aimed at ``destination``, or ``None`` if unbuildable.
+
+    A protocol that resolves to a headerless exporter is unbuildable too: the
+    console fallback would swallow the tenant's credentials and print its spans to
+    the proxy's stdout while the operator's exporter stands down for them.
+    """
+    kind: Final = destination.protocol or "otlp_http"
+    if exporter_transport(kind) == "headerless":
+        verbose_logger.debug("OTel V2 fan-out: no OTLP transport for protocol %r at %s", kind, destination.endpoint)
+        return None
+    try:
+        spec: Final = ExporterSpec(
+            kind=kind,
+            endpoint=destination.endpoint,
+            headers=destination.header_string(),
+            owner=None,
+        )
+        return _processor_for(_exporter_from_spec(spec), use_simple=False)
+    except Exception as exc:  # noqa: BLE001  # a malformed destination must not break the request or the other destinations
+        verbose_logger.debug("OTel V2 fan-out: no processor for %s: %s", destination.endpoint, exc)
+        return None
+
+
+def _shutdown_quietly(processor: SpanProcessor) -> None:
+    try:
+        processor.shutdown()
+    except Exception as exc:  # noqa: BLE001  # defensive: shedding a spare processor must not raise
+        verbose_logger.debug("OTel V2 fan-out: discarding processor failed: %s", exc)
+
+
+class _OverriddenBackendFilter(SpanProcessor):
+    """Hold a span back from ``owner``'s operator-level exporter when the request
+    pointed ``owner`` at a tenant's own account.
+
+    Wrapping is the only place this works: ``SynchronousMultiSpanProcessor.on_end``
+    ignores return values, so a sibling processor can never veto the export.
+
+    Under ``additive`` mode nothing is suppressed, so the wrapper passes every span
+    straight through and the operator keeps its copy.
+    """
+
+    def __init__(self, inner: SpanProcessor, owner: str) -> None:
+        self._inner: Final = inner
+        self._owner: Final = owner
+
+    def on_start(self, span: SDKSpan, parent_context: Context | None = None) -> None:
+        self._inner.on_start(span, parent_context)
+
+    def on_end(self, span: ReadableSpan) -> None:
+        if self._owner in suppressed_backends():
+            return
+        self._inner.on_end(span)
+
+    def shutdown(self) -> None:
+        self._inner.shutdown()
+
+    def force_flush(self, timeout_millis: int = 30000) -> bool:
+        return self._inner.force_flush(timeout_millis)
+
+
 def build_span_exporter(config: OpenTelemetryV2Config) -> SpanExporter:
     """Build a single exporter from the top-level config fields.
 
@@ -444,6 +1024,7 @@ def build_tracer_provider(
     exporter: SpanExporter | None = None,
     baggage_processor: SpanProcessor | None = None,
     use_simple_processor: bool | None = None,
+    tenant_overrides: bool = False,
 ) -> TracerProvider:
     """Build the shared :class:`TracerProvider`.
 
@@ -452,6 +1033,13 @@ def build_tracer_provider(
     ``config.exporters`` entry — this is what fans spans out to multiple
     backends. ``exporter`` and ``use_simple_processor`` are explicit overrides:
     pass a single exporter to attach exactly that one (used by tests).
+
+    ``tenant_overrides`` wraps each owned exporter so a request that pointed that
+    backend at a key's or team's own account skips it. Every v2 logger's provider
+    wants it, since any of them may own the overridden backend; delivering to the
+    tenant is a separate job, done once by :func:`attach_tenant_fan_out`. The
+    per-tenant providers this same function builds must leave it off, or they would
+    filter out the very spans they exist to carry.
     """
     provider: Final = TracerProvider(resource=build_resource(config))
     if baggage_processor is None:
@@ -468,15 +1056,107 @@ def build_tracer_provider(
         if spec.requires_headers and not spec.headers:
             continue
         exp = _exporter_from_spec(spec)
+        processor = _processor_for(
+            exp,
+            (spec.use_simple_processor if spec.use_simple_processor is not None else use_simple_processor),
+        )
+        owner = spec.owner.value if spec.owner is not None else None
         provider.add_span_processor(
-            _processor_for(
-                exp,
-                (spec.use_simple_processor if spec.use_simple_processor is not None else use_simple_processor),
-            )
+            _OverriddenBackendFilter(processor, owner) if tenant_overrides and owner is not None else processor
         )
     return provider
 
 
+_FAN_OUT_ATTACH_LOCK: Final = threading.Lock()
+
+
+def attach_tenant_fan_out(provider: TracerProvider, *configs: OpenTelemetryV2Config) -> None:
+    """Give ``provider`` the fan-out that delivers spans to key/team destinations.
+
+    Called on the one provider published as the OTel global, and idempotent so a
+    second publish (a test, a re-initialized proxy) cannot double-export. Concurrent
+    first calls (requests racing to anchor before any publish) serialize on one lock
+    so exactly one fan-out lands. ``configs`` name the operator's own exporters, one
+    config per v2 logger since each keeps its own provider and still writes its
+    account, so an additive destination pointing at any of them is delivered once
+    rather than twice.
+    """
+    with _FAN_OUT_ATTACH_LOCK:
+        if any(isinstance(processor, TenantFanOutSpanProcessor) for processor in _attached_processors(provider)):
+            return
+        provider.add_span_processor(TenantFanOutSpanProcessor(operator_sinks=operator_sink_keys(*configs)))
+
+
+def deliverable_destinations(
+    destinations: Iterable["OtelDestination"],
+    provider: trace.TracerProvider | None = None,
+) -> tuple["OtelDestination", ...]:
+    """The destinations a request can anchor, given what is published to carry them.
+
+    Anchoring a destination is what tells the operator's own exporter to stand down
+    for that backend, so one nothing can deliver has to be dropped here: with no
+    fan-out attached, or with an exporter that will not build, the request keeps
+    exactly the routing it would have had without any override.
+    """
+    fan_out: Final = next(
+        (
+            processor
+            for processor in _attached_processors(provider if provider is not None else trace.get_tracer_provider())
+            if isinstance(processor, TenantFanOutSpanProcessor)
+        ),
+        None,
+    )
+    return fan_out.deliverable(destinations) if fan_out is not None else ()
+
+
+def operator_sink_keys(*configs: OpenTelemetryV2Config) -> frozenset[_SinkKey]:
+    """The accounts the operator's own exporters write to, in destination terms.
+
+    Every v2 logger's config counts, since each logger exports through its own
+    provider. An exporter with no endpoint of its own resolves one from the
+    environment at export time, so it has no comparable identity and is left out,
+    and so is one that never reaches the wire: a console kind ignores the endpoint,
+    and a header-gated spec with no credentials is skipped when the provider is built.
+    """
+    return frozenset(
+        key
+        for config in configs
+        for spec in config.exporters
+        if _exports_to_the_wire(spec) and (key := _sink_key(spec.endpoint, parse_headers(spec.headers))) is not None
+    )
+
+
+def _exports_to_the_wire(spec: ExporterSpec) -> bool:
+    """Whether ``build_tracer_provider`` gives ``spec`` an exporter that sends OTLP."""
+    return exporter_transport(spec.kind) != "headerless" and not (spec.requires_headers and not spec.headers)
+
+
+def _sink_key(endpoint: str | None, headers: Mapping[str, str]) -> "_SinkKey | None":
+    """The account an exporter writes to, or ``None`` when it has no fixed one.
+
+    Normalized on the three counts that make one account look like two: the operator's
+    spec carries the signal path a tenant destination leaves for the exporter to
+    append, header names survive one round trip lowercased and the other not, and one
+    credential answers to more than one name (see :data:`_CREDENTIAL_ALIASES`).
+    """
+    normalized: Final = _otlp_traces_endpoint(endpoint)
+    if normalized is None:
+        return None
+    return (normalized, tuple(sorted((_credential_name(name), value) for name, value in headers.items())))
+
+
+def _credential_name(header: str) -> str:
+    """The credential a header carries, under whichever name the backend spells it."""
+    normalized: Final = header.strip().lower().replace("-", "_")
+    return _CREDENTIAL_ALIASES.get(normalized, normalized)
+
+
+def _attached_processors(provider: trace.TracerProvider) -> "tuple[SpanProcessor, ...]":
+    """The processors already on ``provider``, or empty when the SDK hides them."""
+    multi: Final = getattr(provider, "_active_span_processor", None)
+    return tuple(getattr(multi, "_span_processors", ()))
+
+
 def get_tracer(provider: TracerProvider, name: str = "litellm") -> Tracer:
     # Stamp the instrumentation scope with the LiteLLM package version so every
     # emitted span carries a deterministic ``scope.version`` (the standard OTel

litellm/integrations/otel/plumbing/routing.py

@@ -25,6 +25,7 @@ from opentelemetry.trace import Tracer
 from litellm._logging import verbose_logger
 from litellm.constants import OTEL_SERVICE_NAME_METADATA_KEYS
 from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.plumbing.context import destination_backends
 from litellm.integrations.otel.plumbing.providers import (
     build_tracer_provider,
     exporter_transport,
@@ -231,10 +232,21 @@ class TenantTracerCache:
         concurrent overflow eviction can't shut it down between selection and
         the caller's span start. The caller must ``release`` it exactly once.
         """
+        # A backend with a destination is delivered by the fan-out processor, which
+        # carries the whole trace and already carries this tenant's credentials and
+        # service name. Routing here too would detach this span onto a second provider,
+        # so the tenant would get the request tree plus a stray one-span trace.
+        if self._callback_name is not None and self._callback_name in destination_backends():
+            return TenantRoute(tracer=default, detached=False)
         credential_headers: Final = self._credential_headers(dynamic_params)
         project_headers: Final = self._project_headers(auth_metadata)
         service_name: Final = tenant_service_name(auth_metadata)
-        if not credential_headers and not project_headers and service_name is None:
+        tenant_account: Final = bool(credential_headers) or bool(project_headers)
+        # A service name on its own only relabels the operator's own backend, so moving
+        # the span to a second provider for it while some other backend has a
+        # destination would drop the model call out of the trace the fan-out delivers.
+        # The destination stamps the same service name itself.
+        if not tenant_account and (service_name is None or destination_backends()):
             return TenantRoute(tracer=default, detached=False)
         # A fixed per-integration region endpoint (New Relic us/eu), never a
         # caller-supplied host; ``None`` keeps the preset's own endpoint.
@@ -255,7 +267,7 @@ class TenantTracerCache:
             _shutdown_provider(evicted)
         return TenantRoute(
             tracer=get_tracer(provider, self._tracer_name),
-            detached=bool(project_headers) or bool(credential_headers),
+            detached=tenant_account,
             provider=provider,
         )
 

litellm/integrations/otel/presets/agentops.py

@@ -39,6 +39,7 @@ class _AgentOpsSettings(BaseSettings):
 def agentops_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
     """Build the AgentOps config without any network I/O.
 

litellm/integrations/otel/presets/arize.py

@@ -26,10 +26,12 @@ class _ArizeSettings(BaseSettings):
 def arize_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
+    base: Final = config_overrides or OpenTelemetryV2Config()
+    mappers: Final = ensure_mappers(base.mapper_names, "openinference")
     arize_cfg: Final = _V1ArizeLogger.get_arize_config()
     headers: Final = _arize_headers(arize_cfg)
-    base: Final = config_overrides or OpenTelemetryV2Config()
     return base.model_copy(
         update={
             "exporters": [
@@ -41,7 +43,7 @@ def arize_preset(
                     owner=ExporterOwner.ARIZE_AX,
                 ),
             ],
-            "mapper_names": ensure_mappers(base.mapper_names, "openinference"),
+            "mapper_names": mappers,
             "resource_attributes": {
                 **base.resource_attributes,
                 **({"model_id": arize_cfg.project_name} if arize_cfg.project_name else {}),

litellm/integrations/otel/presets/base.py

@@ -18,6 +18,18 @@ class Preset(Protocol):
 
     ``config_overrides`` lets one preset layer onto another's config (or onto
     test-supplied defaults); the factory calls presets with no arguments.
+
+    ``allow_missing_credentials`` lets a credential-mandatory backend (langfuse and
+    weave) degrade to an exporter-less, mapper-only config instead of raising when the
+    operator set no env credentials of their own. That is a real
+    deployment: every team brings its own account and the operator keeps none, and
+    without it the whole V2 path silently falls back to the legacy integration, so
+    no team destination is ever reached. Credential-optional backends ignore it.
     """
 
-    def __call__(self, *, config_overrides: OpenTelemetryV2Config | None = None) -> OpenTelemetryV2Config: ...
+    def __call__(
+        self,
+        *,
+        config_overrides: OpenTelemetryV2Config | None = None,
+        allow_missing_credentials: bool = False,
+    ) -> OpenTelemetryV2Config: ...

litellm/integrations/otel/presets/destinations.py

@@ -0,0 +1,152 @@
+"""Map a key's or team's callback vars to the OTLP destination its traces export to.
+
+Header building is delegated to each preset's existing ``*_dynamic_headers`` builder,
+so a destination authenticates exactly the way the per-request tracer route already
+did; only the endpoint and transport need a per-backend rule.
+"""
+
+import os
+from collections.abc import Callable, Mapping
+from functools import lru_cache
+from types import MappingProxyType
+from typing import Final
+
+import litellm
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.destination import OtelDestination
+from litellm.litellm_core_utils.url_utils import is_url_destination_allowed_by_host
+from litellm.types.utils import StandardCallbackDynamicParams
+
+#: An endpoint plus the OTLP transport to reach it with, or ``None`` when the backend
+#: names no destination. The transport is ``None`` where the backend has only one.
+_Destination = tuple[str, str | None]
+
+
+@lru_cache(maxsize=128)
+def _warn_host_not_allowlisted(host: str) -> None:
+    """Cached so one misconfigured team logs once rather than once per request."""
+    verbose_logger.warning(
+        "OTel V2: not exporting to key/team Langfuse host '%s'. Add it to "
+        "litellm_settings.provider_url_destination_allowed_hosts to permit it",
+        host,
+    )
+
+
+def _langfuse_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
+    """The tenant's own Langfuse host, else the operator's, else Langfuse US cloud.
+
+    A host the tenant named has to be allowlisted by the operator, the same way a
+    URL-valued ``model`` is: anyone who can mint a key can write it, and it becomes an
+    endpoint the proxy posts the request's whole trace to, carrying the tenant's own
+    credentials. The operator's own ``LANGFUSE_HOST`` is not checked, since an internal
+    collector there is a deployment choice.
+    """
+    from litellm.integrations.langfuse.langfuse_otel import (
+        LANGFUSE_CLOUD_US_ENDPOINT,
+        LangfuseOtelLogger,
+    )
+
+    tenant_host: Final = params.get("langfuse_host") or None
+    host: Final = tenant_host or LangfuseOtelLogger._get_langfuse_otel_host()  # pyright: ignore[reportPrivateUsage]  # reuse the backend's own env host resolver rather than duplicating it
+    if not host:
+        return (LANGFUSE_CLOUD_US_ENDPOINT, None)
+    normalized: Final = host if host.startswith("http") else f"https://{host}"
+    endpoint: Final = f"{normalized.rstrip('/')}/api/public/otel"
+    if tenant_host is None:
+        return (endpoint, None)
+    if not is_url_destination_allowed_by_host(endpoint, litellm.provider_url_destination_allowed_hosts):
+        _warn_host_not_allowlisted(host)
+        return None
+    return (endpoint, None)
+
+
+def _arize_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
+    from litellm.integrations.arize.arize import ArizeLogger
+
+    config: Final = ArizeLogger.get_arize_config()
+    return (config.endpoint, config.protocol)
+
+
+def _weave_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
+    from litellm.integrations.weave.weave_otel import weave_otel_endpoint
+
+    return (weave_otel_endpoint(os.environ.get("WANDB_HOST")), None)
+
+
+def _newrelic_destination(params: StandardCallbackDynamicParams) -> "_Destination | None":
+    from litellm.integrations.otel.presets.newrelic import newrelic_dynamic_endpoint
+
+    endpoint: Final = newrelic_dynamic_endpoint(params)
+    return (endpoint, None) if endpoint else None
+
+
+#: Callback name -> destination resolver. A backend is destination-capable exactly
+#: when it appears here AND in ``DYNAMIC_HEADERS_BY_CALLBACK``: without a header
+#: builder the destination would carry no tenant credentials, and the exporter
+#: would post the tenant's traffic to the operator's account.
+_DESTINATION_BY_CALLBACK: Final[Mapping[str, Callable[[StandardCallbackDynamicParams], "_Destination | None"]]] = (
+    MappingProxyType(
+        {
+            "langfuse_otel": _langfuse_destination,
+            "arize": _arize_destination,
+            "weave_otel": _weave_destination,
+            "newrelic": _newrelic_destination,
+        }
+    )
+)
+
+#: Headers a destination must carry to authenticate. Several dynamic-header builders
+#: gate each credential independently, so a half-configured backend yields a non-empty
+#: but unusable header set; accepting it would suppress the operator's own exporter and
+#: send the request's whole trace where it cannot be stored.
+_REQUIRED_HEADERS_BY_CALLBACK: Final[Mapping[str, frozenset[str]]] = MappingProxyType(
+    {
+        "langfuse_otel": frozenset({"Authorization"}),
+        "arize": frozenset({"arize-space-id", "api_key"}),
+        "weave_otel": frozenset({"Authorization", "project_id"}),
+        "newrelic": frozenset({"api-key"}),
+    }
+)
+
+_NO_ATTRS: Final[Mapping[str, str]] = MappingProxyType({})
+
+
+def destination_capable_backends() -> frozenset[str]:
+    """Backends a key or team can point at its own account."""
+    from litellm.integrations.otel.presets import DYNAMIC_HEADERS_BY_CALLBACK
+
+    return frozenset(_DESTINATION_BY_CALLBACK) & frozenset(DYNAMIC_HEADERS_BY_CALLBACK)
+
+
+def destination_for(
+    callback_name: str,
+    params: StandardCallbackDynamicParams,
+    service_name: str | None = None,
+) -> OtelDestination | None:
+    """The destination ``params`` names for ``callback_name``, or ``None``.
+
+    ``None`` means the caller configured nothing usable for this backend, so the
+    request keeps the operator's global exporters. ``service_name`` is the key's or
+    team's ``otel_service_name``, which the per-request tracer route applies when the
+    backend is not overridden and the destination has to apply once it is.
+    """
+    from litellm.integrations.otel.presets import DYNAMIC_HEADERS_BY_CALLBACK
+
+    header_builder: Final = DYNAMIC_HEADERS_BY_CALLBACK.get(callback_name)
+    destination_builder: Final = _DESTINATION_BY_CALLBACK.get(callback_name)
+    if header_builder is None or destination_builder is None:
+        return None
+    headers: Final = header_builder(params)
+    if not headers or not _REQUIRED_HEADERS_BY_CALLBACK[callback_name] <= frozenset(headers):
+        return None
+    resolved: Final = destination_builder(params)
+    if resolved is None:
+        return None
+    endpoint, protocol = resolved
+    return OtelDestination(
+        endpoint=endpoint,
+        headers=MappingProxyType(dict(headers)),  # mutable-ok: MappingProxyType needs a concrete mapping to wrap
+        resource_attributes=MappingProxyType({"service.name": service_name}) if service_name else _NO_ATTRS,
+        callback_name=callback_name,
+        protocol=protocol,
+    )

litellm/integrations/otel/presets/langfuse.py

@@ -10,17 +10,32 @@ from litellm.integrations.otel.model.config import (
     ExporterSpec,
     OpenTelemetryV2Config,
 )
-from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.integrations.otel.presets.utils import (
+    credential_gated_exporters,
+    ensure_mappers,
+)
 from litellm.types.utils import StandardCallbackDynamicParams
 
 
 def langfuse_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
-    cfg: Final = _V1Langfuse.get_langfuse_otel_config()
-    kind: Final = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
     base: Final = config_overrides or OpenTelemetryV2Config()
+    mappers: Final = ensure_mappers(base.mapper_names, "langfuse")
+    try:
+        cfg: Final = _V1Langfuse.get_langfuse_otel_config()
+    except Exception:
+        if not allow_missing_credentials:
+            raise
+        return base.model_copy(
+            update={  # mutable-ok: pydantic model_copy takes a plain update mapping
+                "exporters": credential_gated_exporters(base.exporters, ExporterOwner.LANGFUSE_OTEL),
+                "mapper_names": mappers,
+            }
+        )
+    kind: Final = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
     return base.model_copy(
         update={
             "exporters": [
@@ -32,7 +47,7 @@ def langfuse_preset(
                     owner=ExporterOwner.LANGFUSE_OTEL,
                 ),
             ],
-            "mapper_names": ensure_mappers(base.mapper_names, "langfuse"),
+            "mapper_names": mappers,
         }
     )
 

litellm/integrations/otel/presets/langtrace.py

@@ -9,6 +9,7 @@ from litellm.integrations.otel.presets.utils import ensure_mappers
 def langtrace_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
     """Compose the Langtrace mapper on top of the customer's OTLP destination.
 

litellm/integrations/otel/presets/levo.py

@@ -13,6 +13,7 @@ from litellm.integrations.otel.model.config import (
 def levo_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
     cfg: Final = _V1Levo.get_levo_config()
     base: Final = config_overrides or OpenTelemetryV2Config()

litellm/integrations/otel/presets/newrelic.py

@@ -44,6 +44,7 @@ class _NewRelicSettings(BaseSettings):
 def newrelic_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
     settings: Final = _NewRelicSettings()
     base: Final = config_overrides or OpenTelemetryV2Config()

litellm/integrations/otel/presets/phoenix.py

@@ -60,6 +60,7 @@ def phoenix_project_headers(auth_metadata: Mapping[str, str] | None) -> Mapping[
 def phoenix_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
     cfg: Final = _V1Phoenix.get_arize_phoenix_config()
     headers: Final = cfg.otlp_auth_headers if hasattr(cfg, "otlp_auth_headers") else None

litellm/integrations/otel/presets/utils.py

@@ -3,6 +3,8 @@
 from collections.abc import Iterable
 from typing import Final
 
+from litellm.integrations.otel.model.config import ExporterOwner, ExporterSpec
+
 
 def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
     """Return ``mapper_names`` with each of ``names`` appended if not already present.
@@ -15,3 +17,32 @@ def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
         if name not in result:
             result.append(name)
     return result
+
+
+def credential_gated_exporters(
+    exporters: "Iterable[ExporterSpec]", owner: "ExporterOwner"
+) -> "tuple[ExporterSpec, ...]":
+    """``exporters`` with the operator's destination replaced by a header-gated one.
+
+    Used when a credential-mandatory backend is asked to build without the operator's
+    own credentials, so only key/team destinations receive spans. Two things have to
+    happen for that to mean "export nowhere": the placeholder console spec that
+    ``OpenTelemetryV2Config`` folds in for an empty exporter list is dropped, or every
+    span would be printed to stdout, and the gated spec keeps the owner so the
+    override filter still recognises which backend this provider speaks for.
+    """
+    return (
+        *(spec for spec in exporters if not is_unconfigured_placeholder(spec)),
+        ExporterSpec(owner=owner, requires_headers=True),
+    )
+
+
+def is_unconfigured_placeholder(spec: "ExporterSpec") -> bool:
+    """Whether ``spec`` is the one ``_normalize`` folds in when nothing was configured.
+
+    No field set is what says the operator asked for nothing: an exporter they did
+    configure survives, even ``OTEL_EXPORTER=console`` whose value matches the default,
+    and so does the gated spec this module appends, which would otherwise eat itself
+    when one preset layers onto another.
+    """
+    return not spec.model_fields_set

litellm/integrations/otel/presets/weave.py

@@ -7,7 +7,10 @@ from litellm.integrations.otel.model.config import (
     ExporterSpec,
     OpenTelemetryV2Config,
 )
-from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.integrations.otel.presets.utils import (
+    credential_gated_exporters,
+    ensure_mappers,
+)
 from litellm.integrations.weave.weave_otel import (
     _get_weave_authorization_header,
     get_weave_otel_config,
@@ -18,9 +21,21 @@ from litellm.types.utils import StandardCallbackDynamicParams
 def weave_preset(
     *,
     config_overrides: OpenTelemetryV2Config | None = None,
+    allow_missing_credentials: bool = False,
 ) -> OpenTelemetryV2Config:
-    weave_cfg: Final = get_weave_otel_config()
     base: Final = config_overrides or OpenTelemetryV2Config()
+    mappers: Final = ensure_mappers(base.mapper_names, "openinference", "weave")
+    try:
+        weave_cfg: Final = get_weave_otel_config()
+    except Exception:
+        if not allow_missing_credentials:
+            raise
+        return base.model_copy(
+            update={  # mutable-ok: pydantic model_copy takes a plain update mapping
+                "exporters": credential_gated_exporters(base.exporters, ExporterOwner.WEAVE_OTEL),
+                "mapper_names": mappers,
+            }
+        )
     return base.model_copy(
         update={
             "exporters": [
@@ -33,7 +48,7 @@ def weave_preset(
                 ),
             ],
             # Weave consumes OpenInference + a small Weave-specific overlay.
-            "mapper_names": ensure_mappers(base.mapper_names, "openinference", "weave"),
+            "mapper_names": mappers,
         }
     )
 

litellm/integrations/weave/weave_otel.py

@@ -117,6 +117,14 @@ def _get_weave_authorization_header(api_key: str) -> str:
     return f"Basic {auth_header}"
 
 
+def weave_otel_endpoint(host: str | None) -> str:
+    """The OTLP traces endpoint for a self-managed ``host``, else Weave cloud."""
+    if not host:
+        return WEAVE_BASE_URL + WEAVE_OTEL_ENDPOINT
+    normalized: Final = host if host.startswith("http") else f"https://{host}"
+    return normalized.rstrip("/") + WEAVE_OTEL_ENDPOINT
+
+
 def get_weave_otel_config() -> WeaveOtelConfig:
     """
     Retrieves the Weave OpenTelemetry configuration based on environment variables.
@@ -134,7 +142,6 @@ def get_weave_otel_config() -> WeaveOtelConfig:
     """
     api_key: Final = os.getenv("WANDB_API_KEY")
     project_id: Final = os.getenv("WANDB_PROJECT_ID")
-    host = os.getenv("WANDB_HOST")
 
     if not api_key:
         raise ValueError("WANDB_API_KEY must be set for Weave OpenTelemetry integration.")
@@ -144,15 +151,8 @@ def get_weave_otel_config() -> WeaveOtelConfig:
             "WANDB_PROJECT_ID must be set for Weave OpenTelemetry integration. Format: <entity>/<project_name>"
         )
 
-    if host:
-        if not host.startswith("http"):
-            host = "https://" + host
-        # Self-managed instances use a different path
-        endpoint = host.rstrip("/") + WEAVE_OTEL_ENDPOINT
-        verbose_logger.debug("Using Weave OTEL endpoint from host: %s", endpoint)
-    else:
-        endpoint = WEAVE_BASE_URL + WEAVE_OTEL_ENDPOINT
-        verbose_logger.debug("Using Weave cloud endpoint: %s", endpoint)
+    endpoint: Final = weave_otel_endpoint(os.getenv("WANDB_HOST"))
+    verbose_logger.debug("Using Weave OTEL endpoint: %s", endpoint)
 
     # Weave uses Basic auth with format: api:<WANDB_API_KEY>
     auth_header: Final = _get_weave_authorization_header(api_key=api_key)

litellm/litellm_core_utils/litellm_logging.py

@@ -202,6 +202,7 @@ if TYPE_CHECKING:
     from mcp.types import EmbeddedResource, ImageContent, TextContent
 
     from litellm.integrations.otel.logger import OpenTelemetryV2
+    from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
     from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
 try:
     from litellm_enterprise.enterprise_callbacks.callback_controls import (
@@ -4850,31 +4851,83 @@ def _maybe_construct_otel_v2(callback_name: str, _in_memory_loggers: list[Custom
 
     Returns ``None`` when V2 is off OR when there's no preset registered for
     ``callback_name`` — callers should then fall through to the legacy path.
+
+    A preset that needs operator credentials it cannot find is allowed to build
+    only when this request has a key/team destination for that backend and another
+    V2 logger is already registered to carry the fan-out. The resulting logger keeps
+    only its credential-gated exporter, while the registered logger owns operator
+    delivery. Without that carrier, a preset that raises or that ends up with nothing
+    but its gated exporter and the default console placeholder returns ``None``, so the
+    caller falls through to the legacy path exactly as before V2 landed.
     """
     from litellm.integrations.otel.model.config import is_otel_v2_enabled
 
     if not is_otel_v2_enabled():
         return None
     from litellm.integrations.otel.logger import OpenTelemetryV2, build_otel_v2_logger
+    from litellm.integrations.otel.plumbing.context import destination_backends
     from litellm.integrations.otel.presets import PRESET_BY_CALLBACK
 
     preset_fn: Final = PRESET_BY_CALLBACK.get(callback_name)
     if preset_fn is None:
         return None
+    serves_a_destination: Final = callback_name in destination_backends()
+    has_v2_logger: Final = any(isinstance(callback, OpenTelemetryV2) for callback in _in_memory_loggers)
+    carried: Final = serves_a_destination and has_v2_logger
     for callback in _in_memory_loggers:
-        if isinstance(callback, OpenTelemetryV2) and getattr(callback, "callback_name", None) == callback_name:
+        if (
+            isinstance(callback, OpenTelemetryV2)
+            and getattr(callback, "callback_name", None) == callback_name
+            and (serves_a_destination or not _exports_nowhere(callback.config))
+        ):
             return callback
     try:
-        config: Final = preset_fn()
+        built: Final = preset_fn(allow_missing_credentials=carried)
     except Exception:
         # If env vars are missing or the preset raises, defer to the legacy path
         # so customers get the same error story they had before V2 landed.
         return None
+    gated: Final = _is_credential_gated(built)
+    if gated and not carried and not _has_operator_exporter(built):
+        return None
+    config: Final = _only_the_gated_exporter(built) if gated and carried else built
+    if _exports_nowhere(config):
+        verbose_logger.warning(
+            "OTel V2: no operator credentials for '%s'; only key/team destinations will receive its traces",
+            callback_name,
+        )
     v2_logger: Final = build_otel_v2_logger(config=config, callback_name=callback_name)
     _in_memory_loggers.append(v2_logger)
     return v2_logger
 
 
+def _exports_nowhere(config: "OpenTelemetryV2Config") -> bool:
+    """Whether every exporter in ``config`` is waiting on credentials it never got."""
+    return all(_is_gated(spec) for spec in config.exporters)
+
+
+def _is_credential_gated(config: "OpenTelemetryV2Config") -> bool:
+    """Whether the preset built without the operator's own credentials for its backend."""
+    return any(_is_gated(spec) for spec in config.exporters)
+
+
+def _has_operator_exporter(config: "OpenTelemetryV2Config") -> bool:
+    """Whether the operator configured somewhere real to export, beyond the default console placeholder."""
+    from litellm.integrations.otel.presets.utils import is_unconfigured_placeholder
+
+    return any(not _is_gated(spec) and not is_unconfigured_placeholder(spec) for spec in config.exporters)
+
+
+def _only_the_gated_exporter(config: "OpenTelemetryV2Config") -> "OpenTelemetryV2Config":
+    return config.model_copy(
+        update={"exporters": [spec for spec in config.exporters if _is_gated(spec)]}  # mutable-ok: model_copy update
+    )
+
+
+def _is_gated(spec: "ExporterSpec") -> bool:
+    return spec.requires_headers and not spec.headers
+
+
 def _maybe_auto_initialize_arize_phoenix(_in_memory_loggers: list[CustomLogger]) -> None:
     """
     Auto-initialize ArizePhoenixLogger when Phoenix env vars are detected.

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -780,6 +780,7 @@ class PromptTokensDetailsResult(TypedDict):
     image_count: int
     video_length_seconds: float
     audio_length_seconds: float
+    query_count: int
 
 
 def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
@@ -828,6 +829,7 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
         )
         or 0.0
     )
+    query_count: Final = _coerce_token_count(getattr(usage.prompt_tokens_details, "query_count", 0))
 
     return PromptTokensDetailsResult(
         cache_hit_tokens=cache_hit_tokens,
@@ -841,6 +843,7 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
         image_count=image_count,
         video_length_seconds=float(video_length_seconds),
         audio_length_seconds=float(audio_length_seconds),
+        query_count=query_count,
     )
 
 
@@ -978,6 +981,11 @@ def _calculate_input_cost(
             prompt_tokens_details["audio_length_seconds"],
         )
 
+    if prompt_tokens_details["query_count"]:
+        prompt_cost += calculate_cost_component(
+            model_info, "input_cost_per_query", prompt_tokens_details["query_count"]
+        )
+
     return prompt_cost
 
 
@@ -1149,6 +1157,7 @@ def generic_cost_per_token(
         image_count=0,
         video_length_seconds=0.0,
         audio_length_seconds=0.0,
+        query_count=0,
     )
     if usage.prompt_tokens_details:
         prompt_tokens_details = parse_prompt_tokens_details(usage)

litellm/litellm_core_utils/prompt_templates/image_handling.py

@@ -183,6 +183,7 @@ class _RemoteSource:
 class RemoteMedia:
     url: str
     fields: Mapping[str, object]
+    part_type: str
 
 
 _NO_FIELDS: Final[Mapping[str, object]] = MappingProxyType({})
@@ -192,6 +193,10 @@ def inline_every_remote_url(_media: RemoteMedia) -> bool:
     return True
 
 
+def inline_remote_image_urls(media: RemoteMedia) -> bool:
+    return media.part_type == "image_url"
+
+
 def _parse_remote_image(fields: Mapping[str, object]) -> _RemoteImage | None:
     if fields.get("type") != "image_url":
         return None
@@ -223,11 +228,11 @@ def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | _RemoteSour
 def _remote_media(remote: _RemoteImage | _RemoteFile | _RemoteSource) -> RemoteMedia:
     match remote:
         case _RemoteImage(_, image_url, url):
-            return RemoteMedia(url, image_url if image_url is not None else _NO_FIELDS)
+            return RemoteMedia(url, image_url if image_url is not None else _NO_FIELDS, "image_url")
         case _RemoteFile(_, file, url):
-            return RemoteMedia(url, file)
-        case _RemoteSource(_, source, url):
-            return RemoteMedia(url, source)
+            return RemoteMedia(url, file, "file")
+        case _RemoteSource(part, source, url):
+            return RemoteMedia(url, source, str(part.get("type")))
 
 
 _PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"})

litellm/llms/anthropic/chat/handler.py

@@ -368,27 +368,92 @@ class AnthropicChatCompletion(BaseLLM):
         if config is None:
             raise ValueError(f"Provider config not found for model: {model} and provider: {custom_llm_provider}")
 
-        def build_request() -> tuple[dict, dict]:  # mutable-ok: rewritten in place downstream
-            """Translate the request the Python way, returning `(headers, data)`.
+        transform_params: Final = {**optional_params, "is_vertex_request": is_vertex_request}
+
+        def finish_request(request_data: dict) -> tuple[dict, dict]:  # mutable-ok: rewritten in place downstream
+            """Filter beta headers and emit pre_call, returning `(headers, data)`.
 
             The pair stays mutable because the streaming path rewrites it in
-            place (`data["stream"] = True`) before sending.
-
-            Shared by the normal path and by the Rust path's fallback, which
-            builds it only when the Rust call did not serve the request.
+            place (`data["stream"] = True`) before sending. A Rust attempt that
+            declined already emitted pre_call for this request, so skip it there.
             """
-            request_data: Final = config.transform_request(
-                model=model,
-                messages=messages,
-                optional_params={**optional_params, "is_vertex_request": is_vertex_request},
-                litellm_params=litellm_params,
-                headers=headers,
-            )
-            return update_request_with_filtered_beta(
+            request_headers, data = update_request_with_filtered_beta(
                 headers=headers,
                 request_data=request_data,
                 provider=custom_llm_provider,
             )
+            if not serves_via_rust:
+                logging_obj.pre_call(
+                    input=messages,
+                    api_key=api_key,
+                    additional_args={
+                        "complete_input_dict": data,
+                        "api_base": api_base,
+                        "headers": request_headers,
+                    },
+                )
+            print_verbose(f"_is_function_call: {_is_function_call}")
+            return request_headers, data
+
+        async def acompletion_dispatch() -> "ModelResponse | CustomStreamWrapper":
+            """Translate then send, so the provider config can inline remote media off the event loop."""
+            request_headers, data = finish_request(
+                await config.async_transform_request(
+                    model=model,
+                    messages=messages,
+                    optional_params=transform_params,
+                    litellm_params=litellm_params,
+                    headers=headers,
+                )
+            )
+            if (
+                stream is True
+            ):  # if function call - fake the streaming (need complete blocks for output parsing in openai format)
+                print_verbose("makes async anthropic streaming POST request")
+                data["stream"] = stream
+                return await self.acompletion_stream_function(
+                    model=model,
+                    messages=messages,
+                    data=data,
+                    api_base=api_base,
+                    custom_prompt_dict=custom_prompt_dict,
+                    model_response=model_response,
+                    print_verbose=print_verbose,
+                    encoding=encoding,
+                    api_key=api_key,
+                    logging_obj=logging_obj,
+                    optional_params=optional_params,
+                    stream=stream,
+                    _is_function_call=_is_function_call,
+                    json_mode=json_mode,
+                    litellm_params=litellm_params,
+                    logger_fn=logger_fn,
+                    headers=request_headers,
+                    timeout=timeout,
+                    client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
+                )
+            return await self.acompletion_function(
+                model=model,
+                messages=messages,
+                data=data,
+                api_base=api_base,
+                custom_prompt_dict=custom_prompt_dict,
+                model_response=model_response,
+                print_verbose=print_verbose,
+                encoding=encoding,
+                api_key=api_key,
+                provider_config=config,
+                logging_obj=logging_obj,
+                optional_params=optional_params,
+                stream=stream,
+                _is_function_call=_is_function_call,
+                litellm_params=litellm_params,
+                logger_fn=logger_fn,
+                headers=request_headers,
+                client=client,
+                json_mode=json_mode,
+                timeout=timeout,
+            )
 
         # The Rust core owns the whole call for the subset it accepts, so ask
         # before transforming: whichever path runs emits pre_call exactly once.
@@ -424,35 +489,6 @@ class AnthropicChatCompletion(BaseLLM):
                 additional_args=rust_logging_args,
             )
             if acompletion is True:
-
-                async def python_fallback() -> "ModelResponse | CustomStreamWrapper":
-                    # pre_call already fired for this request above. The Rust
-                    # path only declines before the provider is called, so this
-                    # is the same attempt continuing, not a second one.
-                    fallback_headers, fallback_data = build_request()
-                    return await self.acompletion_function(
-                        model=model,
-                        messages=messages,
-                        data=fallback_data,
-                        api_base=api_base,
-                        custom_prompt_dict=custom_prompt_dict,
-                        model_response=model_response,
-                        print_verbose=print_verbose,
-                        encoding=encoding,
-                        api_key=api_key,
-                        provider_config=config,
-                        logging_obj=logging_obj,
-                        optional_params=optional_params,
-                        stream=stream,
-                        _is_function_call=_is_function_call,
-                        litellm_params=litellm_params,
-                        logger_fn=logger_fn,
-                        headers=fallback_headers,
-                        client=client,
-                        json_mode=json_mode,
-                        timeout=timeout,
-                    )
-
                 return rust_chat_completions_bridge.achat_completions_or_fallback(
                     model=model,
                     messages=messages,
@@ -464,7 +500,7 @@ class AnthropicChatCompletion(BaseLLM):
                     extra_headers=headers,
                     timeout=timeout,
                     on_response=log_rust_post_call,
-                    python_fallback=python_fallback,
+                    python_fallback=acompletion_dispatch,
                 )
             rust_response: Final = rust_chat_completions_bridge.chat_completions(
                 model=model,
@@ -481,74 +517,18 @@ class AnthropicChatCompletion(BaseLLM):
             if rust_response is not None:
                 return rust_response
 
-        headers, data = build_request()
-
-        ## LOGGING
-        # Reaching here with `serves_via_rust` set means the Rust attempt
-        # declined at call time, before the provider was called, and already
-        # logged this request. That is the same attempt continuing.
-        if not serves_via_rust:
-            logging_obj.pre_call(
-                input=messages,
-                api_key=api_key,
-                additional_args={
-                    "complete_input_dict": data,
-                    "api_base": api_base,
-                    "headers": headers,
-                },
-            )
-        print_verbose(f"_is_function_call: {_is_function_call}")
         if acompletion is True:
-            if (
-                stream is True
-            ):  # if function call - fake the streaming (need complete blocks for output parsing in openai format)
-                print_verbose("makes async anthropic streaming POST request")
-                data["stream"] = stream
-                return self.acompletion_stream_function(
-                    model=model,
-                    messages=messages,
-                    data=data,
-                    api_base=api_base,
-                    custom_prompt_dict=custom_prompt_dict,
-                    model_response=model_response,
-                    print_verbose=print_verbose,
-                    encoding=encoding,
-                    api_key=api_key,
-                    logging_obj=logging_obj,
-                    optional_params=optional_params,
-                    stream=stream,
-                    _is_function_call=_is_function_call,
-                    json_mode=json_mode,
-                    litellm_params=litellm_params,
-                    logger_fn=logger_fn,
-                    headers=headers,
-                    timeout=timeout,
-                    client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
-                )
-            else:
-                return self.acompletion_function(
-                    model=model,
-                    messages=messages,
-                    data=data,
-                    api_base=api_base,
-                    custom_prompt_dict=custom_prompt_dict,
-                    model_response=model_response,
-                    print_verbose=print_verbose,
-                    encoding=encoding,
-                    api_key=api_key,
-                    provider_config=config,
-                    logging_obj=logging_obj,
-                    optional_params=optional_params,
-                    stream=stream,
-                    _is_function_call=_is_function_call,
-                    litellm_params=litellm_params,
-                    logger_fn=logger_fn,
-                    headers=headers,
-                    client=client,
-                    json_mode=json_mode,
-                    timeout=timeout,
-                )
+            return acompletion_dispatch()
         else:
+            headers, data = finish_request(
+                config.transform_request(
+                    model=model,
+                    messages=messages,
+                    optional_params=transform_params,
+                    litellm_params=litellm_params,
+                    headers=headers,
+                )
+            )
             ## COMPLETION CALL
             if (
                 stream is True

litellm/llms/anthropic/chat/transformation.py

@@ -26,6 +26,11 @@ from litellm.litellm_core_utils.core_helpers import map_finish_reason
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     sanitize_input_schema_for_anthropic,
 )
+from litellm.litellm_core_utils.prompt_templates.image_handling import (
+    RemoteMedia,
+    async_inline_remote_media,
+    inline_remote_image_urls,
+)
 from litellm.llms.base_llm.base_utils import type_to_response_format_param
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.types.llms.anthropic import (
@@ -1840,6 +1845,25 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                 break
         return headers
 
+    def inlines_remote_media(self, media: RemoteMedia) -> bool:
+        return inline_remote_image_urls(media) and media.url.startswith("http://")
+
+    async def async_transform_request(
+        self,
+        model: str,
+        messages: list[AllMessageValues],  # mutable-ok: BaseConfig signature
+        optional_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        litellm_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        headers: dict[str, object],  # mutable-ok: BaseConfig signature
+    ) -> dict[str, object]:  # mutable-ok: BaseConfig signature
+        return self.transform_request(
+            model=model,
+            messages=await async_inline_remote_media(messages, should_inline=self.inlines_remote_media),
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
     def transform_request(
         self,
         model: str,

litellm/llms/base_llm/vector_store/transformation.py

@@ -121,6 +121,9 @@ class RouterVectorStoreEmbeddingExecutor:
 
 
 class BaseVectorStoreConfig:
+    def validate_create_vector_store(self) -> None:
+        return None
+
     def get_supported_openai_params(self, model: str) -> list[VECTOR_STORE_OPENAI_PARAMS]:
         return []
 

litellm/llms/bedrock/embed/embedding.py

@@ -35,7 +35,7 @@ from .amazon_titan_multimodal_transformation import (
 )
 from .amazon_titan_v2_transformation import AmazonTitanV2Config
 from .cohere_transformation import BedrockCohereEmbeddingConfig
-from .twelvelabs_marengo_transformation import TwelveLabsMarengoEmbeddingConfig
+from .twelvelabs_marengo_transformation import TwelveLabsMarengoEmbeddingConfig, drop_params_enabled
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
@@ -239,7 +239,7 @@ class BedrockEmbedding(BaseAWSLLM):
                 returned_response = AmazonTitanG1Config()._transform_response(response_list=response_list, model=model)
             elif provider == "twelvelabs":
                 returned_response = TwelveLabsMarengoEmbeddingConfig()._transform_response(
-                    response_list=response_list, model=model
+                    response_list=response_list, model=model, batch_data=batch_data
                 )
             elif provider == "nova":
                 returned_response = AmazonNovaEmbeddingConfig()._transform_response(
@@ -484,12 +484,13 @@ class BedrockEmbedding(BaseAWSLLM):
         elif provider == "twelvelabs":
             batch_data = []
             for i in input:
-                twelvelabs_request = TwelveLabsMarengoEmbeddingConfig()._transform_request(
+                twelvelabs_request = TwelveLabsMarengoEmbeddingConfig(model=model)._transform_request(
                     input=i,
                     inference_params=inference_params,
                     async_invoke_route=has_async_invoke,
                     model_id=modelId,
                     output_s3_uri=inference_params.get("output_s3_uri"),
+                    drop_params=drop_params_enabled(litellm_params),
                 )
                 batch_data.append(twelvelabs_request)
         elif provider == "nova":

litellm/llms/bedrock/embed/twelvelabs_marengo_3_transformation.py

@@ -0,0 +1,239 @@
+"""
+Request builder for Bedrock TwelveLabs Marengo Embed 3.0, whose payload nests the input under a key named after
+``inputType`` instead of the flat 2.7 layout.
+
+Docs - https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-marengo-3.html
+"""
+
+from collections.abc import Mapping
+from types import MappingProxyType
+from typing import Final
+
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
+from typing_extensions import assert_never
+
+from litellm.llms.bedrock.common_utils import BedrockError
+from litellm.types.llms.bedrock import (
+    TWELVELABS_MARENGO_3_EMBEDDING_OPTIONS,
+    TWELVELABS_MARENGO_3_EMBEDDING_SCOPES,
+    TWELVELABS_MARENGO_3_EMBEDDING_TYPES,
+    TWELVELABS_MARENGO_3_INPUT_TYPES,
+    TwelveLabsMarengo3AudioRequest,
+    TwelveLabsMarengo3EmbeddingRequest,
+    TwelveLabsMarengo3ImageRequest,
+    TwelveLabsMarengo3MultiInputRequest,
+    TwelveLabsMarengo3NamedMediaSource,
+    TwelveLabsMarengo3RequestBase,
+    TwelveLabsMarengo3Segmentation,
+    TwelveLabsMarengo3TextImageRequest,
+    TwelveLabsMarengo3TextRequest,
+    TwelveLabsMarengo3TimedMediaInput,
+    TwelveLabsMarengo3TimedMediaOptions,
+    TwelveLabsMarengo3VideoRequest,
+    TwelveLabsMediaSource,
+    TwelveLabsS3Location,
+)
+from litellm.utils import get_base64_str
+
+MARENGO_3_MODEL_MARKER: Final = "marengo-embed-3-"
+S3_URI_PREFIX: Final = "s3://"
+TIMED_MEDIA_OPTION_FIELDS: Final = MappingProxyType(
+    {
+        "startSec": True,
+        "endSec": True,
+        "segmentation": True,
+        "embeddingOption": True,
+        "embeddingType": True,
+        "embeddingScope": True,
+    }
+)
+TIMED_MEDIA_OPTIONS: Final = TypeAdapter(TwelveLabsMarengo3TimedMediaOptions)
+TIMED_INPUT_TYPES: Final = frozenset({"video", "audio"})
+MARENGO_2_7_ONLY_PARAMS: Final = ("textTruncate", "lengthSec", "useFixedLengthSec", "minClipSec")
+MARENGO_2_7_ONLY_FIELDS: Final = MappingProxyType({name: True for name in MARENGO_2_7_ONLY_PARAMS})
+
+
+def is_marengo_3_model(model: str | None) -> bool:
+    return MARENGO_3_MODEL_MARKER in (model or "")
+
+
+class Marengo3Params(BaseModel):
+    model_config = ConfigDict(extra="ignore", frozen=True)
+
+    inputType: TWELVELABS_MARENGO_3_INPUT_TYPES | None = None
+    input_type: TWELVELABS_MARENGO_3_INPUT_TYPES | None = None
+    media_source: str | None = None
+    media_sources: Mapping[str, str] | None = None
+    bucketOwner: str | None = None
+    startSec: float | None = None
+    endSec: float | None = None
+    segmentation: TwelveLabsMarengo3Segmentation | None = None
+    embeddingOption: tuple[TWELVELABS_MARENGO_3_EMBEDDING_OPTIONS, ...] | None = None
+    embeddingType: tuple[TWELVELABS_MARENGO_3_EMBEDDING_TYPES, ...] | None = None
+    embeddingScope: tuple[TWELVELABS_MARENGO_3_EMBEDDING_SCOPES, ...] | None = None
+    inferenceId: str | None = None
+    textTruncate: object = None
+    lengthSec: object = None
+    useFixedLengthSec: object = None
+    minClipSec: object = None
+
+    @property
+    def resolved_input_type(self) -> TWELVELABS_MARENGO_3_INPUT_TYPES:
+        return self.inputType or self.input_type or "text"
+
+    def timed_media_options(self) -> TwelveLabsMarengo3TimedMediaOptions:
+        return TIMED_MEDIA_OPTIONS.validate_python(self.given_timed_media_options())
+
+    def given_timed_media_options(self) -> dict[str, object]:
+        return self.model_dump(include=TIMED_MEDIA_OPTION_FIELDS, exclude_none=True)
+
+    def given_2_7_only_params(self) -> dict[str, object]:
+        return self.model_dump(include=MARENGO_2_7_ONLY_FIELDS, exclude_none=True)
+
+
+def _require_bucket_owner(bucket_owner: str | None) -> str:
+    if bucket_owner is None:
+        raise BedrockError(
+            status_code=400,
+            message="s3:// media requires the 'bucketOwner' parameter, the account id that owns the bucket",
+        )
+    return bucket_owner
+
+
+def _media_source(media: str, bucket_owner: str | None) -> TwelveLabsMediaSource:
+    if not media.startswith(S3_URI_PREFIX):
+        inline: Final[TwelveLabsMediaSource] = {"base64String": get_base64_str(media)}
+        return inline
+    s3_location: Final[TwelveLabsS3Location] = {"uri": media, "bucketOwner": _require_bucket_owner(bucket_owner)}
+    remote: Final[TwelveLabsMediaSource] = {"s3Location": s3_location}
+    return remote
+
+
+def _named_media_source(name: str, media: str, bucket_owner: str | None) -> TwelveLabsMarengo3NamedMediaSource:
+    named: Final[TwelveLabsMarengo3NamedMediaSource] = {
+        "name": name,
+        "mediaType": "image",
+        **_media_source(media, bucket_owner),
+    }
+    return named
+
+
+def _timed_media_input(media: str, params: Marengo3Params) -> TwelveLabsMarengo3TimedMediaInput:
+    timed: Final[TwelveLabsMarengo3TimedMediaInput] = {
+        "mediaSource": _media_source(media, params.bucketOwner),
+        **params.timed_media_options(),
+    }
+    return timed
+
+
+def _describe(error: ValidationError) -> str:
+    return "; ".join(
+        f"{'.'.join(str(part) for part in problem['loc'])}: {problem['msg']}" for problem in error.errors()
+    )
+
+
+def _validated_params(inference_params: Mapping[str, object]) -> Marengo3Params:
+    try:
+        return Marengo3Params.model_validate(inference_params)
+    except ValidationError as error:
+        raise BedrockError(status_code=400, message=f"Invalid Marengo 3.0 parameters: {_describe(error)}") from error
+
+
+def _reject_unless_dropped(given: Mapping[str, object], drop_params: bool, reason: str) -> None:
+    if not given or drop_params:
+        return
+    raise BedrockError(status_code=400, message=f"{reason} {', '.join(given)}; set drop_params to drop them")
+
+
+def _require(value: str | None, input_type: str, param_name: str) -> str:
+    if value is None:
+        raise BedrockError(status_code=400, message=f"Input type '{input_type}' requires the '{param_name}' parameter")
+    return value
+
+
+def _require_media_sources(value: Mapping[str, str] | None) -> Mapping[str, str]:
+    if not value:
+        raise BedrockError(
+            status_code=400,
+            message="Input type 'multi_input' requires a non-empty 'media_sources' mapping of name to media",
+        )
+    return value
+
+
+def _request_base(inference_id: str | None) -> TwelveLabsMarengo3RequestBase:
+    if inference_id is None:
+        anonymous: Final[TwelveLabsMarengo3RequestBase] = {}
+        return anonymous
+    identified: Final[TwelveLabsMarengo3RequestBase] = {"inferenceId": inference_id}
+    return identified
+
+
+def build_marengo_3_request(
+    input: str, inference_params: Mapping[str, object], drop_params: bool = False
+) -> TwelveLabsMarengo3EmbeddingRequest:
+    params: Final = _validated_params(inference_params)
+    base: Final = _request_base(params.inferenceId)
+    input_type: Final = params.resolved_input_type
+    _reject_unless_dropped(
+        params.given_2_7_only_params(), drop_params, "Marengo 3.0 does not accept the Marengo 2.7 parameters"
+    )
+    if input_type not in TIMED_INPUT_TYPES:
+        _reject_unless_dropped(
+            params.given_timed_media_options(), drop_params, f"Input type '{input_type}' does not accept"
+        )
+    match input_type:
+        case "text":
+            text_request: Final[TwelveLabsMarengo3TextRequest] = {
+                **base,
+                "inputType": "text",
+                "text": {"inputText": input},
+            }
+            return text_request
+        case "image":
+            image_request: Final[TwelveLabsMarengo3ImageRequest] = {
+                **base,
+                "inputType": "image",
+                "image": {"mediaSource": _media_source(input, params.bucketOwner)},
+            }
+            return image_request
+        case "video":
+            video_request: Final[TwelveLabsMarengo3VideoRequest] = {
+                **base,
+                "inputType": "video",
+                "video": _timed_media_input(input, params),
+            }
+            return video_request
+        case "audio":
+            audio_request: Final[TwelveLabsMarengo3AudioRequest] = {
+                **base,
+                "inputType": "audio",
+                "audio": _timed_media_input(input, params),
+            }
+            return audio_request
+        case "text_image":
+            text_image_request: Final[TwelveLabsMarengo3TextImageRequest] = {
+                **base,
+                "inputType": "text_image",
+                "text_image": {
+                    "inputText": input,
+                    "mediaSource": _media_source(
+                        _require(params.media_source, input_type, "media_source"), params.bucketOwner
+                    ),
+                },
+            }
+            return text_image_request
+        case "multi_input":
+            media_sources: Final = tuple(
+                _named_media_source(name, media, params.bucketOwner)
+                for name, media in _require_media_sources(params.media_sources).items()
+            )
+            multi_input_request: Final[TwelveLabsMarengo3MultiInputRequest] = {
+                **base,
+                "inputType": "multi_input",
+                "multi_input": {"inputText": input, "mediaSources": media_sources}
+                if input
+                else {"mediaSources": media_sources},
+            }
+            return multi_input_request
+        case _:
+            assert_never(input_type)

litellm/llms/bedrock/embed/twelvelabs_marengo_transformation.py

@@ -4,19 +4,120 @@ Transformation logic from OpenAI /v1/embeddings format to Bedrock TwelveLabs Mar
 Why separate file? Make it easy to see how transformation works
 
 Docs - https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-marengo.html
+Marengo 3.0 docs - https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-marengo-3.html
 """
 
+from collections.abc import Mapping
 from typing import Final, cast
 
+from pydantic import BaseModel, ConfigDict, TypeAdapter
+from typing_extensions import assert_never
+
+import litellm
+from litellm.llms.bedrock.embed.twelvelabs_marengo_3_transformation import (
+    MARENGO_2_7_ONLY_PARAMS,
+    build_marengo_3_request,
+    is_marengo_3_model,
+)
 from litellm.types.llms.bedrock import (
     TWELVELABS_EMBEDDING_INPUT_TYPES,
+    TWELVELABS_MARENGO_3_INPUT_TYPES,
     TwelveLabsAsyncInvokeRequest,
+    TwelveLabsMarengo3EmbeddingRequest,
     TwelveLabsMarengoEmbeddingRequest,
     TwelveLabsOutputDataConfig,
     TwelveLabsS3Location,
     TwelveLabsS3OutputDataConfig,
 )
-from litellm.types.utils import Embedding, EmbeddingResponse, Usage
+from litellm.types.utils import Embedding, EmbeddingResponse, PromptTokensDetailsWrapper, Usage
+
+
+class MarengoEmbeddingItem(BaseModel):
+    model_config = ConfigDict(extra="ignore", frozen=True)
+
+    embedding: tuple[float, ...] | None = None
+
+
+class MarengoInvokeResponse(BaseModel):
+    model_config = ConfigDict(extra="ignore", frozen=True)
+
+    data: tuple[MarengoEmbeddingItem, ...] = ()
+    embedding: tuple[float, ...] | None = None
+    embeddings: tuple[MarengoEmbeddingItem, ...] = ()
+
+    def vectors(self) -> tuple[tuple[float, ...], ...]:
+        if self.data:
+            return tuple(item.embedding for item in self.data if item.embedding is not None)
+        if self.embedding is not None:
+            return (self.embedding,)
+        return tuple(item.embedding for item in self.embeddings if item.embedding is not None)
+
+
+class MarengoBilledMultiInput(BaseModel):
+    model_config = ConfigDict(extra="ignore", frozen=True)
+
+    inputText: str | None = None
+    mediaSources: tuple[Mapping[str, object], ...] = ()
+
+
+class MarengoBilledRequest(BaseModel):
+    model_config = ConfigDict(extra="ignore", frozen=True)
+
+    inputType: TWELVELABS_MARENGO_3_INPUT_TYPES | None = None
+    multi_input: MarengoBilledMultiInput | None = None
+
+
+INVOKE_RESPONSES: Final = TypeAdapter(tuple[MarengoInvokeResponse, ...])
+BILLED_REQUESTS: Final = TypeAdapter(tuple[MarengoBilledRequest, ...])
+
+
+def _billed_units(request: MarengoBilledRequest) -> tuple[int, int]:
+    input_type: Final = request.inputType
+    match input_type:
+        case "text":
+            return (1, 0)
+        case "image":
+            return (0, 1)
+        case "text_image":
+            return (1, 1)
+        case "multi_input":
+            multi_input: Final = request.multi_input or MarengoBilledMultiInput()
+            return (1 if multi_input.inputText else 0, len(multi_input.mediaSources))
+        case "video" | "audio" | None:
+            return (0, 0)
+        case _:
+            assert_never(input_type)
+
+
+def _billed_usage(batch_data: list[dict] | None) -> Usage:
+    units: Final = tuple(_billed_units(request) for request in BILLED_REQUESTS.validate_python(batch_data or ()))
+    query_count: Final = sum(text_requests for text_requests, _ in units)
+    image_count: Final = sum(images for _, images in units)
+    details: Final = (
+        PromptTokensDetailsWrapper(query_count=query_count or None, image_count=image_count or None)
+        if query_count or image_count
+        else None
+    )
+    return Usage(prompt_tokens=0, completion_tokens=0, total_tokens=0, prompt_tokens_details=details)
+
+
+MARENGO_SHARED_PARAMS: Final = (
+    "encoding_format",
+    "embeddingOption",
+    "startSec",
+    "input_type",
+    "endSec",
+    "segmentation",
+    "embeddingType",
+    "embeddingScope",
+    "inferenceId",
+    "media_source",
+    "media_sources",
+)
+
+
+def drop_params_enabled(litellm_params: Mapping[str, object]) -> bool:
+    return litellm.drop_params is True or litellm_params.get("drop_params") is True
 
 
 class TwelveLabsMarengoEmbeddingConfig:
@@ -26,28 +127,24 @@ class TwelveLabsMarengoEmbeddingConfig:
     Supports text, image, video, and audio inputs.
     - InvokeModel: text and image inputs
     - StartAsyncInvoke: video, audio, image, and text inputs
+
+    Marengo 3.0 (model ids containing "marengo-embed-3") nests the input under a key named after inputType and
+    adds the text_image and multi_input input types; that payload is built by build_marengo_3_request.
     """
 
-    def __init__(self) -> None:
-        pass
+    def __init__(self, model: str | None = None) -> None:
+        self.is_marengo_3: Final = is_marengo_3_model(model)
 
     def get_supported_openai_params(self) -> list[str]:
-        return [
-            "encoding_format",
-            "textTruncate",
-            "embeddingOption",
-            "startSec",
-            "lengthSec",
-            "useFixedLengthSec",
-            "minClipSec",
-            "input_type",
-        ]
+        if self.is_marengo_3:
+            return list(MARENGO_SHARED_PARAMS)
+        return [*MARENGO_SHARED_PARAMS, *MARENGO_2_7_ONLY_PARAMS]
 
     def map_openai_params(self, non_default_params: dict, optional_params: dict) -> dict:
         for k, v in non_default_params.items():
             if k == "encoding_format":
                 # TwelveLabs doesn't have encoding_format, but we can map it to embeddingOption
-                if v == "float":
+                if v == "float" and not self.is_marengo_3:
                     optional_params["embeddingOption"] = ["visual-text", "visual-image"]
             elif k == "textTruncate":
                 optional_params["textTruncate"] = v
@@ -56,7 +153,19 @@ class TwelveLabsMarengoEmbeddingConfig:
             elif k == "input_type":
                 # Map input_type to inputType for Bedrock
                 optional_params["inputType"] = v
-            elif k in ["startSec", "lengthSec", "useFixedLengthSec", "minClipSec"]:
+            elif k in (
+                "startSec",
+                "lengthSec",
+                "useFixedLengthSec",
+                "minClipSec",
+                "endSec",
+                "segmentation",
+                "embeddingType",
+                "embeddingScope",
+                "inferenceId",
+                "media_source",
+                "media_sources",
+            ):
                 optional_params[k] = v
         return optional_params
 
@@ -77,7 +186,8 @@ class TwelveLabsMarengoEmbeddingConfig:
         async_invoke_route: bool = False,
         model_id: str | None = None,
         output_s3_uri: str | None = None,
-    ) -> TwelveLabsMarengoEmbeddingRequest | TwelveLabsAsyncInvokeRequest:
+        drop_params: bool = False,
+    ) -> TwelveLabsMarengoEmbeddingRequest | TwelveLabsMarengo3EmbeddingRequest | TwelveLabsAsyncInvokeRequest:
         """
         Transform OpenAI-style input to TwelveLabs Marengo format/async-invoke format.
 
@@ -87,20 +197,29 @@ class TwelveLabsMarengoEmbeddingConfig:
         - Video inputs (async-invoke only)
         - Audio inputs (async-invoke only)
         - S3 URLs for all media types (async-invoke only)
+        - Marengo 3.0 only: text_image and multi_input inputs (nested payload)
         """
-        # Get input_type or default to "text"
         input_type: Final = cast(
             TWELVELABS_EMBEDDING_INPUT_TYPES,
             inference_params.get("inputType") or inference_params.get("input_type") or "text",
         )
 
-        # Validate that async-invoke is used for video/audio
         if input_type in ["video", "audio"] and not async_invoke_route:
             raise ValueError(
                 f"Input type '{input_type}' requires async_invoke route. "
                 f"Use model format: 'bedrock/async_invoke/model_id'"
             )
 
+        if self.is_marengo_3:
+            marengo_3_request: Final = build_marengo_3_request(
+                input=input, inference_params=inference_params, drop_params=drop_params
+            )
+            if async_invoke_route and model_id:
+                return self._wrap_async_invoke_request(
+                    model_input=marengo_3_request, model_id=model_id, output_s3_uri=output_s3_uri
+                )
+            return marengo_3_request
+
         transformed_request: Final[TwelveLabsMarengoEmbeddingRequest] = {"inputType": input_type}
 
         if input_type == "text":
@@ -154,7 +273,7 @@ class TwelveLabsMarengoEmbeddingConfig:
 
     def _wrap_async_invoke_request(
         self,
-        model_input: TwelveLabsMarengoEmbeddingRequest,
+        model_input: TwelveLabsMarengoEmbeddingRequest | TwelveLabsMarengo3EmbeddingRequest,
         model_id: str,
         output_s3_uri: str | None = None,
     ) -> TwelveLabsAsyncInvokeRequest:
@@ -188,62 +307,16 @@ class TwelveLabsMarengoEmbeddingConfig:
             ),
         )
 
-    def _transform_response(self, response_list: list[dict], model: str) -> EmbeddingResponse:
-        """
-        Transform TwelveLabs response to OpenAI format.
-        Handles the actual TwelveLabs response format: {"data": [{"embedding": [...]}]}
-        """
-        embeddings: Final[list[Embedding]] = []
-        total_tokens = 0
-
-        for response in response_list:
-            # TwelveLabs response format has a "data" field containing the embeddings
-            if "data" in response and isinstance(response["data"], list):
-                for item in response["data"]:
-                    if "embedding" in item:
-                        # Single embedding response
-                        embedding = Embedding(
-                            embedding=item["embedding"],
-                            index=len(embeddings),
-                            object="embedding",
-                        )
-                        embeddings.append(embedding)
-
-                        # Estimate token count (rough approximation)
-                        if "inputTextTokenCount" in item:
-                            total_tokens += item["inputTextTokenCount"]
-                        else:
-                            # Rough estimate: 1 token per 4 characters for text, or use embedding size
-                            total_tokens += len(item["embedding"]) // 4
-            elif "embedding" in response:
-                # Direct embedding response (fallback for other formats)
-                embedding = Embedding(
-                    embedding=response["embedding"],
-                    index=len(embeddings),
-                    object="embedding",
-                )
-                embeddings.append(embedding)
-
-                # Estimate token count (rough approximation)
-                if "inputTextTokenCount" in response:
-                    total_tokens += response["inputTextTokenCount"]
-                else:
-                    # Rough estimate: 1 token per 4 characters for text
-                    total_tokens += len(response.get("inputText", "")) // 4
-            elif "embeddings" in response:
-                # Multiple embeddings response (from video/audio)
-                for i, emb in enumerate(response["embeddings"]):
-                    embedding = Embedding(
-                        embedding=emb["embedding"],
-                        index=len(embeddings),
-                        object="embedding",
-                    )
-                    embeddings.append(embedding)
-                    total_tokens += len(emb["embedding"]) // 4  # Rough estimate
-
-        usage: Final = Usage(prompt_tokens=total_tokens, total_tokens=total_tokens)
-
-        return EmbeddingResponse(data=embeddings, model=model, usage=usage)
+    def _transform_response(
+        self, response_list: list[dict], model: str, batch_data: list[dict] | None = None
+    ) -> EmbeddingResponse:
+        vectors: Final = tuple(
+            vector for response in INVOKE_RESPONSES.validate_python(response_list) for vector in response.vectors()
+        )
+        embeddings: Final = [
+            Embedding(embedding=list(vector), index=index, object="embedding") for index, vector in enumerate(vectors)
+        ]
+        return EmbeddingResponse(data=embeddings, model=model, usage=_billed_usage(batch_data))
 
     def _transform_async_invoke_response(self, response: dict, model: str) -> EmbeddingResponse:
         """

litellm/llms/custom_httpx/llm_http_handler.py

@@ -9826,7 +9826,7 @@ class BaseLLMHTTPHandler:
                 vector_store_search_optional_params=vector_store_search_optional_params,
                 api_base=api_base,
                 litellm_logging_obj=logging_obj,
-                litellm_params=dict(litellm_params),
+                litellm_params=MappingProxyType(dict(litellm_params, timeout=timeout)),
                 extra_body=extra_body,
                 embedding_executor=embedding_executor,
             )
@@ -9871,6 +9871,12 @@ class BaseLLMHTTPHandler:
                 data=request_data,
                 timeout=timeout,
             )
+        except httpx.TimeoutException:
+            raise vector_store_provider_config.get_error_class(
+                error_message="Vector store search exceeded the caller timeout.",
+                status_code=408,
+                headers=httpx.Headers(),
+            ) from None
         except Exception as e:
             raise self._handle_error(e=e, provider_config=vector_store_provider_config)
 
@@ -9955,7 +9961,7 @@ class BaseLLMHTTPHandler:
                 vector_store_search_optional_params=vector_store_search_optional_params,
                 api_base=api_base,
                 litellm_logging_obj=logging_obj,
-                litellm_params=dict(litellm_params),
+                litellm_params=MappingProxyType(dict(litellm_params, timeout=timeout)),
                 extra_body=extra_body,
                 embedding_executor=embedding_executor,
             )
@@ -10000,7 +10006,14 @@ class BaseLLMHTTPHandler:
                 url=url,
                 headers=headers,
                 data=request_data,
+                timeout=timeout,
             )
+        except httpx.TimeoutException:
+            raise vector_store_provider_config.get_error_class(
+                error_message="Vector store search exceeded the caller timeout.",
+                status_code=408,
+                headers=httpx.Headers(),
+            ) from None
         except Exception as e:
             raise self._handle_error(e=e, provider_config=vector_store_provider_config)
 
@@ -10030,6 +10043,8 @@ class BaseLLMHTTPHandler:
         else:
             async_httpx_client = client
 
+        vector_store_provider_config.validate_create_vector_store()
+
         headers: Final = vector_store_provider_config.validate_environment(
             headers=extra_headers or {}, litellm_params=litellm_params
         )
@@ -10100,6 +10115,8 @@ class BaseLLMHTTPHandler:
         else:
             sync_httpx_client = client
 
+        vector_store_provider_config.validate_create_vector_store()
+
         headers: Final = vector_store_provider_config.validate_environment(
             headers=extra_headers or {}, litellm_params=litellm_params
         )

litellm/llms/mongodb/common_utils.py

@@ -1,303 +0,0 @@
-"""Shared helpers for the MongoDB integrations. pymongo lives in the optional ``mongodb`` extra,
-so every import of it is deferred to call time."""
-
-import asyncio
-import threading
-import weakref
-from asyncio import AbstractEventLoop
-from collections import OrderedDict
-from collections.abc import Callable, Mapping
-from dataclasses import dataclass
-from types import MappingProxyType
-from typing import TYPE_CHECKING, Final, TypeAlias, TypeVar
-
-from litellm.exceptions import BadRequestError, ServiceUnavailableError, Timeout
-
-if TYPE_CHECKING:
-    from pymongo import AsyncMongoClient, MongoClient
-
-PYMONGO_INSTALL_HINT: Final = (
-    "The MongoDB vector store requires the 'pymongo' package. "
-    "Run 'pip install litellm[mongodb]' (or 'pip install pymongo') to install it."
-)
-
-MONGODB_PROVIDER: Final = "mongodb"
-
-
-def config_error(message: str) -> BadRequestError:
-    """400 rather than the 500 a bare ValueError becomes once litellm.exception_type wraps it."""
-    return BadRequestError(message=message, model=None, llm_provider=MONGODB_PROVIDER)
-
-
-def timeout_error(message: str) -> Timeout:
-    return Timeout(message=message, model=None, llm_provider=MONGODB_PROVIDER)
-
-
-def unavailable_error(message: str) -> ServiceUnavailableError:
-    """litellm only retries 408, 409, 429 and 5xx, so a 400 here would make a failover permanent."""
-    return ServiceUnavailableError(message=message, model=None, llm_provider=MONGODB_PROVIDER)
-
-
-DEFAULT_CONNECT_TIMEOUT_MS: Final = 10_000
-DEFAULT_SOCKET_TIMEOUT_MS: Final = 30_000
-DEFAULT_SERVER_SELECTION_TIMEOUT_MS: Final = 10_000
-
-_MAX_CACHED_CLIENTS: Final = 32
-
-_APP_NAME: Final = "litellm"
-
-
-@dataclass(frozen=True, slots=True)
-class MongoClientKey:
-    connection_string: str
-    connect_timeout_ms: int
-    socket_timeout_ms: int
-    server_selection_timeout_ms: int
-
-
-SyncClientFactory: TypeAlias = Callable[..., "MongoClient"]
-AsyncClientFactory: TypeAlias = Callable[..., "AsyncMongoClient"]
-
-_K = TypeVar("_K")
-_V = TypeVar("_V")
-
-_AsyncClientCacheKey: TypeAlias = tuple[MongoClientKey, int]
-# CPython recycles id() aggressively, so the id alone would hand a new loop a closed loop's client
-_AsyncClientEntry: TypeAlias = tuple["weakref.ref[AbstractEventLoop]", "AsyncMongoClient"]
-
-_SyncClientCache: TypeAlias = "OrderedDict[MongoClientKey, MongoClient]"
-_AsyncClientCache: TypeAlias = "OrderedDict[_AsyncClientCacheKey, _AsyncClientEntry]"
-
-_sync_clients: Final[_SyncClientCache] = OrderedDict()  # mutable-ok: process-level client cache
-_async_clients: Final[_AsyncClientCache] = OrderedDict()  # mutable-ok: same cache, per loop
-# async searches reach the sync client through executor threads, so both caches are shared state
-_cache_lock: Final = threading.Lock()
-
-
-def _store_bounded(cache: "OrderedDict[_K, _V]", cache_key: "_K", value: "_V") -> None:
-    """Eviction only drops this cache's reference; an in-flight search keeps its client alive."""
-    with _cache_lock:
-        cache[cache_key] = value  # mutable-ok: an LRU cache is mutable state by definition
-        cache.move_to_end(cache_key)
-        while len(cache) > _MAX_CACHED_CLIENTS:
-            cache.popitem(last=False)
-
-
-def _mark_used(cache: "OrderedDict[_K, _V]", cache_key: "_K") -> None:
-    with _cache_lock:
-        if cache_key in cache:
-            cache.move_to_end(cache_key)
-
-
-def import_sync_mongo_client() -> "type[MongoClient]":
-    try:
-        from pymongo import MongoClient as SyncMongoClient
-    except ImportError as e:
-        raise config_error(PYMONGO_INSTALL_HINT) from e
-    return SyncMongoClient
-
-
-def import_async_mongo_client() -> "type[AsyncMongoClient]":
-    try:
-        from pymongo import AsyncMongoClient as AsyncMongoClientClass
-    except ImportError as e:
-        raise config_error(PYMONGO_INSTALL_HINT) from e
-    return AsyncMongoClientClass
-
-
-def _client_kwargs(key: MongoClientKey) -> Mapping[str, object]:
-    return MappingProxyType(
-        {
-            "connectTimeoutMS": key.connect_timeout_ms,
-            "socketTimeoutMS": key.socket_timeout_ms,
-            "serverSelectionTimeoutMS": key.server_selection_timeout_ms,
-            "appname": _APP_NAME,
-        }
-    )
-
-
-def get_sync_client(key: MongoClientKey, client_class: SyncClientFactory | None = None) -> "MongoClient":
-    cached: Final = _sync_clients.get(key)
-    if cached is not None:
-        _mark_used(_sync_clients, key)
-        return cached
-    build: Final = client_class if client_class is not None else import_sync_mongo_client()
-    client: Final = build(key.connection_string, **_client_kwargs(key))
-    _store_bounded(_sync_clients, key, client)
-    return client
-
-
-def _purge_dead_loops() -> None:
-    """A cached client holds its loop alive, so a closed loop's entry would pin that client and its
-    sockets for the life of the process."""
-    with _cache_lock:
-        for stale in tuple(
-            cache_key
-            for cache_key, (loop_ref, _) in _async_clients.items()
-            if (cached_loop := loop_ref()) is None or cached_loop.is_closed()
-        ):
-            del _async_clients[stale]
-
-
-def get_async_client(key: MongoClientKey, client_class: AsyncClientFactory | None = None) -> "AsyncMongoClient":
-    """Async clients bind to the loop that created them, so the cache is keyed per loop."""
-    loop: Final = asyncio.get_running_loop()
-    loop_key: Final = (key, id(loop))
-    cached: Final = _async_clients.get(loop_key)
-    if cached is not None and cached[0]() is loop:
-        _mark_used(_async_clients, loop_key)
-        return cached[1]
-    _purge_dead_loops()
-    build: Final = client_class if client_class is not None else import_async_mongo_client()
-    client: Final = build(key.connection_string, **_client_kwargs(key))
-    _store_bounded(_async_clients, loop_key, (weakref.ref(loop), client))
-    return client
-
-
-def reset_client_cache() -> None:
-    with _cache_lock:
-        _sync_clients.clear()
-        _async_clients.clear()
-
-
-_AUTHENTICATION_FAILED_CODE: Final = 18
-_UNAUTHORIZED_CODE: Final = 13
-# Atlas reports a rejected user as code 8000 "AtlasError" where a self-managed mongod reports 18
-_AUTHENTICATION_MESSAGE_MARKERS: Final = ("bad auth", "authentication failed", "not authorized")
-_RESOLUTION_TIMEOUT_MARKERS: Final = ("resolution lifetime expired", "dns operation timed out")
-_UNKNOWN_HOSTNAME_MARKERS: Final = ("dns query name does not exist", "name or service not known")
-_CREDENTIAL_ESCAPING_MARKERS: Final = ("must be escaped according to rfc 3986", "bad database name")
-
-
-def _index_hint(index_name: str, database: str, collection: str) -> str:
-    return (
-        f"No queryable MongoDB Vector Search index named '{index_name}' was found on "
-        f"'{database}.{collection}'. Confirm the index exists on that exact collection, that its "
-        "status is READY rather than still building, and that the vector store id matches the index name."
-    )
-
-
-def missing_index_error(index_name: str, database: str, collection: str) -> BadRequestError:
-    """$vectorSearch against a missing index, database or collection returns zero documents rather
-    than failing, so an empty result set is checked against the catalogue and reported as this."""
-    return config_error(
-        f"{_index_hint(index_name, database, collection)} A vector search against a database, "
-        "collection or index that does not exist returns no results rather than an error, so this "
-        "was reported as an empty result set by MongoDB."
-    )
-
-
-def index_not_ready_error(index_name: str, database: str, collection: str, status: str) -> BadRequestError:
-    return config_error(
-        f"The MongoDB Vector Search index '{index_name}' on '{database}.{collection}' is not queryable "
-        f"yet; its status is {status}. Searches against it return no results until the build finishes."
-    )
-
-
-def translate_mongo_error(error: Exception, index_name: str, database: str, collection: str) -> Exception:
-    """Returns the exception to raise, so callers keep the driver error as ``__cause__``."""
-    try:
-        from pymongo.errors import (
-            ConfigurationError,
-            ConnectionFailure,
-            ExecutionTimeout,
-            InvalidOperation,
-            NetworkTimeout,
-            OperationFailure,
-            ServerSelectionTimeoutError,
-        )
-    except ImportError:
-        return error
-
-    if isinstance(error, ServerSelectionTimeoutError):
-        return timeout_error(
-            "Could not reach the MongoDB deployment before the timeout. On Atlas this is usually the "
-            "project's IP access list not containing this host, or a paused cluster. On a self-managed "
-            "deployment it is usually the host or port in the URI, or a firewall between this process "
-            f"and mongod. Either way it can also be an unresolvable hostname. Driver detail: {error}"
-        )
-    # ExecutionTimeout subclasses OperationFailure, so it has to be matched before it
-    if isinstance(error, (NetworkTimeout, ExecutionTimeout)):
-        return timeout_error(
-            f"The MongoDB vector search against '{database}.{collection}' timed out before returning. "
-            f"Driver detail: {error}"
-        )
-    # ServerSelectionTimeoutError and NetworkTimeout also subclass ConnectionFailure, so this only
-    # sees what those branches left
-    if isinstance(error, ConnectionFailure):
-        return unavailable_error(
-            f"The connection to '{database}.{collection}' was dropped or refused. That is usually a "
-            "replica set failover or a restarted node, so the search is worth retrying. If it keeps "
-            "happening: on Atlas the usual cause is a connection string with no username and password, "
-            "or a TLS failure, so confirm the URI is the one Atlas shows under Connect, Drivers; on a "
-            "self-managed deployment, check that mongod is listening on the host and port in the URI. "
-            f"Driver detail: {error}"
-        )
-    if isinstance(error, OperationFailure):
-        code: Final = error.code
-        detail: Final = str(error).lower()
-        if code in (_AUTHENTICATION_FAILED_CODE, _UNAUTHORIZED_CODE) or any(
-            marker in detail for marker in _AUTHENTICATION_MESSAGE_MARKERS
-        ):
-            return config_error(
-                "MongoDB rejected the credentials in mongodb_connection_string, or the database user "
-                f"lacks read access to '{database}.{collection}'. Driver detail: {error.details}"
-            )
-        if "dimension" in detail:
-            return config_error(
-                "The query embedding does not match the vector dimensions the index was built for. "
-                "litellm_embedding_model must be the same model that produced the stored vectors. "
-                f"Driver detail: {error}"
-            )
-        if "is not indexed as vector" in detail:
-            return config_error(
-                "mongodb_embedding_field names a field the MongoDB Vector Search index does not cover. "
-                f"It must match the 'path' the index '{index_name}' was created on. Driver detail: {error}"
-            )
-        if "index" in detail and ("not found" in detail or "does not exist" in detail or "unknown" in detail):
-            return config_error(f"{_index_hint(index_name, database, collection)} Driver detail: {error}")
-        return config_error(
-            f"MongoDB rejected the vector search against '{database}.{collection}' using index "
-            f"'{index_name}'. Driver detail: {error}"
-        )
-    if isinstance(error, ConfigurationError):
-        configuration_detail: Final = str(error).lower()
-        if any(marker in configuration_detail for marker in _RESOLUTION_TIMEOUT_MARKERS):
-            return timeout_error(
-                "The DNS lookup for the cluster in mongodb_connection_string did not finish in time. "
-                "A mongodb+srv:// URI needs an SRV lookup before any connection is attempted, so this "
-                f"is DNS or the configured timeout, not MongoDB. Driver detail: {error}"
-            )
-        if any(marker in configuration_detail for marker in _UNKNOWN_HOSTNAME_MARKERS):
-            return config_error(
-                "The hostname in mongodb_connection_string does not exist in DNS. On Atlas, check the "
-                "cluster name against the URI shown under Connect, Drivers. On a self-managed deployment, "
-                f"check that the hostname resolves from this process. Driver detail: {error}"
-            )
-        if any(marker in configuration_detail for marker in _CREDENTIAL_ESCAPING_MARKERS):
-            return config_error(
-                "mongodb_connection_string could not be parsed. A username or password containing "
-                "'@', '/', ':' or '%' has to be percent-encoded per RFC 3986, so 'p@ss/word' becomes "
-                "'p%40ss%2Fword'. If the credentials are already encoded, check the database name in "
-                f"the URI path instead. Driver detail: {error}"
-            )
-        return config_error(
-            f"mongodb_connection_string is not a usable MongoDB connection string. Driver detail: {error}"
-        )
-    if isinstance(error, InvalidOperation):
-        return config_error(f"The MongoDB client was already closed or is unusable. Driver detail: {error}")
-    # An unreadable tlsCAFile or tlsCertificateKeyFile raises OSError, not a PyMongoError
-    if isinstance(error, OSError) and error.filename:
-        return config_error(
-            f"'{error.filename}', named by a TLS option in mongodb_connection_string, could not be read. "
-            "Check that tlsCAFile and tlsCertificateKeyFile point at files this process can open; inside "
-            f"a container that is the path in the container, not on the host. Driver detail: {error}"
-        )
-    # pymongo raises a plain ValueError, not a PyMongoError, for an unusable port
-    if isinstance(error, ValueError):
-        return config_error(
-            "The host and port in mongodb_connection_string could not be parsed. If the port is a "
-            "number between 0 and 65535, the cause is usually an unescaped ':' in the password, which "
-            f"has to be percent-encoded per RFC 3986 as '%3A'. Driver detail: {error}"
-        )
-    return error

litellm/llms/mongodb/vector_stores/transformation.py

@@ -1,37 +1,29 @@
-"""MongoDB Vector Search has no HTTP query API, so this is a direct provider that runs the
-``$vectorSearch`` aggregation through pymongo. ``vector_store_id`` is the search index name."""
-
-from collections.abc import Callable, Mapping, Sequence
+from collections.abc import Mapping, Sequence
+from ipaddress import ip_address
+from math import isfinite
 from types import MappingProxyType
-from typing import TYPE_CHECKING, Final, NoReturn
+from typing import TYPE_CHECKING, Final, Literal, NoReturn
+from urllib.parse import quote, urlsplit
 
 import httpx
-from pydantic import BaseModel, ConfigDict
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 
+from litellm.exceptions import AuthenticationError, BadRequestError, ServiceUnavailableError, Timeout
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.vector_store.transformation import (
-    BaseDirectVectorStoreConfig,
+    BaseQueryEmbeddingVectorStoreConfig,
     LiteLLMVectorStoreEmbeddingExecutor,
     VectorStoreEmbeddingExecutor,
 )
-from litellm.llms.mongodb.common_utils import (
-    DEFAULT_CONNECT_TIMEOUT_MS,
-    DEFAULT_SERVER_SELECTION_TIMEOUT_MS,
-    DEFAULT_SOCKET_TIMEOUT_MS,
-    MongoClientKey,
-    config_error,
-    get_async_client,
-    get_sync_client,
-    index_not_ready_error,
-    missing_index_error,
-    translate_mongo_error,
-)
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import EmbeddingResponse
 from litellm.types.vector_stores import (
+    BaseVectorStoreAuthCredentials,
     VectorStoreCreateOptionalRequestParams,
-    VectorStoreResultContent,
+    VectorStoreIndexEndpoints,
     VectorStoreSearchOptionalRequestParams,
     VectorStoreSearchResponse,
-    VectorStoreSearchResult,
 )
 
 if TYPE_CHECKING:
@@ -39,26 +31,45 @@ if TYPE_CHECKING:
 
 DEFAULT_EMBEDDING_FIELD_NAME: Final = "embedding"
 DEFAULT_TEXT_FIELD_NAME: Final = "text"
-SCORE_FIELD_NAME: Final = "score"
-
 DEFAULT_MAX_NUM_RESULTS: Final = 10
 MIN_MAX_NUM_RESULTS: Final = 1
 MAX_MAX_NUM_RESULTS: Final = 50
-
 NUM_CANDIDATES_MULTIPLIER: Final = 10
 MIN_NUM_CANDIDATES: Final = 100
 MAX_NUM_CANDIDATES: Final = 10_000
-
 MAX_QUERY_CHARACTERS: Final = 32_000
-
 _EMPTY_EMBEDDING_CONFIG: Final = MappingProxyType({})
-
 _SEARCH_ONLY_MESSAGE: Final = (
     "MongoDB vector store is search-only. Create the collection and its MongoDB Vector Search "
     "index in MongoDB directly, then register it here by index name."
 )
 
 
+def config_error(message: str) -> BadRequestError:
+    return BadRequestError(message=message, model=None, llm_provider="mongodb")
+
+
+class _Content(BaseModel):
+    model_config = ConfigDict(frozen=True, strict=True)
+    type: Literal["text"]
+    text: str
+
+
+class _Result(BaseModel):
+    model_config = ConfigDict(frozen=True, strict=True, allow_inf_nan=False)
+    score: float | None
+    content: Sequence[_Content]
+    file_id: str | None
+    filename: str | None
+
+
+class _SearchResponse(BaseModel):
+    model_config = ConfigDict(frozen=True, strict=True)
+    object: Literal["vector_store.search_results.page"]
+    search_query: str
+    data: Sequence[_Result]
+
+
 class _MongoDBSearchParams(BaseModel):
     """Typed view over the vector store's litellm_params; unrelated keys are ignored."""
 
@@ -66,7 +77,6 @@ class _MongoDBSearchParams(BaseModel):
 
     litellm_embedding_model: str | None = None
     litellm_embedding_config: Mapping[str, object] | None = None
-    mongodb_connection_string: str | None = None
     mongodb_database: str | None = None
     mongodb_collection: str | None = None
     mongodb_text_field: str | None = None
@@ -91,21 +101,6 @@ class _MongoDBSearchParams(BaseModel):
             )
         return self.litellm_embedding_model
 
-    def require_connection_string(self) -> str:
-        if not self.mongodb_connection_string:
-            raise config_error(
-                "mongodb_connection_string is required in litellm_params for the MongoDB vector store. "
-                "Example: mongodb+srv://<user>:<password>@<cluster>.mongodb.net for Atlas, or "
-                "mongodb://<user>:<password>@<host>:27017 for a self-managed deployment"
-            )
-        scheme: Final = self.mongodb_connection_string.split("://", 1)[0].lower()
-        if scheme not in ("mongodb", "mongodb+srv"):
-            raise config_error(
-                "mongodb_connection_string must start with 'mongodb://' or 'mongodb+srv://', "
-                f"got '{self.mongodb_connection_string.split('://', 1)[0]}://'"
-            )
-        return self.mongodb_connection_string
-
     def require_database(self) -> str:
         if not self.mongodb_database:
             raise config_error(
@@ -127,30 +122,28 @@ _MONGODB_PARAM_PREFIX: Final = "mongodb_"
 _KNOWN_MONGODB_PARAMS: Final = frozenset(
     name for name in _MongoDBSearchParams.model_fields if name.startswith(_MONGODB_PARAM_PREFIX)
 )
+_RESPONSE_ADAPTER: Final = TypeAdapter(VectorStoreSearchResponse)
 
 
-class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig):
-    def __init__(
-        self,
-        embedding_executor: VectorStoreEmbeddingExecutor | None = None,
-        sync_client_factory: Callable[[MongoClientKey], object] | None = None,
-        async_client_factory: Callable[[MongoClientKey], object] | None = None,
-    ) -> None:
-        super().__init__()
-        self.embedding_executor: Final[VectorStoreEmbeddingExecutor] = (
-            embedding_executor if embedding_executor is not None else LiteLLMVectorStoreEmbeddingExecutor()
-        )
-        self.sync_client_factory: Final[Callable[[MongoClientKey], object]] = (
-            sync_client_factory if sync_client_factory is not None else get_sync_client
-        )
-        self.async_client_factory: Final[Callable[[MongoClientKey], object]] = (
-            async_client_factory if async_client_factory is not None else get_async_client
-        )
+class MongoDBVectorStoreConfig(BaseQueryEmbeddingVectorStoreConfig):
+    def __init__(self, embedding_executor: VectorStoreEmbeddingExecutor | None = None) -> None:
+        self.embedding_executor: Final = embedding_executor or LiteLLMVectorStoreEmbeddingExecutor()
+
+    def get_auth_credentials(self, litellm_params: Mapping[str, object]) -> BaseVectorStoreAuthCredentials:
+        return BaseVectorStoreAuthCredentials()
+
+    def get_vector_store_endpoints_by_type(self) -> VectorStoreIndexEndpoints:
+        return VectorStoreIndexEndpoints(read=[], write=[])  # mutable-ok: the TypedDict declares list fields
 
     @staticmethod
     def _reject_unknown_params(litellm_params: Mapping[str, object]) -> None:
         """Without this a mistyped mongodb_collection reads as 'mongodb_collection is required',
         naming a key the reader can see they have set."""
+        if litellm_params.get("mongodb_connection_string") is not None:
+            raise config_error(
+                "MongoDB vector stores now use the BETA sidecar. Move mongodb_connection_string to "
+                "MONGODB_CONNECTION_STRING in the sidecar, remove it from LiteLLM, and configure api_base and api_key."
+            )
         unknown: Final = sorted(
             key for key in litellm_params if key.startswith(_MONGODB_PARAM_PREFIX) and key not in _KNOWN_MONGODB_PARAMS
         )
@@ -191,239 +184,203 @@ class MongoDBVectorStoreConfig(BaseDirectVectorStoreConfig):
             return configured
         return min(max(limit * NUM_CANDIDATES_MULTIPLIER, MIN_NUM_CANDIDATES), MAX_NUM_CANDIDATES)
 
-    @staticmethod
-    def _timeout_ms(timeout: float | httpx.Timeout | None) -> tuple[int, int]:
-        """The connect and socket budgets pymongo is built with, in that order."""
-        if isinstance(timeout, httpx.Timeout):
-            return (
-                int((timeout.connect or DEFAULT_CONNECT_TIMEOUT_MS / 1000) * 1000),
-                int((timeout.read or DEFAULT_SOCKET_TIMEOUT_MS / 1000) * 1000),
+    def validate_environment(
+        self, headers: Mapping[str, object], litellm_params: GenericLiteLLMParams | None
+    ) -> dict[str, object]:  # mutable-ok: the shared HTTP handler requires writable headers
+        if litellm_params is None:
+            raise config_error("Configure api_base and api_key for the MongoDB BETA sidecar.")
+        self._reject_unknown_params(MappingProxyType(dict(litellm_params)))
+        api_key: Final = litellm_params.api_key or get_secret_str("MONGODB_SIDECAR_API_KEY")
+        if not api_key:
+            raise config_error("MongoDB sidecar api_key is required. Set api_key or MONGODB_SIDECAR_API_KEY.")
+        return {
+            **headers,
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }  # mutable-ok: writable HTTP headers
+
+    def get_complete_url(self, api_base: str | None, litellm_params: Mapping[str, object]) -> str:
+        if not api_base:
+            raise config_error("MongoDB sidecar api_base is required, for example http://127.0.0.1:8080.")
+        try:
+            parsed: Final = urlsplit(api_base)
+            valid: Final = parsed.scheme in ("http", "https") and bool(parsed.hostname) and parsed.port != 0
+        except ValueError:
+            raise config_error("MongoDB sidecar api_base must be a valid HTTP or HTTPS URL.") from None
+        if not valid or parsed.username or parsed.password or parsed.query or parsed.fragment:
+            raise config_error(
+                "MongoDB sidecar api_base must be an HTTP or HTTPS URL without credentials, query, or fragment."
             )
-        if timeout is None:
-            return DEFAULT_CONNECT_TIMEOUT_MS, DEFAULT_SOCKET_TIMEOUT_MS
-        return min(int(float(timeout) * 1000), DEFAULT_CONNECT_TIMEOUT_MS), int(float(timeout) * 1000)
+        if parsed.scheme == "http":
+            try:
+                loopback: Final = ip_address(parsed.hostname or "").is_loopback
+            except ValueError:
+                raise config_error(
+                    "MongoDB sidecar requires HTTPS. HTTP is supported only for a loopback IP such as 127.0.0.1."
+                ) from None
+            if not loopback:
+                raise config_error(
+                    "MongoDB sidecar requires HTTPS. HTTP is supported only for a loopback IP such as 127.0.0.1."
+                )
+        return api_base.rstrip("/")
+
+    @staticmethod
+    def _timeout_ms(value: object) -> int:
+        seconds: Final = value.read if isinstance(value, httpx.Timeout) else value
+        if seconds is None:
+            return 30_000
+        if not isinstance(seconds, (int, float)) or not isfinite(seconds) or seconds <= 0:
+            raise config_error("MongoDB search timeout must be a positive finite number.")
+        try:
+            return max(1, int(seconds * 1000))
+        except (ValueError, OverflowError):
+            raise config_error("MongoDB search timeout must be a positive finite number.") from None
 
     @classmethod
-    def _client_key(cls, params: _MongoDBSearchParams, timeout: float | httpx.Timeout | None) -> MongoClientKey:
-        connect_ms, socket_ms = cls._timeout_ms(timeout)
-        return MongoClientKey(
-            connection_string=params.require_connection_string(),
-            connect_timeout_ms=connect_ms,
-            socket_timeout_ms=socket_ms,
-            server_selection_timeout_ms=min(connect_ms, DEFAULT_SERVER_SELECTION_TIMEOUT_MS),
-        )
+    def _params(
+        cls,
+        litellm_params: Mapping[str, object],
+        optional_params: VectorStoreSearchOptionalRequestParams,
+        extra_body: Mapping[str, object] | None,
+    ) -> _MongoDBSearchParams:
+        cls._reject_unknown_params(litellm_params)
+        if extra_body:
+            raise config_error("MongoDB vector store does not support extra_body overrides.")
+        for unsupported in ("filters", "ranking_options", "rewrite_query"):
+            if optional_params.get(unsupported) is not None:
+                raise config_error(f"MongoDB vector store does not support the {unsupported} parameter.")
+        try:
+            params: Final = _MongoDBSearchParams.model_validate(litellm_params)
+        except ValidationError:
+            raise config_error(
+                "Invalid MongoDB vector-store configuration. Check the database, collection, fields, and candidate count."
+            ) from None
+        params.require_database()
+        params.require_collection()
+        params.require_embedding_model()
+        cls._num_candidates(cls._limit(optional_params), params.mongodb_num_candidates)
+        cls._timeout_ms(litellm_params.get("timeout"))
+        return params
 
     @classmethod
-    def _pipeline(
+    def _request(
         cls,
         vector_store_id: str,
-        query_vector: Sequence[float],
+        query_text: str,
         params: _MongoDBSearchParams,
-        vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
-    ) -> Sequence[Mapping[str, object]]:
-        if vector_store_search_optional_params.get("filters") is not None:
+        optional_params: VectorStoreSearchOptionalRequestParams,
+        api_base: str,
+        embedding_response: EmbeddingResponse,
+        timeout: object,
+    ) -> tuple[str, dict[str, object]]:  # mutable-ok: the provider contract returns a writable JSON request body
+        if not embedding_response.data:
             raise config_error(
-                "MongoDB vector store does not support the filters parameter yet. "
-                "Restrict the collection or the MongoDB Vector Search index definition instead."
+                "The embedding model returned no embedding for the search query. Check litellm_embedding_model."
             )
-        if vector_store_search_optional_params.get("ranking_options") is not None:
-            raise config_error(
-                "MongoDB vector store does not support the ranking_options parameter yet. "
-                "Every result already carries the vectorSearchScore, so filter or re-rank "
-                "on that rather than having the threshold silently ignored."
-            )
-        if vector_store_search_optional_params.get("rewrite_query") is not None:
-            raise config_error(
-                "MongoDB vector store does not support the rewrite_query parameter. The query is "
-                "embedded exactly as sent; rewrite it before calling if you need that."
-            )
-        limit: Final = cls._limit(vector_store_search_optional_params)
-        search: Final = MappingProxyType(
-            {
-                "index": vector_store_id,
-                "path": params.embedding_field,
-                "queryVector": tuple(query_vector),
-                "numCandidates": cls._num_candidates(limit, params.mongodb_num_candidates),
-                "limit": limit,
-            }
-        )
-        projection: Final = MappingProxyType(
-            {params.text_field: 1, SCORE_FIELD_NAME: MappingProxyType({"$meta": "vectorSearchScore"})}
-        )
-        return [  # mutable-ok: pymongo rejects any non-list pipeline in common.validate_list
-            MappingProxyType({"$vectorSearch": search}),
-            MappingProxyType({"$project": projection}),
-        ]
-
-    @classmethod
-    def _field_value(cls, document: Mapping[str, object], dotted_path: str) -> str | None:
-        """None means absent, which is what separates a mistyped field from genuinely empty text."""
-        head, _, rest = dotted_path.partition(".")
-        if head not in document:
-            return None
-        value: Final = document[head]
-        if not rest:
-            return None if value is None else str(value)
-        return cls._field_value(value, rest) if isinstance(value, Mapping) else None
-
-    @classmethod
-    def _to_result(cls, document: Mapping[str, object], text_field: str) -> VectorStoreSearchResult:
-        document_id: Final = document.get("_id")
-        identifier: Final = None if document_id is None else str(document_id)
-        content: Final = [  # mutable-ok: VectorStoreSearchResult declares a list of content parts
-            VectorStoreResultContent(text=cls._field_value(document, text_field) or "", type="text")
-        ]
-        raw_score: Final = document.get(SCORE_FIELD_NAME)
-        return VectorStoreSearchResult(
-            score=float(raw_score) if isinstance(raw_score, (int, float)) else None,
-            content=content,
-            file_id=identifier,
-            filename=identifier,
+        vector: Final = embedding_response.data[0]["embedding"]
+        if not vector or any(not isinstance(value, (float, int)) or not isfinite(value) for value in vector):
+            raise config_error("The embedding model must return a non-empty, finite query vector.")
+        limit: Final = cls._limit(optional_params)
+        return (
+            f"{api_base}/v1/vector_stores/{quote(vector_store_id, safe='')}/search",
+            {  # mutable-ok: JSON transport requires a dict
+                "query": query_text,
+                "query_vector": tuple(vector),
+                "mongodb_database": params.require_database(),
+                "mongodb_collection": params.require_collection(),
+                "mongodb_embedding_field": params.embedding_field,
+                "mongodb_text_field": params.text_field,
+                "mongodb_num_candidates": cls._num_candidates(limit, params.mongodb_num_candidates),
+                "max_num_results": limit,
+                "timeout_ms": cls._timeout_ms(timeout),
+            },
         )
 
-    @classmethod
-    def _raise_for_missing_text_field(
-        cls, documents: Sequence[Mapping[str, object]], text_field: str, database: str, collection: str
-    ) -> None:
-        """$vectorSearch matches documents carrying no text, so a mistyped mongodb_text_field
-        returns well-scored results with empty content instead of failing."""
-        if documents and all(cls._field_value(document, text_field) is None for document in documents):
-            raise config_error(
-                f"None of the {len(documents)} matched documents in '{database}.{collection}' has a "
-                f"'{text_field}' field, so every result would carry empty text. Set mongodb_text_field "
-                "to the field holding the readable text; it accepts a dotted path such as metadata.body."
-            )
-
-    @classmethod
-    def _to_response(
-        cls, documents: Sequence[Mapping[str, object]], query_text: str, text_field: str
-    ) -> VectorStoreSearchResponse:
-        return VectorStoreSearchResponse(
-            object="vector_store.search_results.page",
-            search_query=query_text,
-            data=[  # mutable-ok: VectorStoreSearchResponse declares data as a list
-                cls._to_result(document, text_field) for document in documents
-            ],
-        )
-
-    @staticmethod
-    def _raise_for_unusable_index(
-        catalogue: Sequence[Mapping[str, object]], index_name: str, database: str, collection: str
-    ) -> None:
-        """mongod returns zero documents both for a query that matched nothing and for a missing
-        database, collection or index, so the catalogue decides which one happened."""
-        if not catalogue:
-            raise missing_index_error(index_name, database, collection)
-        entry: Final = catalogue[0]
-        if not entry.get("queryable"):
-            raise index_not_ready_error(index_name, database, collection, str(entry.get("status") or "unknown"))
-
-    @staticmethod
-    def _embedding_vector(embedding_response: EmbeddingResponse) -> Sequence[float]:
-        data: Final = embedding_response.data
-        if not data:
-            raise config_error(
-                "The embedding model returned no embedding for the search query, so there is nothing "
-                "to search MongoDB with. Check the embedding deployment named by litellm_embedding_model."
-            )
-        return data[0]["embedding"]
-
-    def execute_search_vector_store_request(
+    def transform_search_vector_store_request(
         self,
         vector_store_id: str,
         query: str | Sequence[str],
         vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
+        api_base: str,
         litellm_logging_obj: "LiteLLMLoggingObj",
         litellm_params: Mapping[str, object],
+        extra_body: Mapping[str, object] | None = None,
         embedding_executor: VectorStoreEmbeddingExecutor | None = None,
-        timeout: float | httpx.Timeout | None = None,
-    ) -> VectorStoreSearchResponse:
-        self._reject_unknown_params(litellm_params)
-        params: Final = _MongoDBSearchParams.model_validate(litellm_params)
+    ) -> tuple[str, dict[str, object]]:  # mutable-ok: the provider contract returns a writable JSON request body
+        params: Final = self._params(litellm_params, vector_store_search_optional_params, extra_body)
         query_text: Final = self._query_text(query)
-        key: Final = self._client_key(params, timeout)
-        database: Final = params.require_database()
-        collection: Final = params.require_collection()
-
-        embedding_response: Final = (embedding_executor or self.embedding_executor).embed(
-            params.require_embedding_model(),
+        response: Final = (embedding_executor or self.embedding_executor).embed(
+            params.require_embedding_model(), query_text, params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG
+        )
+        return self._request(
+            vector_store_id,
             query_text,
-            params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG,
-        )
-        pipeline: Final = self._pipeline(
-            vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params
+            params,
+            vector_store_search_optional_params,
+            api_base,
+            response,
+            litellm_params.get("timeout"),
         )
 
-        try:
-            client: Final = self.sync_client_factory(key)
-            target: Final = client[database][collection]  # pyright: ignore[reportIndexIssue]  # factory is typed as returning object so injected doubles are accepted
-            documents: Final = tuple(target.aggregate(pipeline))
-        except Exception as e:
-            raise translate_mongo_error(e, index_name=vector_store_id, database=database, collection=collection) from e
-        if not documents:
-            try:
-                catalogue: Final = tuple(target.list_search_indexes(vector_store_id))
-            except Exception as e:
-                raise translate_mongo_error(
-                    e, index_name=vector_store_id, database=database, collection=collection
-                ) from e
-            self._raise_for_unusable_index(catalogue, vector_store_id, database, collection)
-        self._raise_for_missing_text_field(documents, params.text_field, database, collection)
-        return self._to_response(documents, query_text, params.text_field)
-
-    async def aexecute_search_vector_store_request(
+    async def atransform_search_vector_store_request(
         self,
         vector_store_id: str,
         query: str | Sequence[str],
         vector_store_search_optional_params: VectorStoreSearchOptionalRequestParams,
+        api_base: str,
         litellm_logging_obj: "LiteLLMLoggingObj",
         litellm_params: Mapping[str, object],
+        extra_body: Mapping[str, object] | None = None,
         embedding_executor: VectorStoreEmbeddingExecutor | None = None,
-        timeout: float | httpx.Timeout | None = None,
-    ) -> VectorStoreSearchResponse:
-        self._reject_unknown_params(litellm_params)
-        params: Final = _MongoDBSearchParams.model_validate(litellm_params)
+    ) -> tuple[str, dict[str, object]]:  # mutable-ok: the provider contract returns a writable JSON request body
+        params: Final = self._params(litellm_params, vector_store_search_optional_params, extra_body)
         query_text: Final = self._query_text(query)
-        key: Final = self._client_key(params, timeout)
-        database: Final = params.require_database()
-        collection: Final = params.require_collection()
-
-        embedding_response: Final = await (embedding_executor or self.embedding_executor).aembed(
-            params.require_embedding_model(),
+        response: Final = await (embedding_executor or self.embedding_executor).aembed(
+            params.require_embedding_model(), query_text, params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG
+        )
+        return self._request(
+            vector_store_id,
             query_text,
-            params.litellm_embedding_config or _EMPTY_EMBEDDING_CONFIG,
-        )
-        pipeline: Final = self._pipeline(
-            vector_store_id, self._embedding_vector(embedding_response), params, vector_store_search_optional_params
+            params,
+            vector_store_search_optional_params,
+            api_base,
+            response,
+            litellm_params.get("timeout"),
         )
 
+    def transform_search_vector_store_response(
+        self, response: httpx.Response, litellm_logging_obj: "LiteLLMLoggingObj"
+    ) -> VectorStoreSearchResponse:
         try:
-            client: Final = self.async_client_factory(key)
-            target: Final = client[database][collection]  # pyright: ignore[reportIndexIssue]  # factory is typed as returning object so injected doubles are accepted
-            cursor: Final = await target.aggregate(pipeline)
-            documents: Final = [  # mutable-ok: an async comprehension cannot build a tuple directly
-                document async for document in cursor
-            ]
-        except Exception as e:
-            raise translate_mongo_error(e, index_name=vector_store_id, database=database, collection=collection) from e
-        if not documents:
-            try:
-                index_cursor: Final = await target.list_search_indexes(vector_store_id)
-                catalogue: Final = [  # mutable-ok: an async comprehension cannot build a tuple directly
-                    entry async for entry in index_cursor
-                ]
-            except Exception as e:
-                raise translate_mongo_error(
-                    e, index_name=vector_store_id, database=database, collection=collection
-                ) from e
-            self._raise_for_unusable_index(catalogue, vector_store_id, database, collection)
-        self._raise_for_missing_text_field(documents, params.text_field, database, collection)
-        return self._to_response(documents, query_text, params.text_field)
+            validated: Final = _SearchResponse.model_validate_json(response.content)
+            return _RESPONSE_ADAPTER.validate_python(validated.model_dump())
+        except ValidationError:
+            raise ServiceUnavailableError(
+                message="MongoDB sidecar returned an invalid search response. Check the sidecar version and deployment.",
+                model=None,
+                llm_provider="mongodb",
+            ) from None
+
+    def get_error_class(
+        self, error_message: str, status_code: int, headers: Mapping[str, object] | httpx.Headers
+    ) -> BaseLLMException:
+        if status_code == 400:
+            raise config_error(error_message)
+        if status_code == 401:
+            raise AuthenticationError(message="MongoDB sidecar rejected api_key.", model=None, llm_provider="mongodb")
+        if status_code == 408:
+            raise Timeout(message=error_message, model=None, llm_provider="mongodb")
+        raise ServiceUnavailableError(
+            message="MongoDB sidecar is unavailable. Check its address, health, and logs.",
+            model=None,
+            llm_provider="mongodb",
+        )
+
+    def validate_create_vector_store(self) -> NoReturn:
+        raise config_error(_SEARCH_ONLY_MESSAGE)
 
     def transform_create_vector_store_request(
-        self,
-        vector_store_create_optional_params: VectorStoreCreateOptionalRequestParams,
-        api_base: str,
+        self, vector_store_create_optional_params: VectorStoreCreateOptionalRequestParams, api_base: str
     ) -> NoReturn:
         raise config_error(_SEARCH_ONLY_MESSAGE)
 

litellm/llms/ollama/completion/transformation.py

@@ -16,6 +16,10 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     custom_prompt,
     ollama_pt,
 )
+from litellm.litellm_core_utils.prompt_templates.image_handling import (
+    async_inline_remote_media,
+    inline_remote_image_urls,
+)
 from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionUsageBlock
@@ -344,6 +348,26 @@ class OllamaConfig(BaseConfig):
         )
         return model_response
 
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return True
+
+    async def async_transform_request(
+        self,
+        model: str,
+        messages: list[AllMessageValues],  # mutable-ok: BaseConfig signature
+        optional_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        litellm_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        headers: dict[str, object],  # mutable-ok: BaseConfig signature
+    ) -> dict[str, object]:  # mutable-ok: BaseConfig signature
+        return self.transform_request(
+            model=model,
+            messages=await async_inline_remote_media(messages, should_inline=inline_remote_image_urls),
+            optional_params=optional_params,
+            litellm_params=litellm_params,
+            headers=headers,
+        )
+
     def transform_request(
         self,
         model: str,

litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/transformation.py

@@ -5,6 +5,7 @@ from typing import TYPE_CHECKING, Final
 import httpx
 
 import litellm
+from litellm.litellm_core_utils.prompt_templates.image_handling import RemoteMedia, inline_remote_image_urls
 from litellm.llms.base_llm.chat.transformation import LiteLLMLoggingObj
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import ModelResponse
@@ -51,6 +52,9 @@ class VertexAIAnthropicConfig(AnthropicConfig):
     def custom_llm_provider(self) -> str | None:
         return "vertex_ai"
 
+    def inlines_remote_media(self, media: RemoteMedia) -> bool:
+        return inline_remote_image_urls(media)
+
     def should_strip_billing_metadata(self) -> bool:
         return True
 

litellm/llms/watsonx/chat/transformation.py

@@ -157,11 +157,9 @@ class IBMWatsonXChatConfig(IBMWatsonXMixin, OpenAIGPTConfig):
     @staticmethod
     async def aapply_prompt_template(model: str, messages: list[dict[str, str]]) -> str | None:
         """Apply prompt template (async version)"""
-        import litellm
         from litellm.litellm_core_utils.prompt_templates.factory import (
             ahf_chat_template,
             custom_prompt,
-            hf_chat_template,
             ibm_granite_pt,
             mistral_instruct_pt,
         )
@@ -179,11 +177,7 @@ class IBMWatsonXChatConfig(IBMWatsonXMixin, OpenAIGPTConfig):
             else:
                 hf_model = model
             try:
-                # Use sync if cached, async if not
-                if hf_model in litellm.known_tokenizer_config:
-                    result = hf_chat_template(model=hf_model, messages=messages)
-                else:
-                    result = await ahf_chat_template(model=hf_model, messages=messages)
+                result = await ahf_chat_template(model=hf_model, messages=messages)
                 # Return result if it's truthy (not None and not empty string)
                 # The caller (_aconvert_watsonx_messages_core) will handle None/empty by falling back to default
                 if result:

litellm/llms/watsonx/completion/transformation.py

@@ -16,6 +16,7 @@ from ..common_utils import (
     IBMWatsonXMixin,
     WatsonXAIError,
     _get_api_params,
+    aconvert_watsonx_messages_to_prompt,
     convert_watsonx_messages_to_prompt,
 )
 
@@ -236,7 +237,11 @@ class IBMWatsonXAIConfig(IBMWatsonXMixin, BaseConfig):
             **watsonx_auth_payload,
         }
 
-    async def atransform_request(
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return True
+
+    async def async_transform_request(
         self,
         model: str,
         messages: list[AllMessageValues],
@@ -244,11 +249,6 @@ class IBMWatsonXAIConfig(IBMWatsonXMixin, BaseConfig):
         litellm_params: dict,
         headers: dict,
     ) -> dict:
-        """Async version of transform_request"""
-        from litellm.llms.watsonx.common_utils import (
-            aconvert_watsonx_messages_to_prompt,
-        )
-
         provider: Final = model.split("/")[0]
         prompt: Final = await aconvert_watsonx_messages_to_prompt(
             model=model, messages=messages, provider=provider, custom_prompt_dict={}

litellm/main.py

@@ -6545,7 +6545,7 @@ def embedding(
                 client=client,
                 timeout=timeout,
                 aembedding=aembedding,
-                litellm_params={},
+                litellm_params=litellm_params_dict,
                 api_base=api_base,
                 print_verbose=print_verbose,
                 extra_headers=headers,

litellm/model_prices_and_context_window_backup.json

@@ -650,7 +650,10 @@
     },
     "twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
         "litellm_provider": "bedrock",
         "max_input_tokens": 77,
         "max_tokens": 77,
@@ -662,7 +665,7 @@
     },
     "us.twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
         "input_cost_per_video_per_second": 0.0007,
         "input_cost_per_audio_per_second": 0.00014,
         "input_cost_per_image": 0.0001,
@@ -677,7 +680,7 @@
     },
     "eu.twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
         "input_cost_per_video_per_second": 0.0007,
         "input_cost_per_audio_per_second": 0.00014,
         "input_cost_per_image": 0.0001,
@@ -690,6 +693,48 @@
         "supports_embedding_image_input": true,
         "supports_image_input": true
     },
+    "twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
+    "us.twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
+    "eu.twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
     "twelvelabs.pegasus-1-2-v1:0": {
         "input_cost_per_video_per_second": 0.00049,
         "output_cost_per_token": 7.5e-06,

litellm/proxy/_experimental/mcp_server/server.py

@@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, Any, Final, Protocol
 
 import httpx
 from fastapi import FastAPI, HTTPException
-from pydantic import AnyUrl, ConfigDict
+from pydantic import AnyUrl, ConfigDict, TypeAdapter, ValidationError
 from starlette.requests import Request as StarletteRequest
 from starlette.responses import JSONResponse
 from starlette.types import Message, Receive, Scope, Send
@@ -108,9 +108,9 @@ _MAX_STATEFUL_SESSIONS_PER_OWNER: Final = 100
 # prevents an authenticated client from forcing the proxy to buffer an
 # arbitrarily large body just to make a routing decision.
 _MCP_ROUTING_PEEK_MAX_BYTES: Final = 4096
-# ASGI scope key holding the tracing span of the request carrying an MCP
-# message, written on the request task and read back by the message handler.
+# ASGI scope keys carrying OTel request state into a stateful MCP message handler.
 _MCP_TRANSPORT_SPAN_SCOPE_KEY: Final = "litellm_otel_transport_span"
+_MCP_DESTINATIONS_SCOPE_KEY: Final = "litellm_otel_request_destinations"
 
 
 def _invalidate_byok_cred_cache(user_id: str, server_id: str) -> None:
@@ -328,18 +328,17 @@ def _otel_publish_transport_span_on_scope(scope: Scope) -> None:
         scope[_MCP_TRANSPORT_SPAN_SCOPE_KEY] = span
 
 
-def _otel_transport_span_from_message(req_ctx: object) -> object:
-    """The tracing span of the HTTP request that carried this MCP message.
-
-    Read off that request's ASGI scope, reached through the ``Request`` the
-    streamable-HTTP transport attaches to each message, so it is this message's
-    transport and not whichever request happens to have touched the session last.
-    Returns whatever the scope holds; the otel plumbing validates it."""
+def _otel_value_from_message_scope(req_ctx: object, key: str) -> object:
     request: Final = getattr(req_ctx, "request", None)
     scope: Final = getattr(request, "scope", None)
     if not isinstance(scope, Mapping):
         return None
-    return scope.get(_MCP_TRANSPORT_SPAN_SCOPE_KEY)
+    return scope.get(key)
+
+
+def _otel_transport_span_from_message(req_ctx: object) -> object:
+    """The tracing span of the HTTP request that carried this MCP message."""
+    return _otel_value_from_message_scope(req_ctx, _MCP_TRANSPORT_SPAN_SCOPE_KEY)
 
 
 def _otel_set_mcp_transport_span(span: object) -> object:
@@ -372,6 +371,44 @@ def _otel_reset_mcp_transport_span(token: object) -> None:
         return
 
 
+def _otel_publish_request_destinations_on_scope(scope: Scope) -> None:
+    try:
+        from litellm.integrations.otel.plumbing.context import request_destinations
+
+        scope[_MCP_DESTINATIONS_SCOPE_KEY] = request_destinations()
+    except ImportError:
+        return
+
+
+def _otel_set_mcp_request_destinations(req_ctx: object) -> object:
+    destinations: Final = _otel_value_from_message_scope(req_ctx, _MCP_DESTINATIONS_SCOPE_KEY)
+    if not isinstance(destinations, tuple):
+        return None
+    try:
+        from litellm.integrations.otel.model.destination import OtelDestination
+        from litellm.integrations.otel.plumbing.context import set_request_destinations
+
+        destination_adapter: Final[TypeAdapter[tuple[OtelDestination, ...]]] = TypeAdapter(
+            tuple[OtelDestination, ...],
+            config=ConfigDict(revalidate_instances="always"),
+        )
+        validated_destinations: Final = destination_adapter.validate_python(destinations, strict=True)
+        return set_request_destinations(validated_destinations)
+    except (ImportError, ValidationError):
+        return None
+
+
+def _otel_reset_mcp_request_destinations(token: object) -> None:
+    if token is None:
+        return
+    try:
+        from litellm.integrations.otel.plumbing.context import reset_request_destinations
+
+        reset_request_destinations(token)
+    except ImportError:
+        return
+
+
 def _proxy_exception_to_http_exception(exc: ProxyException) -> HTTPException:
     """Map a ``ProxyException`` to an ``HTTPException`` that preserves its real
     status code and headers.
@@ -763,10 +800,12 @@ if MCP_AVAILABLE:
             _session_reset_token = active_mcp_session_var.set(req_ctx.session)
         _trace_token = None
         _transport_token = None
+        _destinations_token = None
 
         try:
             _trace_token = _otel_set_mcp_trace_carrier(_mcp_meta_trace_carrier(req_ctx))
             _transport_token = _otel_set_mcp_transport_span(_otel_transport_span_from_message(req_ctx))
+            _destinations_token = _otel_set_mcp_request_destinations(req_ctx)
             # Get user authentication from context variable
             (
                 user_api_key_auth,
@@ -828,6 +867,7 @@ if MCP_AVAILABLE:
             # This prevents the HTTP stream from failing and allows the client to get a response
             return []
         finally:
+            _otel_reset_mcp_request_destinations(_destinations_token)
             _otel_reset_mcp_transport_span(_transport_token)
             _otel_reset_mcp_trace_carrier(_trace_token)
             if _session_reset_token is not None:
@@ -1021,10 +1061,12 @@ if MCP_AVAILABLE:
             _session_reset_token = active_mcp_session_var.set(req_ctx.session)
         _trace_token = None
         _transport_token = None
+        _destinations_token = None
 
         try:
             _trace_token = _otel_set_mcp_trace_carrier(_mcp_meta_trace_carrier(req_ctx))
             _transport_token = _otel_set_mcp_transport_span(_otel_transport_span_from_message(req_ctx))
+            _destinations_token = _otel_set_mcp_request_destinations(req_ctx)
             # Validate arguments
             (
                 user_api_key_auth,
@@ -1163,6 +1205,7 @@ if MCP_AVAILABLE:
 
             return response
         finally:
+            _otel_reset_mcp_request_destinations(_destinations_token)
             _otel_reset_mcp_transport_span(_transport_token)
             _otel_reset_mcp_trace_carrier(_trace_token)
             if _session_reset_token is not None:
@@ -4493,6 +4536,7 @@ if MCP_AVAILABLE:
 
             async def _dispatch() -> None:
                 _otel_publish_transport_span_on_scope(scope)
+                _otel_publish_request_destinations_on_scope(scope)
                 auth_user: Final = _set_or_update_auth_context(
                     user_api_key_auth=user_api_key_auth,
                     mcp_auth_header=mcp_auth_header,

litellm/proxy/_types.py

@@ -835,6 +835,14 @@ class LiteLLMRoutes(enum.Enum):
         "/team/daily/activity/aggregated",
         "/team/spend/by_user",
         "/team/{team_id}/members/me",
+        # POST/GET the team's logging callbacks, and DELETE one of them. Every
+        # handler calls _verify_team_access, which admits only a proxy admin, an
+        # org admin for the team, or an admin of this team.
+        #
+        # team_id is a free-form string, so it spells these with the same path
+        # converter the router uses; the gate matches that converter.
+        "/team/{team_id:path}/callback",
+        "/team/{team_id:path}/callback/{callback_name}",
         "/model/new",
         "/model/update",
         "/model/delete",
@@ -3584,6 +3592,7 @@ class AllCallbacks(LiteLLMPydanticObjectBase):
         ui_callback_name="OpenTelemetry",
         litellm_callback_params=[
             "OTEL_EXPORTER",
+            "OTEL_EXPORTER_OTLP_PROTOCOL",
             "OTEL_ENDPOINT",
             "OTEL_TRACES_ENDPOINT",
             "OTEL_HEADERS",

litellm/proxy/agent_endpoints/a2a_endpoints.py

@@ -144,31 +144,32 @@ def _validate_push_notification_url(url: str) -> None:
         raise HTTPException(status_code=400, detail=str(e)) from e
 
 
-def _caller_identity_headers(user_api_key_dict: UserAPIKeyAuth) -> dict[str, str]:
-    headers: Final[dict[str, str]] = {}
-    if user_api_key_dict.user_id:
-        headers["X-LiteLLM-User-Id"] = user_api_key_dict.user_id
-    if user_api_key_dict.team_id:
-        headers["X-LiteLLM-Team-Id"] = user_api_key_dict.team_id
-    return headers
+def _caller_identity_headers(user_api_key_dict: UserAPIKeyAuth) -> Mapping[str, str]:
+    return MappingProxyType(
+        {
+            name: value
+            for name, value in (
+                ("X-LiteLLM-User-Id", user_api_key_dict.user_id),
+                ("X-LiteLLM-Team-Id", user_api_key_dict.team_id),
+            )
+            if value
+        }
+    )
 
 
 def _forwarding_headers(
-    user_api_key_dict: UserAPIKeyAuth,
+    caller_identity: Mapping[str, str],
     request_data: Mapping[str, object],
     agent_extra_headers: Mapping[str, str] | None,
-) -> Mapping[str, str] | None:
-    sanitized: Final = (
-        {k: v for k, v in agent_extra_headers.items() if not k.lower().startswith("x-litellm-")}
-        if agent_extra_headers
-        else None
+) -> dict[str, str] | None:
+    passthrough: Final = tuple(
+        (name, value)
+        for name, value in (agent_extra_headers.items() if agent_extra_headers else ())
+        if not name.lower().startswith("x-litellm-")
     )
-    merged: Final = merge_agent_headers(dynamic_headers=sanitized, static_headers=None) or {}
-    identity: Final = _caller_identity_headers(user_api_key_dict)
     trace_id: Final = request_data.get("litellm_trace_id")
-    if trace_id:
-        identity["X-LiteLLM-Trace-Id"] = str(trace_id)
-    merged.update(identity)
+    trace: Final = (("X-LiteLLM-Trace-Id", str(trace_id)),) if trace_id else ()
+    merged: Final = dict((*passthrough, *caller_identity.items(), *trace))
     return merged or None
 
 
@@ -755,6 +756,7 @@ async def invoke_agent_a2a(
             ProxyBaseLLMRequestProcessing,
         )
 
+        caller_identity: Final = _caller_identity_headers(user_api_key_dict)
         processor: Final = ProxyBaseLLMRequestProcessing(data=body)
         data, logging_obj = await processor.common_processing_pre_call_logic(
             request=request,
@@ -793,9 +795,13 @@ async def invoke_agent_a2a(
                     if header_name:
                         dynamic_headers[header_name] = val
 
-        agent_extra_headers = merge_agent_headers(
-            dynamic_headers=dynamic_headers or None,
-            static_headers=static_headers or None,
+        agent_extra_headers = _forwarding_headers(
+            caller_identity=caller_identity,
+            request_data=data,
+            agent_extra_headers=merge_agent_headers(
+                dynamic_headers=dynamic_headers or None,
+                static_headers=static_headers or None,
+            ),
         )
 
         # Databricks App endpoints require a short-lived OAuth M2M token rather
@@ -942,12 +948,7 @@ async def invoke_agent_a2a(
                 "method": method,
                 "params": params,
             }
-            caller_headers: Final = _forwarding_headers(
-                user_api_key_dict=user_api_key_dict,
-                request_data=data,
-                agent_extra_headers=agent_extra_headers,
-            )
-            result = await _forward_jsonrpc(agent_url, forward_body, extra_headers=caller_headers)
+            result = await _forward_jsonrpc(agent_url, forward_body, extra_headers=agent_extra_headers)
             if method == "agent/getAuthenticatedExtendedCard":
                 card: Final = result.get("result")
                 if isinstance(card, dict):
@@ -988,16 +989,11 @@ async def invoke_agent_a2a(
                 "method": method,
                 "params": params,
             }
-            sse_caller_headers: Final = _forwarding_headers(
-                user_api_key_dict=user_api_key_dict,
-                request_data=data,
-                agent_extra_headers=agent_extra_headers,
-            )
             return await _forward_jsonrpc_sse(
                 agent_url,
                 forward_body,
                 request_id=request_id,
-                extra_headers=sse_caller_headers,
+                extra_headers=agent_extra_headers,
                 proxy_logging_obj=proxy_logging_obj,
                 user_api_key_dict=user_api_key_dict,
                 request_data=data,

litellm/proxy/anthropic_endpoints/endpoints.py

@@ -25,6 +25,11 @@ from litellm.proxy.common_request_processing import (
     proxy_exception_from_http_exception,
 )
 from litellm.proxy.common_utils.http_parsing_utils import _read_request_body
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.types.utils import TokenCountResponse
 
 router: Final = APIRouter()
@@ -243,9 +248,9 @@ async def anthropic_response(
         return _anthropic_error_json_response(
             ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
                 headers=headers,
             ),
             request,

litellm/proxy/auth/route_checks.py

@@ -497,10 +497,22 @@ class RouteChecks:
 
         def _placeholder_to_regex(match: re.Match) -> str:
             placeholder: Final = match.group(0).strip("{}")
-            if placeholder.endswith(":path"):
-                # allow "/" in the placeholder value, but don't eat the route suffix after ":"
-                return r"[^:]+"
-            return r"[^/]+"
+            if not placeholder.endswith(":path"):
+                return r"[^/]+"
+            # A ":path" placeholder takes whatever the router's own path
+            # converter takes, slashes and colons alike, so an id spelled with
+            # either (or both) still matches the template it was mounted under.
+            #
+            # Unless the template puts a ":" literal of its own after the
+            # placeholder: the Google routes end in ":generateContent" and
+            # friends, and there the value has to stop before that suffix
+            # rather than swallow it and match a different verb.
+            #
+            # "[\s\S]" rather than ".", because "." stops at a newline and the
+            # path converter does not: a %0A anywhere in the value would leave
+            # the route unmatched here while still reaching the handler, which
+            # turns this gate into a bypass for the lists built on it.
+            return r"[^:]+" if ":" in match.string[match.end() :] else r"[\s\S]+"
 
         pattern = re.sub(r"\{[^}]+\}", _placeholder_to_regex, pattern)
         # Anchor the pattern to match the entire string

litellm/proxy/auth/user_api_key_auth.py

@@ -2837,6 +2837,43 @@ async def _authorize_authenticated_request(
 
 
 @tracer.wrap()
+def _seed_request_destinations(user_api_key_dict: UserAPIKeyAuth, request: Request | None = None) -> None:
+    """Anchor the OTLP destinations this key or team overrides its traces to.
+
+    Called inside the ``auth`` phase span so that span reaches the tenant's account
+    as well, and on the request task so the ``ContextVar`` is inherited by the logging
+    tasks that close the LLM span. Best-effort: trace routing must never fail auth.
+
+    ``request`` carries the headers, so a backend this request disabled with
+    ``x-litellm-disable-callbacks`` resolves to no destination.
+
+    Only destinations the published fan-out can build are anchored. Anchoring one is
+    what tells the operator's exporter to hold that backend's spans back under
+    ``override``, so an unbuildable one would leave the span with nowhere to go.
+
+    The ``postgres`` spans under ``auth`` close before this runs, because they are the
+    reads that resolve the identity being read here. They never reach the tenant's
+    account, and they are never withheld from the operator's backend, whichever mode
+    is set.
+    """
+    try:
+        from litellm.integrations.otel.logger import fan_out_provider
+        from litellm.integrations.otel.plumbing.context import set_request_destinations
+        from litellm.integrations.otel.plumbing.providers import deliverable_destinations
+        from litellm.proxy.litellm_pre_call_utils import (
+            resolve_tenant_otel_destinations,
+        )
+
+        set_request_destinations(
+            deliverable_destinations(
+                resolve_tenant_otel_destinations(user_api_key_dict, _safe_get_request_headers(request)),
+                fan_out_provider(),
+            )
+        )
+    except Exception as exc:  # noqa: BLE001  # telemetry routing is best-effort and must never break authentication
+        verbose_proxy_logger.debug("OTel V2: tenant destination resolution failed: %s", exc)
+
+
 async def user_api_key_auth(
     request: Request,
     api_key: str = fastapi.Security(api_key_header),
@@ -2883,6 +2920,7 @@ async def user_api_key_auth(
                 raise body_parse_exception
             raise
         user_api_key_auth_obj.budget_reservation = None
+        _seed_request_destinations(user_api_key_auth_obj, request)
 
         # A body that never parsed is authenticated (so the trace carries identity
         # and this ``auth`` span) but not authorized: there is no model to check it

litellm/proxy/common_request_processing.py

@@ -54,6 +54,12 @@ from litellm.proxy.common_utils.callback_utils import (
     get_logging_caching_headers,
     get_remaining_tokens_and_requests_from_request_data,
 )
+from litellm.proxy.common_utils.openai_error_payload import (
+    attribute_of,
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.proxy.common_utils.sse_keepalive import (
     SSE_COMMENT_PING_BYTES,
     coerce_keepalive_interval,
@@ -464,46 +470,6 @@ def _stream_usage_tracking_updates(
     }
 
 
-def _getattr_object(value: object, name: str, default: object = None) -> object:
-    return getattr(value, name, default)
-
-
-_OPENAI_ERROR_TYPE_BY_STATUS: Final[Mapping[int, str]] = MappingProxyType(
-    {
-        status.HTTP_401_UNAUTHORIZED: "authentication_error",
-        status.HTTP_403_FORBIDDEN: "permission_error",
-        status.HTTP_429_TOO_MANY_REQUESTS: "rate_limit_error",
-    }
-)
-
-
-def _error_status_code(exc: object, default: int) -> int:
-    """The HTTP status an exception carries, or ``default`` when it carries none."""
-    carried: Final = _getattr_object(exc, "status_code")
-    return carried if isinstance(carried, int) and not isinstance(carried, bool) else default
-
-
-def _openai_error_type(exc: object, status_code: int) -> str:
-    """OpenAI types ``error.type`` as a required string, so an exception carrying none
-    falls back to the type its status code stands for."""
-    carried: Final = _getattr_object(exc, "type")
-    if isinstance(carried, str):
-        return carried
-    mapped: Final = _OPENAI_ERROR_TYPE_BY_STATUS.get(status_code)
-    if mapped is not None:
-        return mapped
-    if status_code < status.HTTP_500_INTERNAL_SERVER_ERROR:
-        return "invalid_request_error"
-    return "internal_server_error"
-
-
-def _openai_error_param(exc: object) -> str | None:
-    """OpenAI types ``error.param`` as nullable, so an exception carrying none
-    serializes as JSON ``null``."""
-    carried: Final = _getattr_object(exc, "param")
-    return carried if isinstance(carried, str) else None
-
-
 class _UpstreamHttpResponse(Protocol):
     @property
     def status_code(self) -> int: ...
@@ -573,15 +539,15 @@ def serialize_http_exception_detail(
 
 
 def proxy_exception_from_http_exception(exc: HTTPException, headers: dict[str, str]) -> ProxyException:
-    raw_detail: Final = _getattr_object(exc, "detail", str(exc))
+    raw_detail: Final = attribute_of(exc, "detail", str(exc))
     message, structured_fields = serialize_http_exception_detail(raw_detail)
     existing_fields: Final = getattr(exc, "provider_specific_fields", None) or {}
     merged_fields: Final = {**existing_fields, **structured_fields} if structured_fields else (existing_fields or None)
-    error_status: Final = _error_status_code(exc, status.HTTP_400_BAD_REQUEST)
+    error_status: Final = error_status_code(exc, status.HTTP_400_BAD_REQUEST)
     return ProxyException(
         message=message,
-        type=_openai_error_type(exc, error_status),
-        param=_openai_error_param(exc),
+        type=openai_error_type(exc, error_status),
+        param=openai_error_param(exc),
         code=error_status,
         provider_specific_fields=merged_fields,
         headers=headers,
@@ -865,8 +831,8 @@ def sse_error_payload(exc: BaseException) -> tuple[int, Mapping[str, object]]:
     are byte-identical.
     """
     # Preserve status code from HTTPException (e.g. guardrail blocks)
-    error_status: Final = _error_status_code(exc, status.HTTP_500_INTERNAL_SERVER_ERROR)
-    raw_detail: Final = _getattr_object(exc, "detail", "Error processing stream start")
+    error_status: Final = error_status_code(exc, status.HTTP_500_INTERNAL_SERVER_ERROR)
+    raw_detail: Final = attribute_of(exc, "detail", "Error processing stream start")
     message, structured_fields = serialize_http_exception_detail(raw_detail)
 
     existing_fields: Final = getattr(exc, "provider_specific_fields", None) or {}
@@ -874,8 +840,8 @@ def sse_error_payload(exc: BaseException) -> tuple[int, Mapping[str, object]]:
 
     error_obj: Final = {
         "message": message,
-        "type": _openai_error_type(exc, error_status),
-        "param": _openai_error_param(exc),
+        "type": openai_error_type(exc, error_status),
+        "param": openai_error_param(exc),
         "code": str(error_status),
     }
     if not merged_fields:
@@ -2815,10 +2781,10 @@ class ProxyBaseLLMRequestProcessing:
         ``ResponsesAPIResponse`` directly. Handle both shapes so the
         container-ownership recording path can walk ``.output`` either way.
         """
-        completed: Final = _getattr_object(stream_response, "completed_response")
+        completed: Final = attribute_of(stream_response, "completed_response")
         if completed is None:
             return None
-        response_obj: Final = _getattr_object(completed, "response")
+        response_obj: Final = attribute_of(completed, "response")
         if response_obj is not None:
             return response_obj
         return completed
@@ -3468,7 +3434,7 @@ class ProxyBaseLLMRequestProcessing:
         headers = getattr(e, "headers", None) or {}
         if not headers:
             # Try to get headers from e.response.headers (httpx.Response)
-            _response: Final = _getattr_object(e, "response")
+            _response: Final = attribute_of(e, "response")
             if _response is not None:
                 _response_headers: Final = getattr(_response, "headers", None)
                 if _response_headers:
@@ -3543,8 +3509,8 @@ class ProxyBaseLLMRequestProcessing:
             _code = status.HTTP_500_INTERNAL_SERVER_ERROR
         raise ProxyException(
             message=redact_internal_details_from_client_message(getattr(e, "message", error_msg)),
-            type=_openai_error_type(e, _code),
-            param=_openai_error_param(e),
+            type=openai_error_type(e, _code),
+            param=openai_error_param(e),
             openai_code=getattr(e, "code", None),
             code=_code,
             provider_specific_fields=getattr(e, "provider_specific_fields", None),
@@ -3754,11 +3720,11 @@ class ProxyBaseLLMRequestProcessing:
 
             if isinstance(e, HTTPException):
                 raise e
-            stream_error_status: Final = _error_status_code(e, status.HTTP_500_INTERNAL_SERVER_ERROR)
+            stream_error_status: Final = error_status_code(e, status.HTTP_500_INTERNAL_SERVER_ERROR)
             proxy_exception: Final = ProxyException(
                 message=redact_internal_details_from_client_message(getattr(e, "message", str(e))),
-                type=_openai_error_type(e, stream_error_status),
-                param=_openai_error_param(e),
+                type=openai_error_type(e, stream_error_status),
+                param=openai_error_param(e),
                 code=stream_error_status,
             )
             stream_completed = True

litellm/proxy/common_utils/callback_config_validation.py

@@ -44,6 +44,91 @@ def _langfuse_environment_error(callback_vars: Mapping[str, str]) -> str | None:
     return None
 
 
+# Which credential family a dynamic variable belongs to. The families are the
+# integrations that share one account: every langfuse_* variable configures the
+# same Langfuse project whether it rides the classic callback or the OTel one,
+# and every dd_* variable configures the same Datadog account.
+_VAR_FAMILIES: Final[Mapping[str, str]] = MappingProxyType(
+    {
+        "arize_": "Arize",
+        "dd_": "Datadog",
+        "gcs_": "GCS",
+        "humanloop_": "Humanloop",
+        "langfuse_": "Langfuse",
+        "langsmith_": "LangSmith",
+        "newrelic_": "New Relic",
+        "posthog_": "PostHog",
+        "wandb_": "Weights & Biases",
+        "weave_": "Weights & Biases",
+    }
+)
+
+
+def _family_of(var: str) -> str | None:
+    """The credential family ``var`` configures, or ``None`` if it configures none.
+
+    ``turn_off_message_logging`` and friends belong to no backend, so they carry
+    no credentials anyone could redirect.
+    """
+    return next((family for prefix, family in _VAR_FAMILIES.items() if var.startswith(prefix)), None)
+
+
+def cross_entry_family_error(
+    callback_vars: Mapping[str, str] | None,
+    stored_vars_by_entry: Sequence[Mapping[str, str]],
+) -> str | None:
+    """Reject an entry that changes what a family another entry holds resolves to.
+
+    Every stored entry's variables are flattened into one dict before a request
+    reads them, and the flattened dict is what the exporter authenticates and
+    addresses with. So an entry naming only a destination is enough to redirect
+    credentials that were written somewhere else: a host on a second entry pairs
+    with the key from the first, and the request carries that key to the new
+    host.
+
+    Two rules together keep the flattened dict out of the caller's hands. A
+    variable the family already configures has to keep the value it has, so
+    nothing already in use can be moved. A variable the family does not yet
+    configure may only carry a value the family already holds, which is what lets
+    the same credential go in under its other spelling (``langfuse_secret`` and
+    ``langfuse_secret_key`` are one key) without anything here having to list the
+    spellings. Between them, no value the caller chose can enter the family, and
+    repeating the family as it stands is still allowed -- that is how one
+    integration gets registered for both the success and the failure event.
+
+    A team admin who does want to move a family deletes the entry holding it
+    first, which reveals nothing.
+
+    Only the writers this endpoint newly admits are held to this, because a proxy
+    admin already holds every credential the proxy has.
+
+    ``stored_vars_by_entry`` has to arrive decrypted; the credential values are
+    encrypted at rest and ciphertext never equals the plaintext coming in.
+    """
+    if not callback_vars:
+        return None
+    stored_by_var: Final = {
+        var: value for entry in stored_vars_by_entry for var, value in entry.items() if _family_of(var) is not None
+    }
+    family_values: Final = frozenset(
+        (family, value)
+        for entry in stored_vars_by_entry
+        for var, value in entry.items()
+        if (family := _family_of(var)) is not None
+    )
+    held_families: Final = frozenset(family for family, _ in family_values)
+    return next(
+        (
+            f"{family} is already configured by another callback entry on this team. "
+            f"Remove that entry before setting {var} here."
+            for var, value, family in ((v, callback_vars[v], _family_of(v)) for v in callback_vars)
+            if family in held_families
+            and (stored_by_var[var] != value if var in stored_by_var else (family, value) not in family_values)
+        ),
+        None,
+    )
+
+
 def logging_metadata_config_error(metadata: Mapping[str, object] | None) -> str | None:
     """Validate every ``logging`` entry of a team/key metadata payload."""
     if not metadata:

litellm/proxy/common_utils/openai_error_payload.py

@@ -0,0 +1,52 @@
+"""Shapes the ``error`` object the proxy answers with so it matches OpenAI's contract:
+``type`` is a required string and ``param`` is nullable, neither of which the literal
+string ``"None"`` satisfies."""
+
+from collections.abc import Mapping
+from types import MappingProxyType
+from typing import Final
+
+from fastapi import status
+
+_OPENAI_ERROR_TYPE_BY_STATUS: Final[Mapping[int, str]] = MappingProxyType(
+    {
+        status.HTTP_401_UNAUTHORIZED: "authentication_error",
+        status.HTTP_403_FORBIDDEN: "permission_error",
+        status.HTTP_429_TOO_MANY_REQUESTS: "rate_limit_error",
+    }
+)
+
+
+def attribute_of(value: object, name: str, default: object = None) -> object:
+    return getattr(value, name, default)
+
+
+def error_status_code(exc: object, default: int) -> int:
+    """The HTTP status an exception carries as ``status_code`` or, the way ``ProxyException``
+    stores it, as a stringified ``code``; ``default`` when it carries neither."""
+    carried: Final = attribute_of(exc, "status_code")
+    if isinstance(carried, int) and not isinstance(carried, bool):
+        return carried
+    stringified: Final = attribute_of(exc, "code")
+    return int(stringified) if isinstance(stringified, str) and stringified.isdecimal() else default
+
+
+def openai_error_type(exc: object, status_code: int) -> str:
+    """OpenAI types ``error.type`` as a required string, so an exception carrying none
+    falls back to the type its status code stands for."""
+    carried: Final = attribute_of(exc, "type")
+    if isinstance(carried, str):
+        return carried
+    mapped: Final = _OPENAI_ERROR_TYPE_BY_STATUS.get(status_code)
+    if mapped is not None:
+        return mapped
+    if status_code < status.HTTP_500_INTERNAL_SERVER_ERROR:
+        return "invalid_request_error"
+    return "internal_server_error"
+
+
+def openai_error_param(exc: object) -> str | None:
+    """OpenAI types ``error.param`` as nullable, so an exception carrying none
+    serializes as JSON ``null``."""
+    carried: Final = attribute_of(exc, "param")
+    return carried if isinstance(carried, str) else None

litellm/proxy/db/db_spend_update_writer.py

@@ -1063,7 +1063,7 @@ class DBSpendUpdateWriter:
 
             await enqueue_spend_logs(prisma_client, (payload,))
             if payload.get("call_type") in RESPONSES_SESSION_CALL_TYPES:
-                request_spend_log_flush()
+                request_spend_log_flush(prisma_client)
         else:
             verbose_proxy_logger.debug("prisma_client is None. Skipping writing spend logs to db.")
 

litellm/proxy/db/db_url_settings.py

@@ -64,6 +64,7 @@ DISABLE_PREPARED_STATEMENTS_ENV_VAR: Final = "DATABASE_DISABLE_PREPARED_STATEMEN
 DisablePreparedStatementsFlag = Annotated[
     bool, BeforeValidator(partial(token_auth_flag_enabled, env_var=DISABLE_PREPARED_STATEMENTS_ENV_VAR))
 ]
+MAX_IDLE_CONNECTION_LIFETIME_ENV_VAR: Final = "DATABASE_MAX_IDLE_CONNECTION_LIFETIME"
 
 # schema.prisma pins `provider = "postgresql"`, so these are the only schemes
 # Prisma can actually connect with.
@@ -217,6 +218,9 @@ class DatabaseURLSettings(BaseSettings):
     disable_prepared_statements: DisablePreparedStatementsFlag = Field(
         default=False, validation_alias=DISABLE_PREPARED_STATEMENTS_ENV_VAR
     )
+    max_idle_connection_lifetime: int | None = Field(
+        default=None, validation_alias=MAX_IDLE_CONNECTION_LIFETIME_ENV_VAR
+    )
 
     # Writer
     database_url: str | None = Field(default=None, validation_alias="DATABASE_URL")
@@ -453,6 +457,12 @@ class DatabaseURLSettings(BaseSettings):
                 if url:
                     os.environ[env_var] = add_missing_query_params(url, MappingProxyType({"pgbouncer": "true"}))
 
+        lifetime_params: Final = idle_lifetime_params(self.max_idle_connection_lifetime)
+        for env_var in ("DATABASE_URL", "DIRECT_URL"):
+            url = os.environ.get(env_var)
+            if url:
+                os.environ[env_var] = add_missing_query_params(url, lifetime_params)
+
         # The reader inherits the writer's connection params (pool size, timeouts,
         # pgbouncer mode). Without this the reader pool ignores the configured cap
         # and falls back to Prisma's `num_physical_cpus * 2 + 1` default.

litellm/proxy/db/prisma_client.py

@@ -16,6 +16,7 @@ from datetime import datetime, timedelta
 from typing import Any, Final, Protocol
 
 from litellm._logging import verbose_proxy_logger
+from litellm.proxy.db.db_url_settings import add_missing_query_params, connection_params_from_url
 from litellm.proxy.db.token_auth import (
     DEFAULT_POSTGRES_PORT,
     DatabaseTokenAuth,
@@ -438,7 +439,10 @@ class PrismaWrapper:
             return None
 
         endpoint: Final = self._iam_endpoint if self._iam_endpoint is not None else self._endpoint_from_env()
-        db_url: Final = endpoint.build_url(mint_database_token(auth, endpoint))
+        db_url: Final = add_missing_query_params(
+            endpoint.build_url(mint_database_token(auth, endpoint)),
+            connection_params_from_url(os.environ.get(self._db_url_env_var, "")),
+        )
         os.environ[self._db_url_env_var] = db_url
         return db_url
 

litellm/proxy/image_endpoints/endpoints.py

@@ -20,6 +20,11 @@ from litellm.proxy.common_utils.http_parsing_utils import (
     coerce_numeric_form_fields,
     numeric_form_fields,
 )
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.proxy.route_llm_request import route_request
 from litellm.types.images.main import ImageEditRequestParams
 from litellm.types.llms.openai import ChatCompletionUserMessage
@@ -200,18 +205,18 @@ async def image_generation(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
                 openai_code=getattr(e, "code", None),
-                code=getattr(e, "status_code", 500),
+                code=error_status_code(e, 500),
             )
 
 

litellm/proxy/litellm_pre_call_utils.py

@@ -10,6 +10,7 @@ from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, cast
 
 from fastapi import HTTPException, Request
+from pydantic import TypeAdapter
 from pydantic import ValidationError as PydanticValidationError
 from starlette.datastructures import Headers
 
@@ -28,6 +29,7 @@ from litellm.constants import (
     SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
     SESSION_ID_GENERATED_METADATA_KEY,
     SESSION_ID_OMITTED_METADATA_KEY,
+    X_LITELLM_DISABLE_CALLBACKS,
 )
 from litellm.litellm_core_utils.credential_accessor import CredentialAccessor
 from litellm.litellm_core_utils.initialize_dynamic_callback_params import (
@@ -159,6 +161,7 @@ from litellm.types.utils import (
     CustomPricingLiteLLMParams,
     LlmProviders,
     ProviderSpecificHeader,
+    StandardCallbackDynamicParams,
     StandardLoggingUserAPIKeyMetadata,
     SupportedCacheControls,
 )
@@ -172,6 +175,7 @@ _ENABLE_TEAM_STALE_ALIAS_BYPASS: bool | None = None
 
 
 if TYPE_CHECKING:
+    from litellm.integrations.otel.model.destination import OtelDestination
     from litellm.proxy.proxy_server import ProxyConfig as _ProxyConfig
     from litellm.types.proxy.policy_engine import Policy, PolicyMatchContext
 
@@ -975,6 +979,142 @@ def _get_dynamic_logging_metadata(
     return callback_settings_obj
 
 
+_TENANT_OTEL_PARAMS: Final = TypeAdapter(StandardCallbackDynamicParams)
+
+
+def _tenant_otel_params(callback_vars: Mapping[str, str]) -> StandardCallbackDynamicParams:
+    try:
+        return _TENANT_OTEL_PARAMS.validate_python(callback_vars)
+    except PydanticValidationError:
+        return StandardCallbackDynamicParams()
+
+
+_NO_REQUEST_HEADERS: Final[Mapping[str, str]] = MappingProxyType({})
+
+
+def _dynamically_disabled_backends(
+    user_api_key_dict: UserAPIKeyAuth,
+    request_headers: Mapping[str, str] | None,
+) -> frozenset[str]:
+    """The callbacks this request turned off, read the way dispatch reads them.
+
+    Same sources, precedence, and premium gate ``EnterpriseCallbackControls`` applies
+    before it skips a callback: the ``x-litellm-disable-callbacks`` header wins over the
+    key's stored list, team settings are not a source, and a non-premium proxy honours
+    neither. A destination has to agree with that decision, or a backend the key turned
+    off would still be exported to, now through the fan-out instead of the callback.
+    """
+    from litellm.proxy.proxy_server import premium_user
+
+    if litellm.allow_dynamic_callback_disabling is not True or not premium_user:
+        return frozenset()
+    header: Final = (request_headers if request_headers is not None else _NO_REQUEST_HEADERS).get(
+        X_LITELLM_DISABLE_CALLBACKS
+    )
+    if header is not None:
+        return frozenset(name.strip().lower() for name in header.split(","))
+    metadata: Final = user_api_key_dict.metadata
+    disabled: Final = metadata.get("litellm_disabled_callbacks") if metadata else None
+    if not isinstance(disabled, list):
+        return frozenset()
+    return frozenset(name.lower() for name in disabled if isinstance(name, str))
+
+
+def resolve_tenant_otel_destinations(
+    user_api_key_dict: UserAPIKeyAuth,
+    request_headers: Mapping[str, str] | None = None,
+) -> "tuple[OtelDestination, ...]":
+    """The OTLP destinations this request's key or team config overrides its traces to.
+
+    Key settings win over team settings outright, the same precedence
+    ``_get_dynamic_logging_metadata`` applies, so one caller never exports the same
+    backend to two accounts. An empty key-level list counts as configured, since that
+    is what disabling a key's callbacks writes. Returns empty when OTEL V2 is off, when
+    neither level named a destination-capable backend, or when the config is
+    incomplete, and the request then keeps the operator's own exporters.
+
+    Two entries naming the same backend merge their ``callback_vars`` last-wins, the
+    way ``convert_key_logging_metadata_to_callback`` merges them, so the destination
+    and the per-request tracer routing cannot read one config two ways.
+
+    A ``failure``-only entry is skipped: a destination is resolved during auth, before
+    the request has an outcome, so honouring the filter would mean holding every span
+    back until the call finishes. Those entries keep today's behaviour instead, where
+    the tenant's credentials reach the backend through per-request tracer routing and
+    the operator's exporter is left alone.
+
+    A backend the request disabled dynamically, through the key's
+    ``litellm_disabled_callbacks`` or the ``x-litellm-disable-callbacks`` header in
+    ``request_headers``, resolves to no destination, so the fan-out never carries the
+    request tree to that account and the operator's exporter is never suppressed for
+    it. That leaves the request exactly where it stood before destinations existed:
+    the OTel V2 logger itself is not on the disable list's class registry, so its own
+    span still routes to the tenant's credentials the way it did then.
+    """
+    from litellm.integrations.otel.model.config import is_otel_v2_enabled
+    from litellm.integrations.otel.presets.destinations import destination_for
+
+    if not is_otel_v2_enabled():
+        return ()
+    key_entries: Final = KeyAndTeamLoggingSettings.get_key_dynamic_logging_settings(user_api_key_dict)
+    entries: Final = (
+        key_entries
+        if key_entries is not None
+        else KeyAndTeamLoggingSettings.get_team_dynamic_logging_settings(user_api_key_dict)
+    )
+    if not entries:
+        return ()
+    disabled: Final = _dynamically_disabled_backends(user_api_key_dict, request_headers)
+    callbacks: Final = tuple(
+        callback
+        for item in entries
+        if (callback := _get_validated_callback_metadata(item=item, source="otel-destination")) is not None
+        if callback.callback_type != "failure"
+        if callback.callback_name.lower() not in disabled
+    )
+    return tuple(
+        destination
+        for name in dict.fromkeys(callback.callback_name for callback in callbacks)
+        if (
+            destination := destination_for(
+                name,
+                _tenant_otel_params(
+                    MappingProxyType(
+                        {
+                            var: value
+                            for callback in callbacks
+                            if callback.callback_name == name
+                            for var, value in callback.callback_vars.items()
+                        }
+                    )
+                ),
+                _tenant_service_name(user_api_key_dict),
+            )
+        )
+        is not None
+    )
+
+
+def _tenant_service_name(user_api_key_dict: UserAPIKeyAuth) -> str | None:
+    """The ``service.name`` this key or team configured, the key winning over its team.
+
+    Same fields and same precedence the request-metadata build applies, read straight
+    off the auth object because destinations resolve during auth, before that metadata
+    is assembled.
+    """
+    sources: Final = (user_api_key_dict.metadata, user_api_key_dict.team_metadata)
+    return next(
+        (
+            stripped
+            for source in sources
+            if source
+            for field in OTEL_SERVICE_NAME_METADATA_KEYS
+            if isinstance(value := source.get(field), str) and (stripped := value.strip())
+        ),
+        None,
+    )
+
+
 def clean_headers(
     headers: Headers,
     litellm_key_header_name: str | None = None,

litellm/proxy/management_endpoints/team_callback_endpoints.py

@@ -20,6 +20,7 @@ from litellm.proxy._types import (
     LiteLLM_AuditLogs,
     LiteLLM_TeamTable,
     LitellmTableNames,
+    LitellmUserRoles,
     ProxyErrorTypes,
     ProxyException,
     TeamCallbackDeleteResponse,
@@ -28,7 +29,10 @@ from litellm.proxy._types import (
     UserAPIKeyAuth,
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
-from litellm.proxy.common_utils.callback_config_validation import callback_config_error
+from litellm.proxy.common_utils.callback_config_validation import (
+    callback_config_error,
+    cross_entry_family_error,
+)
 from litellm.proxy.common_utils.callback_utils import (
     _CALLBACK_VAR_ENCRYPTED_PREFIX,
     decrypt_callback_vars,
@@ -230,6 +234,22 @@ def _callback_error(status_code: int, message: str) -> HTTPException:
     )
 
 
+def _unknown_team_error(team_id: str, user_api_key_dict: UserAPIKeyAuth, status_code: int) -> HTTPException:
+    """Report an unknown team without telling an unauthorized caller that it is unknown.
+
+    These routes are reachable by any authenticated caller so that a team admin can
+    get as far as _verify_team_access. A distinct "does not exist" would therefore let
+    any valid key probe which team ids exist, so a caller who could not have managed
+    the team either way gets the same 403 body _verify_team_access raises.
+    """
+    if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN:
+        return _callback_error(status_code, f"Team id = {team_id} does not exist.")
+    return HTTPException(
+        status_code=status.HTTP_403_FORBIDDEN,
+        detail="You do not have access to this team",
+    )
+
+
 @router.post(
     "/team/{team_id:path}/callback",
     tags=["team management"],
@@ -304,10 +324,7 @@ async def add_team_callbacks(
         # Check if team_id exists already
         _existing_team = await prisma_client.get_data(team_id=team_id, table_name="team", query_type="find_unique")
         if _existing_team is None:
-            raise HTTPException(
-                status_code=400,
-                detail={"error": f"Team id = {team_id} does not exist. Please use a different team id."},
-            )
+            raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_400_BAD_REQUEST)
 
         # IDOR guard: only proxy admins / org admins / team admins of THIS
         # team may write callback credentials. Without this, any
@@ -326,6 +343,28 @@ async def add_team_callbacks(
         if team_callback_settings is None or not isinstance(team_callback_settings, list):
             team_callback_settings = []
 
+        # One entry has to own a credential family end to end. The entries are
+        # flattened into one dict before a request reads them, so an entry
+        # naming only a destination would pair with a key written on another
+        # entry and carry it to that destination -- a key a team admin can read
+        # back nowhere. Repeating a value the owning entry already stores is
+        # fine, which is how one integration covers both events. Proxy admins
+        # are exempt: they already hold every credential the proxy has.
+        if user_api_key_dict.user_role != LitellmUserRoles.PROXY_ADMIN:
+            # Decrypted, because the check compares the incoming values against
+            # the stored ones and the credentials are encrypted at rest.
+            decrypted_logging: Final = decrypt_callback_vars(team_metadata).get("logging")
+            stored_entries: Final = decrypted_logging if isinstance(decrypted_logging, list) else ()
+            stored_entry_vars: Final = [  # mutable-ok: read-only input to the check, never stored
+                entry.get("callback_vars") or {} for entry in stored_entries
+            ]
+            family_error: Final = cross_entry_family_error(data.callback_vars, stored_entry_vars)
+            if family_error is not None:
+                raise HTTPException(
+                    status_code=status.HTTP_400_BAD_REQUEST,
+                    detail=family_error,
+                )
+
         ## check if it already exists, for the same callback event
         for callback in team_callback_settings:
             if (
@@ -452,7 +491,7 @@ async def delete_team_callback(
             team_id=team_id, table_name="team", query_type="find_unique"
         )
         if _existing_team is None:
-            raise _callback_error(404, f"Team id = {team_id} does not exist.")
+            raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_404_NOT_FOUND)
 
         # IDOR guard: only proxy admins / org admins / team admins of THIS team may
         # deregister its callbacks, otherwise any authenticated key holder could
@@ -726,10 +765,7 @@ async def get_team_callbacks(
         # Check if team_id exists
         _existing_team = await prisma_client.get_data(team_id=team_id, table_name="team", query_type="find_unique")
         if _existing_team is None:
-            raise HTTPException(
-                status_code=404,
-                detail={"error": f"Team id = {team_id} does not exist."},
-            )
+            raise _unknown_team_error(team_id, user_api_key_dict, status.HTTP_404_NOT_FOUND)
 
         # IDOR guard: callback metadata holds third-party API credentials
         # (Langfuse / Langsmith / GCS). Only proxy admins / org admins /

litellm/proxy/openai_files_endpoints/files_endpoints.py

@@ -45,6 +45,11 @@ from litellm.proxy.common_utils.openai_endpoint_utils import (
     get_custom_llm_provider_from_request_headers,
     get_custom_llm_provider_from_request_query,
 )
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.proxy.openai_files_endpoints.batch_file_validation import (
     check_batch_file_upload,
     raise_batch_file_validation_failure,
@@ -296,22 +301,22 @@ async def route_create_file(
         if managed_files_obj is None:
             raise ProxyException(
                 message="Managed files hook not found",
-                type="None",
-                param="None",
+                type=ProxyErrorTypes.internal_server_error.value,
+                param=None,
                 code=500,
             )
         if llm_router is None:
             raise ProxyException(
                 message="LLM Router not found",
-                type="None",
-                param="None",
+                type=ProxyErrorTypes.internal_server_error.value,
+                param=None,
                 code=500,
             )
         if not isinstance(managed_files_obj, BaseFileEndpoints):
             raise ProxyException(
                 message="Managed files hook is not a BaseFileEndpoints",
-                type="None",
-                param="None",
+                type=ProxyErrorTypes.internal_server_error.value,
+                param=None,
                 code=500,
             )
         # Managed files internally calls llm_router.acreate_file() which includes loadbalancing
@@ -713,17 +718,17 @@ async def create_file(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e.detail)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )
     finally:
         for spool in spools:
@@ -812,22 +817,22 @@ async def get_file_content(
             if managed_files_obj is None:
                 raise ProxyException(
                     message="Managed files hook not found",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             if llm_router is None:
                 raise ProxyException(
                     message="LLM Router not found",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             if not isinstance(managed_files_obj, BaseFileEndpoints):
                 raise ProxyException(
                     message="Managed files hook is not a BaseFileEndpoints",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
 
@@ -1021,17 +1026,17 @@ async def get_file_content(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e.detail)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )
 
 
@@ -1151,15 +1156,15 @@ async def get_file(
             if managed_files_obj is None:
                 raise ProxyException(
                     message="Managed files hook not found",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             if not isinstance(managed_files_obj, BaseFileEndpoints):
                 raise ProxyException(
                     message="Managed files hook is not a BaseFileEndpoints",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             response = await managed_files_obj.afile_retrieve(
@@ -1215,17 +1220,17 @@ async def get_file(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e.detail)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )
 
 
@@ -1355,22 +1360,22 @@ async def delete_file(
             if managed_files_obj is None:
                 raise ProxyException(
                     message="Managed files hook not found",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             if llm_router is None:
                 raise ProxyException(
                     message="LLM Router not found",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
             if not isinstance(managed_files_obj, BaseFileEndpoints):
                 raise ProxyException(
                     message="Managed files hook is not a BaseFileEndpoints",
-                    type="None",
-                    param="None",
+                    type=ProxyErrorTypes.internal_server_error.value,
+                    param=None,
                     code=500,
                 )
 
@@ -1427,17 +1432,17 @@ async def delete_file(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e.detail)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )
 
 
@@ -1629,15 +1634,15 @@ async def list_files(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e.detail)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -78,6 +78,11 @@ from litellm.proxy.common_utils.http_parsing_utils import (
     _read_request_body,
     _safe_get_request_headers,
 )
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.proxy.common_utils.sse_keepalive import (
     wrap_passthrough_sse_bytes_with_keepalive_pings,
 )
@@ -311,9 +316,9 @@ async def chat_completion_pass_through_endpoint(
         error_msg: Final = f"{e}"
         raise ProxyException(
             message=getattr(e, "message", error_msg),
-            type=getattr(e, "type", "None"),
-            param=getattr(e, "param", "None"),
-            code=getattr(e, "status_code", 500),
+            type=openai_error_type(e, error_status_code(e, 500)),
+            param=openai_error_param(e),
+            code=error_status_code(e, 500),
         )
 
 
@@ -1728,18 +1733,18 @@ async def pass_through_request(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(getattr(e, "detail", str(e)))),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
                 headers=custom_headers,
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
                 headers=custom_headers,
             )
 

litellm/proxy/realtime_endpoints/endpoints.py

@@ -17,6 +17,11 @@ from litellm.proxy.common_utils.encrypt_decrypt_utils import (
     encrypt_value_helper,
 )
 from litellm.proxy.common_utils.http_parsing_utils import _read_request_body
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 from litellm.types.realtime import (
     RealtimeClientSecretRequest,
     RealtimeClientSecretResponse,
@@ -304,15 +309,15 @@ async def create_realtime_client_secret(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", http_status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, http_status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, http_status.HTTP_400_BAD_REQUEST),
             )
         raise ProxyException(
             message=getattr(e, "message", str(e)),
-            type=getattr(e, "type", "None"),
-            param=getattr(e, "param", "None"),
-            code=getattr(e, "status_code", 500),
+            type=openai_error_type(e, error_status_code(e, 500)),
+            param=openai_error_param(e),
+            code=error_status_code(e, 500),
         )
 
     if upstream_resp.status_code != 200:
@@ -495,15 +500,15 @@ async def proxy_realtime_calls(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", http_status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, http_status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, http_status.HTTP_400_BAD_REQUEST),
             )
         raise ProxyException(
             message=getattr(e, "message", str(e)),
-            type=getattr(e, "type", "None"),
-            param=getattr(e, "param", "None"),
-            code=getattr(e, "status_code", 500),
+            type=openai_error_type(e, error_status_code(e, 500)),
+            param=openai_error_param(e),
+            code=error_status_code(e, 500),
         )
 
     return Response(
@@ -608,15 +613,15 @@ async def create_realtime_transcription_session(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "detail", getattr(e, "message", str(e))),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", http_status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, http_status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, http_status.HTTP_400_BAD_REQUEST),
             )
         raise ProxyException(
             message=getattr(e, "message", str(e)),
-            type=getattr(e, "type", "None"),
-            param=getattr(e, "param", "None"),
-            code=getattr(e, "status_code", 500),
+            type=openai_error_type(e, error_status_code(e, 500)),
+            param=openai_error_param(e),
+            code=error_status_code(e, 500),
         )
 
     if upstream_resp.status_code != 200:

litellm/proxy/rerank_endpoints/endpoints.py

@@ -11,6 +11,11 @@ from litellm._logging import verbose_proxy_logger
 from litellm.proxy._types import *
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
 
 router: Final = APIRouter()
 
@@ -112,15 +117,15 @@ async def rerank(
         if isinstance(e, HTTPException):
             raise ProxyException(
                 message=getattr(e, "message", str(e)),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", status.HTTP_400_BAD_REQUEST),
+                type=openai_error_type(e, error_status_code(e, status.HTTP_400_BAD_REQUEST)),
+                param=openai_error_param(e),
+                code=error_status_code(e, status.HTTP_400_BAD_REQUEST),
             )
         else:
             error_msg: Final = f"{e}"
             raise ProxyException(
                 message=getattr(e, "message", error_msg),
-                type=getattr(e, "type", "None"),
-                param=getattr(e, "param", "None"),
-                code=getattr(e, "status_code", 500),
+                type=openai_error_type(e, error_status_code(e, 500)),
+                param=openai_error_param(e),
+                code=error_status_code(e, 500),
             )

litellm/proxy/utils.py

@@ -37,6 +37,7 @@ from litellm.proxy._types import (
     SpendLogsMetadata,
     SpendLogsPayload,
 )
+from litellm.proxy.common_utils.openai_error_payload import openai_error_param
 from litellm.proxy.spend_tracking.spend_log_error_logger import spend_log_error
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.proxy.model_listing import ModelInfoResponse
@@ -3806,7 +3807,7 @@ class _StaleReadEngine:
 class PrismaClient:
     spend_log_transactions: list = []
     _spend_log_transactions_lock = asyncio.Lock()
-    spend_log_flush_requested: ClassVar[asyncio.Event] = asyncio.Event()
+    spend_log_flush_requested: "asyncio.Event | None" = None
     spend_log_queue_bytes: ClassVar[int] = 0
     spend_logs_queue_monitor_task: "asyncio.Task[None] | None" = None
     tool_usage_transactions: list["ToolUsageTransaction"] = []
@@ -6532,23 +6533,27 @@ async def enqueue_spend_logs(
         )
 
 
-def request_spend_log_flush() -> None:
-    """Wake the queue monitor now rather than leaving the rows for its next poll.
+def request_spend_log_flush(prisma_client: PrismaClient) -> None:
+    """Wake this client's queue monitor now rather than leaving the rows for its next poll.
 
     The Responses API hands the client an id it can chain from straight away, and that
     lookup reads the DB, so the row cannot sit in this worker's queue for a poll interval.
     Repeated requests coalesce into the monitor's next pass, so the batching holds.
+    A request made before the monitor is running is dropped, and loses nothing: the
+    monitor reads the queue on its first pass, before it ever waits on a request.
     """
-    PrismaClient.spend_log_flush_requested.set()
+    flush_requested: Final = prisma_client.spend_log_flush_requested
+    if flush_requested is not None:
+        flush_requested.set()
 
 
-async def _wait_for_spend_log_flush_request(interval: float) -> bool:
+async def _wait_for_spend_log_flush_request(flush_requested: asyncio.Event, interval: float) -> bool:
     """Wait out ``interval``, returning early and True when a flush was requested."""
     try:
-        await asyncio.wait_for(PrismaClient.spend_log_flush_requested.wait(), timeout=interval)
+        await asyncio.wait_for(flush_requested.wait(), timeout=interval)
     except asyncio.TimeoutError:
         return False
-    PrismaClient.spend_log_flush_requested.clear()
+    flush_requested.clear()
     return True
 
 
@@ -6975,6 +6980,8 @@ async def _monitor_spend_logs_queue(
     max_backoff: Final = 30.0  # Maximum backoff interval in seconds
     backoff_multiplier: Final = 1.5  # Exponential backoff multiplier
     current_interval = base_interval
+    flush_requested: Final = asyncio.Event()
+    prisma_client.spend_log_flush_requested = flush_requested  # rebind-ok: the client owns its monitor's flush signal
 
     verbose_proxy_logger.info(
         "Starting spend logs queue monitor (threshold: %s, poll_interval: %ss)", threshold, base_interval
@@ -7013,7 +7020,7 @@ async def _monitor_spend_logs_queue(
                 # Exponential backoff when no logs to process
                 current_interval = min(current_interval * backoff_multiplier, max_backoff)
 
-            if await _wait_for_spend_log_flush_request(current_interval):
+            if await _wait_for_spend_log_flush_request(flush_requested, current_interval):
                 current_interval = base_interval
         except Exception as e:
             spend_log_error("Error in spend logs queue monitor: %s", str(e), exc=e)
@@ -7398,7 +7405,7 @@ def handle_exception_on_proxy(e: Exception) -> ProxyException:
         return ProxyException(
             message=getattr(e, "detail", f"error({e})"),
             type=ProxyErrorTypes.internal_server_error,
-            param=getattr(e, "param", "None"),
+            param=openai_error_param(e),
             code=getattr(e, "status_code", status.HTTP_500_INTERNAL_SERVER_ERROR),
         )
     elif isinstance(e, ProxyException):
@@ -7407,7 +7414,7 @@ def handle_exception_on_proxy(e: Exception) -> ProxyException:
     return ProxyException(
         message=str(e),
         type=ProxyErrorTypes.internal_server_error,
-        param=getattr(e, "param", "None"),
+        param=openai_error_param(e),
         code=_status_code,
     )
 

litellm/router.py

@@ -3656,6 +3656,13 @@ class Router:
         effective_model_info: Final = kwargs.get("model_info") or deployment.get("model_info") or MappingProxyType({})
         self._set_failed_deployment_id_on_exception(exception, MappingProxyType({"model_info": effective_model_info}))
 
+    @staticmethod
+    def _stamp_retry_skip_deployment_id(exception: Exception, kwargs: Mapping[str, object]) -> None:
+        effective_model_info: Final = kwargs.get("model_info")
+        deployment_id: Final = effective_model_info.get("id") if isinstance(effective_model_info, Mapping) else None
+        if isinstance(deployment_id, str) and deployment_id:
+            exception.retry_skip_deployment_id = deployment_id  # pyright: ignore[reportAttributeAccessIssue]  # dynamic stamp, read by _deployment_ids_to_skip_on_retry
+
     def _update_kwargs_with_default_litellm_params(
         self, kwargs: dict, metadata_variable_name: str | None = "metadata"
     ) -> None:
@@ -4358,6 +4365,7 @@ class Router:
                 model=model,
                 messages=[{"role": "user", "content": "prompt"}],
                 specific_deployment=kwargs.pop("specific_deployment", None),
+                request_kwargs=kwargs,
             )
             self._update_kwargs_with_deployment(deployment=deployment, kwargs=kwargs)
             data: Final = deployment["litellm_params"].copy()
@@ -4388,6 +4396,7 @@ class Router:
             verbose_router_logger.info("litellm.image_generation(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def aimage_generation(self, prompt: str, model: str, **kwargs):
@@ -4472,6 +4481,7 @@ class Router:
             verbose_router_logger.info("litellm.aimage_generation(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def atranscription(self, file: FileTypes, model: str, **kwargs):
@@ -4576,6 +4586,7 @@ class Router:
             verbose_router_logger.info("litellm.atranscription(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def aspeech(self, model: str, input: str, voice: str | None = None, **kwargs):
@@ -4690,6 +4701,7 @@ class Router:
             verbose_router_logger.info("litellm.aspeech(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def arerank(self, model: str, **kwargs):
@@ -4748,6 +4760,7 @@ class Router:
             verbose_router_logger.info("litellm.arerank(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     def text_completion(
@@ -4882,6 +4895,7 @@ class Router:
             verbose_router_logger.info("litellm.atext_completion(model=%s)\x1b[31m Exception %s\x1b[0m", model, e)
             if model is not None:
                 self.fail_calls[model] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def aadapter_completion(
@@ -4972,6 +4986,7 @@ class Router:
             verbose_router_logger.info("litellm.aadapter_completion(model=%s)\x1b[31m Exception %s\x1b[0m", model, e)
             if model is not None:
                 self.fail_calls[model] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def _asearch_with_fallbacks(self, original_function: Callable, **kwargs):
@@ -5754,6 +5769,7 @@ class Router:
                 model=model,
                 input=input,
                 specific_deployment=kwargs.pop("specific_deployment", None),
+                request_kwargs=kwargs,
             )
             self._update_kwargs_with_deployment(deployment=deployment, kwargs=kwargs)
             data: Final = deployment["litellm_params"].copy()
@@ -5792,6 +5808,7 @@ class Router:
             verbose_router_logger.info("litellm.embedding(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def aembedding(
@@ -5879,6 +5896,7 @@ class Router:
             verbose_router_logger.info("litellm.aembedding(model=%s)\x1b[31m Exception %s\x1b[0m", model_name, e)
             if model_name is not None:
                 self.fail_calls[model_name] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     #### FILES API ####
@@ -6252,6 +6270,7 @@ class Router:
             )
             if model is not None:
                 self.fail_calls[model] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def aretrieve_batch(
@@ -6472,6 +6491,7 @@ class Router:
             )
             if model is not None:
                 self.fail_calls[model] += 1
+            self._stamp_retry_skip_deployment_id(e, kwargs)
             raise e
 
     async def alist_batches(
@@ -7589,7 +7609,9 @@ class Router:
 
     @staticmethod
     def _deployment_ids_to_skip_on_retry(exception: Exception, already_skipped: object) -> tuple[str, ...]:
-        failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
+        failed_deployment_id: Final[str | None] = getattr(exception, "retry_skip_deployment_id", None) or getattr(
+            exception, "failed_deployment_id", None
+        )
         status_code: Final = getattr(exception, "status_code", None)
         if not failed_deployment_id or not isinstance(status_code, int):
             return ()

litellm/types/llms/bedrock.py

@@ -1,7 +1,7 @@
 import json
 from collections.abc import Sequence
 from enum import Enum
-from typing import TYPE_CHECKING, Any, Final, Literal
+from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias
 
 from typing_extensions import ReadOnly, Required, TypedDict, override
 
@@ -557,7 +557,7 @@ class AmazonTitanMultimodalEmbeddingResponse(TypedDict):
     message: str  # Specifies any errors that occur during generation.
 
 
-# TwelveLabs Marengo Embed 2.7 types
+# TwelveLabs Marengo Embed types
 TWELVELABS_EMBEDDING_INPUT_TYPES = Literal["text", "image", "video", "audio"]
 TWELVELABS_EMBEDDING_OPTIONS = Literal["visual-text", "visual-image", "audio"]
 
@@ -591,6 +591,113 @@ class TwelveLabsMarengoEmbeddingResponse(TypedDict):
     endSec: float
 
 
+TWELVELABS_MARENGO_3_INPUT_TYPES: TypeAlias = Literal["text", "image", "video", "audio", "text_image", "multi_input"]
+TWELVELABS_MARENGO_3_EMBEDDING_OPTIONS: TypeAlias = Literal["visual", "audio", "transcription"]
+TWELVELABS_MARENGO_3_EMBEDDING_TYPES: TypeAlias = Literal["separate_embedding", "fused_embedding"]
+TWELVELABS_MARENGO_3_EMBEDDING_SCOPES: TypeAlias = Literal["clip", "asset"]
+
+
+class TwelveLabsMarengo3FixedSegmentationConfig(TypedDict):
+    durationSec: ReadOnly[int]
+
+
+class TwelveLabsMarengo3FixedSegmentation(TypedDict):
+    method: ReadOnly[Literal["fixed"]]
+    fixed: ReadOnly[TwelveLabsMarengo3FixedSegmentationConfig]
+
+
+class TwelveLabsMarengo3DynamicSegmentationConfig(TypedDict):
+    minDurationSec: ReadOnly[int]
+
+
+class TwelveLabsMarengo3DynamicSegmentation(TypedDict):
+    method: ReadOnly[Literal["dynamic"]]
+    dynamic: ReadOnly[TwelveLabsMarengo3DynamicSegmentationConfig]
+
+
+TwelveLabsMarengo3Segmentation: TypeAlias = TwelveLabsMarengo3FixedSegmentation | TwelveLabsMarengo3DynamicSegmentation
+
+
+class TwelveLabsMarengo3TextInput(TypedDict):
+    inputText: ReadOnly[str]
+
+
+class TwelveLabsMarengo3ImageInput(TypedDict):
+    mediaSource: ReadOnly[TwelveLabsMediaSource]
+
+
+class TwelveLabsMarengo3TimedMediaOptions(TypedDict, total=False):
+    startSec: ReadOnly[float]
+    endSec: ReadOnly[float]
+    segmentation: ReadOnly[TwelveLabsMarengo3Segmentation]
+    embeddingOption: ReadOnly[Sequence[TWELVELABS_MARENGO_3_EMBEDDING_OPTIONS]]
+    embeddingType: ReadOnly[Sequence[TWELVELABS_MARENGO_3_EMBEDDING_TYPES]]
+    embeddingScope: ReadOnly[Sequence[TWELVELABS_MARENGO_3_EMBEDDING_SCOPES]]
+
+
+class TwelveLabsMarengo3TimedMediaInput(TwelveLabsMarengo3TimedMediaOptions):
+    mediaSource: Required[ReadOnly[TwelveLabsMediaSource]]
+
+
+class TwelveLabsMarengo3TextImageInput(TypedDict):
+    inputText: ReadOnly[str]
+    mediaSource: ReadOnly[TwelveLabsMediaSource]
+
+
+class TwelveLabsMarengo3NamedMediaSource(TwelveLabsMediaSource):
+    name: Required[ReadOnly[str]]
+    mediaType: Required[ReadOnly[Literal["image"]]]
+
+
+class TwelveLabsMarengo3MultiInput(TypedDict, total=False):
+    inputText: ReadOnly[str]
+    mediaSources: Required[ReadOnly[Sequence[TwelveLabsMarengo3NamedMediaSource]]]
+
+
+class TwelveLabsMarengo3RequestBase(TypedDict, total=False):
+    inferenceId: ReadOnly[str]
+
+
+class TwelveLabsMarengo3TextRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["text"]]
+    text: ReadOnly[TwelveLabsMarengo3TextInput]
+
+
+class TwelveLabsMarengo3ImageRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["image"]]
+    image: ReadOnly[TwelveLabsMarengo3ImageInput]
+
+
+class TwelveLabsMarengo3VideoRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["video"]]
+    video: ReadOnly[TwelveLabsMarengo3TimedMediaInput]
+
+
+class TwelveLabsMarengo3AudioRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["audio"]]
+    audio: ReadOnly[TwelveLabsMarengo3TimedMediaInput]
+
+
+class TwelveLabsMarengo3TextImageRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["text_image"]]
+    text_image: ReadOnly[TwelveLabsMarengo3TextImageInput]
+
+
+class TwelveLabsMarengo3MultiInputRequest(TwelveLabsMarengo3RequestBase):
+    inputType: ReadOnly[Literal["multi_input"]]
+    multi_input: ReadOnly[TwelveLabsMarengo3MultiInput]
+
+
+TwelveLabsMarengo3EmbeddingRequest: TypeAlias = (
+    TwelveLabsMarengo3TextRequest
+    | TwelveLabsMarengo3ImageRequest
+    | TwelveLabsMarengo3VideoRequest
+    | TwelveLabsMarengo3AudioRequest
+    | TwelveLabsMarengo3TextImageRequest
+    | TwelveLabsMarengo3MultiInputRequest
+)
+
+
 class TwelveLabsS3OutputDataConfig(TypedDict):
     s3Uri: str
 
@@ -601,7 +708,7 @@ class TwelveLabsOutputDataConfig(TypedDict):
 
 class TwelveLabsAsyncInvokeRequest(TypedDict):
     modelId: str
-    modelInput: TwelveLabsMarengoEmbeddingRequest
+    modelInput: ReadOnly[TwelveLabsMarengoEmbeddingRequest | TwelveLabsMarengo3EmbeddingRequest]
     outputDataConfig: TwelveLabsOutputDataConfig
 
 

litellm/types/utils.py

@@ -272,7 +272,7 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False):
     input_cost_per_token_above_272k_tokens_flex: float | None
     input_cost_per_token_above_512k_tokens: float | None  # MiniMax-M3: prompts >512K priced at 2x input
     input_cost_per_character_above_128k_tokens: float | None  # only for vertex ai models
-    input_cost_per_query: float | None  # only for rerank models
+    input_cost_per_query: float | None  # per-request pricing: rerank, search, and Bedrock Marengo embeddings
     input_cost_per_image: float | None  # only for vertex ai models
     input_cost_per_image_token: float | None  # for gpt-image-1 and similar models
     input_cost_per_video_token: float | None  # for gemini omni models with video input
@@ -1694,6 +1694,9 @@ class PromptTokensDetailsWrapper(
     audio_length_seconds: float | None = None
     """Length of audio sent to the model. Used for multimodal embeddings priced per audio-second."""
 
+    query_count: int | None = None
+    """Number of billable requests sent to the model. Used for embeddings priced per request, such as Bedrock Marengo."""
+
     cache_write_tokens: int | None = None
     """Number of cache write (creation) tokens sent to the model. OpenAI naming (prompt_tokens_details.cache_write_tokens); this is the canonical field."""
 
@@ -1735,6 +1738,8 @@ class PromptTokensDetailsWrapper(
             del self.video_length_seconds
         if self.audio_length_seconds is None:
             del self.audio_length_seconds
+        if self.query_count is None:
+            del self.query_count
         if self.web_search_requests is None:
             del self.web_search_requests
         if self.google_maps_grounding_requests is None:

litellm/utils.py

@@ -3632,7 +3632,7 @@ def get_optional_params_embeddings(
         elif "cohere.embed" in model:
             object = litellm.BedrockCohereEmbeddingConfig()
         elif "twelvelabs" in model or "marengo" in model:
-            object = litellm.TwelveLabsMarengoEmbeddingConfig()
+            object = litellm.TwelveLabsMarengoEmbeddingConfig(model=model)
         elif "nova" in model.lower():
             object = litellm.AmazonNovaEmbeddingConfig()
         else:  # unmapped model
@@ -6043,7 +6043,7 @@ def get_model_info(
             input_cost_per_character_above_128k_tokens: Optional[
                 float
             ]  # only for vertex ai models
-            input_cost_per_query: Optional[float] # only for rerank models
+            input_cost_per_query: Optional[float] # per-request pricing: rerank, search, and Bedrock Marengo embeddings
             input_cost_per_image: Optional[float]  # only for vertex ai models
             input_cost_per_audio_token: Optional[float]
             input_cost_per_audio_per_second: Optional[float]  # only for vertex ai models

model_prices_and_context_window.json

@@ -650,7 +650,10 @@
     },
     "twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
         "litellm_provider": "bedrock",
         "max_input_tokens": 77,
         "max_tokens": 77,
@@ -662,7 +665,7 @@
     },
     "us.twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
         "input_cost_per_video_per_second": 0.0007,
         "input_cost_per_audio_per_second": 0.00014,
         "input_cost_per_image": 0.0001,
@@ -677,7 +680,7 @@
     },
     "eu.twelvelabs.marengo-embed-2-7-v1:0": {
         "deprecation_date": "2026-11-30",
-        "input_cost_per_token": 7e-05,
+        "input_cost_per_query": 7e-05,
         "input_cost_per_video_per_second": 0.0007,
         "input_cost_per_audio_per_second": 0.00014,
         "input_cost_per_image": 0.0001,
@@ -690,6 +693,48 @@
         "supports_embedding_image_input": true,
         "supports_image_input": true
     },
+    "twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
+    "us.twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
+    "eu.twelvelabs.marengo-embed-3-0-v1:0": {
+        "input_cost_per_query": 7e-05,
+        "input_cost_per_video_per_second": 0.0007,
+        "input_cost_per_audio_per_second": 0.00014,
+        "input_cost_per_image": 0.0001,
+        "litellm_provider": "bedrock",
+        "max_input_tokens": 500,
+        "max_tokens": 500,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 512,
+        "supports_embedding_image_input": true,
+        "supports_image_input": true
+    },
     "twelvelabs.pegasus-1-2-v1:0": {
         "input_cost_per_video_per_second": 0.00049,
         "output_cost_per_token": 7.5e-06,

pyproject.toml

@@ -114,7 +114,6 @@ caching = ["diskcache>=5.6.3,<6.0"]
 mcp = ["mcp>=1.28.1,<2.0"]
 # Driver for the MongoDB Atlas vector store; Atlas Vector Search has no HTTP query API.
 # The floor is 4.9 because that is the release AsyncMongoClient landed in.
-mongodb = ["pymongo>=4.9,<5.0"]
 # SAML SSO for the admin UI. python3-saml pulls in xmlsec/lxml, whose wheels
 # bundle the native libxmlsec1/libxml2 libraries, so no system packages are
 # required. Kept out of the base `proxy` extra so it stays optional.

tests/test_litellm/integrations/otel/test_otel_v2_components.py

@@ -6,14 +6,18 @@ import json
 import threading
 from collections.abc import Iterator
 from dataclasses import replace
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from http.server import BaseHTTPRequestHandler, HTTPServer, ThreadingHTTPServer
 
 import pytest
 
 pytest.importorskip("opentelemetry")
 
+from opentelemetry.proto.collector.trace.v1.trace_service_pb2 import (  # noqa: E402
+    ExportTraceServiceRequest,
+)
 from opentelemetry.sdk.metrics import MeterProvider  # noqa: E402
 from opentelemetry.sdk.metrics.export import InMemoryMetricReader  # noqa: E402
+from opentelemetry.sdk.trace import TracerProvider  # noqa: E402
 from opentelemetry.sdk.trace.export import (  # noqa: E402
     BatchSpanProcessor,
     ConsoleSpanExporter,
@@ -541,6 +545,89 @@ def test_build_span_exporter_variants():
     assert "OTLPSpanExporter" in type(http_exporter).__name__
 
 
+def _export_one_trace_to_local_collector(exporter_kind: str) -> tuple[list[dict], tuple[int, int, int]]:
+    """Run a parent/child trace through the configured exporter against a
+    throwaway HTTP collector. Returns the requests as the collector saw them
+    (child first, since it ends first) and (trace_id, parent span_id, child span_id)."""
+    received: list[dict] = []
+
+    class Collector(BaseHTTPRequestHandler):
+        def do_POST(self):
+            body = self.rfile.read(int(self.headers["Content-Length"]))
+            received.append({"path": self.path, "headers": dict(self.headers), "body": body})
+            self.send_response(200)
+            self.end_headers()
+
+        def log_message(self, *_args):
+            pass
+
+    server = HTTPServer(("127.0.0.1", 0), Collector)
+    threading.Thread(target=server.serve_forever, daemon=True).start()
+    try:
+        config = OpenTelemetryV2Config(
+            exporter=exporter_kind,
+            endpoint=f"http://127.0.0.1:{server.server_port}",
+            headers="x-collector-token=secret",
+        )
+        provider = TracerProvider()
+        provider.add_span_processor(SimpleSpanProcessor(providers.build_span_exporter(config)))
+        tracer = provider.get_tracer("test")
+        with tracer.start_as_current_span("parent", kind=SpanKind.SERVER) as parent:
+            with tracer.start_as_current_span("child") as child:
+                ids = (
+                    parent.get_span_context().trace_id,
+                    parent.get_span_context().span_id,
+                    child.get_span_context().span_id,
+                )
+        provider.shutdown()
+    finally:
+        server.shutdown()
+        server.server_close()
+    assert len(received) == 2
+    return received, ids
+
+
+def _only_span(request: dict) -> dict:
+    scope_spans = json.loads(request["body"])["resourceSpans"][0]["scopeSpans"][0]["spans"]
+    assert len(scope_spans) == 1
+    return scope_spans[0]
+
+
+def test_http_json_exporter_posts_otlp_json_to_traces_endpoint():
+    """``http/json`` must put the OTLP/JSON mapping on the wire (camelCase
+    fields, integer enums, hex ids) with a JSON content type, so collectors that
+    cannot decode protobuf can ingest the trace. Headers still travel."""
+    (child_request, parent_request), (trace_id, parent_id, child_id) = _export_one_trace_to_local_collector("http/json")
+
+    assert parent_request["path"] == "/v1/traces"
+    assert parent_request["headers"]["Content-Type"] == "application/json"
+    assert parent_request["headers"]["x-collector-token"] == "secret"
+    parent = _only_span(parent_request)
+    assert parent["name"] == "parent"
+    assert parent["kind"] == 2
+    assert parent["traceId"] == format(trace_id, "032x")
+    assert parent["spanId"] == format(parent_id, "016x")
+    assert "parentSpanId" not in parent
+    child = _only_span(child_request)
+    assert child["traceId"] == format(trace_id, "032x")
+    assert child["spanId"] == format(child_id, "016x")
+    assert child["parentSpanId"] == format(parent_id, "016x")
+
+
+def test_http_protobuf_exporter_still_posts_protobuf():
+    (_child_request, parent_request), (trace_id, _parent_id, _child_id) = _export_one_trace_to_local_collector(
+        "http/protobuf"
+    )
+
+    assert parent_request["path"] == "/v1/traces"
+    assert parent_request["headers"]["Content-Type"] == "application/x-protobuf"
+    assert format(trace_id, "032x").encode() not in parent_request["body"]
+    decoded = ExportTraceServiceRequest.FromString(parent_request["body"])
+    span = decoded.resource_spans[0].scope_spans[0].spans[0]
+    assert span.name == "parent"
+    assert span.trace_id == trace_id.to_bytes(16, "big")
+
+
 @pytest.fixture
 def otlp_collector() -> Iterator[tuple[str, list[str]]]:
     received_paths: list[str] = []
@@ -612,6 +699,21 @@ def test_traces_endpoint_per_exporter_coexists_with_default_normalization(otlp_c
     assert sorted(received_paths) == ["/services/collector/traces", "/v1/traces"]
 
 
+def test_http_json_exporter_honors_traces_endpoint(otlp_collector):
+    base_url, received_paths = otlp_collector
+    cfg = OpenTelemetryV2Config(
+        exporters=[
+            {
+                "kind": "http/json",
+                "endpoint": base_url,
+                "traces_endpoint": f"{base_url}/services/collector/traces",
+            }
+        ]
+    )
+    _export_one_span(cfg)
+    assert received_paths == ["/services/collector/traces"]
+
+
 def test_otlp_metric_exporter_uses_cumulative_histogram_temporality():
     """Histograms must export as cumulative, not delta.
 

tests/test_litellm/integrations/otel/test_otel_v2_logger.py

@@ -2041,7 +2041,7 @@ def test_select_global_otel_v2_logger_builds_one_when_none_registered():
     assert isinstance(chosen, OpenTelemetryV2)
 
 
-def test_publish_global_otel_v2_provider_sets_selected_logger_provider():
+def test_publish_global_otel_v2_provider_sets_selected_logger_provider(monkeypatch):
     """The startup publish must set the OTel global provider to the *selected*
     logger's provider (the preset logger that owns every exporter), so the FastAPI
     server span and the gen-ai spans share one provider and one trace.
@@ -2051,8 +2051,10 @@ def test_publish_global_otel_v2_provider_sets_selected_logger_provider():
     test would otherwise miss: that the published provider is the selected logger's,
     not some other.
     """
+    from litellm.integrations.otel import logger as otel_logger
     from litellm.integrations.otel.logger import publish_global_otel_v2_provider
 
+    monkeypatch.setattr(otel_logger, "_published_v2_provider", None)
     cfg = OpenTelemetryV2Config(exporter="in_memory")
     tp = providers.build_tracer_provider(cfg)
     preset_logger = OpenTelemetryV2(

tests/test_litellm/integrations/test_azure_sentinel.py

@@ -867,6 +867,45 @@ async def test_azure_sentinel_concurrent_threshold_sends_collapse_into_one_attem
     assert getattr(logger, queue_attr) == []
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("queue_attr, send_method, build_payloads", QUEUE_CASES)
+async def test_azure_sentinel_batch_size_bounds_every_request_under_concurrent_events(
+    queue_attr, send_method, build_payloads
+):
+    """Lowering batch_size is the documented way to stay under the ingestion cap, so no request may
+    carry more than batch_size records even when events keep landing while a send is on the wire,
+    and every one of those records still has to arrive exactly once."""
+    logger = _build_logger(batch_size=5)
+    records = build_payloads(40)
+
+    attempts = []
+    first_send_started = asyncio.Event()
+    release_first_send = asyncio.Event()
+
+    async def _on_ingest(data):
+        attempts.append([record["id"] for record in json.loads(data.decode("utf-8"))])
+        if len(attempts) == 1:
+            first_send_started.set()
+            await release_first_send.wait()
+        return _accepted()
+
+    _install_ingestion(logger, _on_ingest)
+
+    sends = [asyncio.create_task(_log(logger, queue_attr, record)) for record in records]
+    await asyncio.wait_for(first_send_started.wait(), timeout=10)
+
+    assert attempts == [[record["id"] for record in records[:5]]]
+    assert getattr(logger, queue_attr) == records[5:]
+
+    release_first_send.set()
+    await asyncio.wait_for(asyncio.gather(*sends), timeout=10)
+    await logger.flush_queue()
+
+    assert max(len(attempt) for attempt in attempts) <= 5
+    assert [record_id for attempt in attempts for record_id in attempt] == [record["id"] for record in records]
+    assert getattr(logger, queue_attr) == []
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("queue_attr, send_method, build_payloads", QUEUE_CASES)
 async def test_azure_sentinel_requeues_a_cancelled_send(

tests/test_litellm/integrations/test_mlflow.py

@@ -195,3 +195,70 @@ def test_mlflow_stream_handler_uses_async_complete_response():
             is final_response
         )
         assert "abc123" not in mlflow_logger._stream_id_to_span
+
+
+def test_mlflow_stream_handler_pops_span_when_end_raises():
+    modules = _mock_mlflow_modules()
+    with patch.dict("sys.modules", modules):
+        from litellm.integrations.mlflow import MlflowLogger
+
+        mlflow_logger = MlflowLogger()
+        mlflow_logger._start_span_or_trace = MagicMock(return_value="mock_span")
+        mlflow_logger._end_span_or_trace = MagicMock(
+            side_effect=TypeError("unexpected keyword argument 'trace_id'")
+        )
+        mlflow_logger._extract_and_set_chat_attributes = MagicMock()
+
+        response_obj = MagicMock()
+        response_obj.choices = []
+
+        kwargs = {
+            "litellm_call_id": "leak123",
+            "complete_streaming_response": MagicMock(),
+        }
+
+        with pytest.raises(TypeError):
+            mlflow_logger._handle_stream_event(
+                kwargs=kwargs,
+                response_obj=response_obj,
+                start_time=datetime.utcnow(),
+                end_time=datetime.utcnow(),
+            )
+
+        assert "leak123" not in mlflow_logger._stream_id_to_span
+
+
+class _Mlflow2StyleClient:
+    """Mimics the mlflow 2.x client signatures, which have no trace_id kwarg."""
+
+    def __init__(self):
+        self.ended_traces = []
+        self.ended_spans = []
+
+    def end_trace(self, request_id, outputs=None, attributes=None, status="OK", end_time_ns=None):
+        self.ended_traces.append(request_id)
+
+    def end_span(self, request_id, span_id, outputs=None, attributes=None, status="OK", end_time_ns=None):
+        self.ended_spans.append((request_id, span_id))
+
+
+def test_mlflow_end_span_or_trace_works_with_mlflow_2x_client():
+    modules = _mock_mlflow_modules()
+    with patch.dict("sys.modules", modules):
+        from litellm.integrations.mlflow import MlflowLogger
+
+        mlflow_logger = MlflowLogger()
+        client = _Mlflow2StyleClient()
+        mlflow_logger._client = client
+
+        root_span = MagicMock(parent_id=None, request_id="req-1")
+        mlflow_logger._end_span_or_trace(
+            span=root_span, outputs="out", end_time_ns=1, status="OK"
+        )
+        assert client.ended_traces == ["req-1"]
+
+        child_span = MagicMock(parent_id="parent-1", request_id="req-2", span_id="span-2")
+        mlflow_logger._end_span_or_trace(
+            span=child_span, outputs="out", end_time_ns=1, status="OK"
+        )
+        assert client.ended_spans == [("req-2", "span-2")]

tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py

@@ -2658,6 +2658,7 @@ def test_cache_writing_cost_with_zero_creation_tokens_and_ephemeral_details():
         "image_count": 0,
         "video_length_seconds": 0.0,
         "audio_length_seconds": 0.0,
+        "query_count": 0,
     }
 
     model_info: ModelInfo = {}
@@ -3239,6 +3240,37 @@ def test_image_count_prevents_text_tokens_fallback(_local_model_cost_map):
     assert completion_cost == 0.0
 
 
+def test_query_count_bills_input_cost_per_query(_local_model_cost_map):
+    usage = Usage(
+        prompt_tokens=0,
+        completion_tokens=0,
+        total_tokens=0,
+        prompt_tokens_details=PromptTokensDetailsWrapper(query_count=3, image_count=1),
+    )
+
+    prompt_cost, completion_cost = generic_cost_per_token(
+        model="us.twelvelabs.marengo-embed-3-0-v1:0",
+        usage=usage,
+        custom_llm_provider="bedrock",
+    )
+
+    assert prompt_cost == pytest.approx(3 * 7e-05 + 1e-04)
+    assert completion_cost == 0.0
+
+
+def test_query_count_is_free_without_a_per_query_price(_local_model_cost_map):
+    usage = Usage(
+        prompt_tokens=0,
+        completion_tokens=0,
+        total_tokens=0,
+        prompt_tokens_details=PromptTokensDetailsWrapper(query_count=1),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(model="text-embedding-3-small", usage=usage, custom_llm_provider="openai")
+
+    assert prompt_cost == 0.0
+
+
 # ---------------------------------------------------------------------------
 # Data-residency (OpenAI regional processing) tests
 # ---------------------------------------------------------------------------

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -6074,15 +6074,17 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa
         for hook in [cb for cb in litellm.callbacks if isinstance(cb, VectorStorePreCallHook)]:
             litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, hook)
 def test_newrelic_dispatch_prefers_otel_v2_when_flag_on(monkeypatch):
-    """With LITELLM_OTEL_V2 on, the "newrelic" callback builds the OTel v2
-    logger (per-team credential routing); with the flag off (default) it keeps
-    the legacy agent-based logger, so existing deployments are untouched."""
+    """With LITELLM_OTEL_V2 on and operator credentials present, the "newrelic"
+    callback builds the OTel v2 logger (per-team credential routing); with the
+    flag off (default) it keeps the legacy agent-based logger, so existing
+    deployments are untouched."""
     from litellm.integrations.otel.logger import OpenTelemetryV2
     from litellm.integrations.otel.model.config import is_otel_v2_enabled
     from litellm.litellm_core_utils import litellm_logging as logging_module
 
     logging_module._in_memory_loggers.clear()
     monkeypatch.setenv("LITELLM_OTEL_V2", "true")
+    monkeypatch.setenv("NEW_RELIC_LICENSE_KEY", "test-license-key")
     is_otel_v2_enabled.cache_clear()
     try:
         v2_logger = logging_module._init_custom_logger_compatible_class(
@@ -6137,6 +6139,7 @@ def test_get_custom_logger_compatible_class_finds_v2_newrelic(monkeypatch):
 
     logging_module._in_memory_loggers.clear()
     monkeypatch.setenv("LITELLM_OTEL_V2", "true")
+    monkeypatch.setenv("NEW_RELIC_LICENSE_KEY", "test-license-key")
     is_otel_v2_enabled.cache_clear()
     try:
         created = logging_module._init_custom_logger_compatible_class(

tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py

@@ -9,7 +9,8 @@ import pytest
 import litellm
 from litellm.constants import RESPONSE_FORMAT_TOOL_NAME
 from litellm.llms.anthropic.chat.handler import ModelResponseIterator, make_call
-from litellm.llms.custom_httpx.http_handler import HTTPHandler
+from litellm._uuid import uuid
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.types.llms.openai import (
     ChatCompletionToolCallChunk,
     ChatCompletionToolCallFunctionChunk,
@@ -85,6 +86,54 @@ def test_anthropic_completion_does_not_send_deployment_default_limits():
     assert "default_api_key_tpm_limit" not in request_body
 
 
+async def test_anthropic_async_completion_inlines_http_images_off_the_event_loop(async_only_image_fetch):
+    http_image_url = f"http://img.example/{uuid.uuid4()}.png"
+    https_image_url = f"https://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-sonnet-4-6",
+                "content": [{"type": "text", "text": "Green"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="anthropic/claude-sonnet-4-6",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": http_image_url}},
+                    {"type": "image_url", "image_url": {"url": https_image_url}},
+                ],
+            }
+        ],
+        api_key="test-key",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [http_image_url]
+    sources = [part["source"] for part in captured["body"]["messages"][0]["content"] if part["type"] == "image"]
+    assert sources == [
+        {"type": "base64", "media_type": "image/png", "data": async_only_image_fetch.base64_png},
+        {"type": "url", "url": https_image_url},
+    ]
+
+
 def test_redacted_thinking_content_block_delta():
     chunk = {
         "type": "content_block_start",

tests/test_litellm/llms/bedrock/embed/test_bedrock_async_invoke_embedding.py

@@ -184,6 +184,45 @@ class TestBedrockAsyncInvokeEmbedding:
             request_url = mock_post.call_args.kwargs.get("url", "")
             assert "/async-invoke" in request_url
 
+    def test_async_invoke_marengo_3_wraps_the_nested_payload_with_the_base_model_id(self):
+        client = HTTPHandler()
+
+        with patch.object(client, "post") as mock_post:
+            mock_response = Mock()
+            mock_response.status_code = 200
+            mock_response.text = json.dumps(async_invoke_response)
+            mock_response.json = lambda: json.loads(mock_response.text)
+            mock_post.return_value = mock_response
+
+            response = litellm.embedding(
+                model="bedrock/async_invoke/twelvelabs.marengo-embed-3-0-v1:0",
+                input="s3://test-bucket/clip.mp4",
+                client=client,
+                aws_region_name="us-east-1",
+                aws_bedrock_runtime_endpoint="https://bedrock-runtime.us-east-1.amazonaws.com",
+                api_key="test-bearer-token-12345",
+                input_type="video",
+                embeddingOption=["visual", "audio"],
+                segmentation={"method": "fixed", "fixed": {"durationSec": 6}},
+                bucketOwner="123456789012",
+                output_s3_uri="s3://test-bucket/async-invoke-output/",
+            )
+
+        assert response._hidden_params._invocation_arn == async_invoke_response["invocationArn"]
+        assert mock_post.call_args.kwargs["url"].endswith("/async-invoke")
+        assert json.loads(mock_post.call_args.kwargs["data"]) == {
+            "modelId": "twelvelabs.marengo-embed-3-0-v1:0",
+            "modelInput": {
+                "inputType": "video",
+                "video": {
+                    "mediaSource": {"s3Location": {"uri": "s3://test-bucket/clip.mp4", "bucketOwner": "123456789012"}},
+                    "segmentation": {"method": "fixed", "fixed": {"durationSec": 6}},
+                    "embeddingOption": ["visual", "audio"],
+                },
+            },
+            "outputDataConfig": {"s3OutputDataConfig": {"s3Uri": "s3://test-bucket/async-invoke-output/"}},
+        }
+
     @pytest.mark.asyncio
     async def test_async_invoke_twelvelabs_embedding_async_with_mock(self):
         """Test async invoke embedding with async calls."""

tests/test_litellm/llms/bedrock/embed/test_bedrock_embedding.py

@@ -5,6 +5,7 @@ from unittest.mock import Mock, patch
 import pytest
 
 import litellm
+from litellm.llms.bedrock.embed.twelvelabs_marengo_transformation import TwelveLabsMarengoEmbeddingConfig
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 
 # Mock responses for different embedding models
@@ -1059,3 +1060,182 @@ def test_bedrock_embedding_bearer_token_never_runs_the_sigv4_credential_chain(mo
 
     assert response.data[0]["embedding"] == titan_embedding_response["embedding"]
     assert mock_post.call_args.kwargs["headers"]["Authorization"] == "Bearer env-bearer-token-12345"
+
+
+marengo_3_embedding_response = {"data": [{"embedding": [0.01 * i for i in range(512)]}]}
+MARENGO_3_DUCK = "data:image/png;base64,ZHVjaw=="
+
+
+@pytest.mark.parametrize(
+    "model,kwargs,expected_body,expected_usage_details",
+    [
+        (
+            "bedrock/us.twelvelabs.marengo-embed-3-0-v1:0",
+            {"input_type": "text"},
+            {"inputType": "text", "text": {"inputText": "a duck on water"}},
+            {"query_count": 1},
+        ),
+        (
+            "bedrock/twelvelabs.marengo-embed-3-0-v1:0",
+            {"input_type": "text"},
+            {"inputType": "text", "text": {"inputText": "a duck on water"}},
+            {"query_count": 1},
+        ),
+        (
+            "bedrock/us.twelvelabs.marengo-embed-3-0-v1:0",
+            {"input_type": "text_image", "media_source": MARENGO_3_DUCK},
+            {
+                "inputType": "text_image",
+                "text_image": {"inputText": "a duck on water", "mediaSource": {"base64String": "ZHVjaw=="}},
+            },
+            {"query_count": 1, "image_count": 1},
+        ),
+        (
+            "bedrock/us.twelvelabs.marengo-embed-3-0-v1:0",
+            {"input_type": "multi_input", "media_sources": {"bird": MARENGO_3_DUCK}},
+            {
+                "inputType": "multi_input",
+                "multi_input": {
+                    "inputText": "a duck on water",
+                    "mediaSources": [{"name": "bird", "mediaType": "image", "base64String": "ZHVjaw=="}],
+                },
+            },
+            {"query_count": 1, "image_count": 1},
+        ),
+    ],
+)
+def test_marengo_3_embedding_sends_the_nested_payload_and_parses_512_dims(
+    model, kwargs, expected_body, expected_usage_details
+):
+    client = HTTPHandler()
+
+    with patch.object(client, "post") as mock_post:
+        mock_response = Mock()
+        mock_response.status_code = 200
+        mock_response.text = json.dumps(marengo_3_embedding_response)
+        mock_response.json = lambda: json.loads(mock_response.text)
+        mock_post.return_value = mock_response
+
+        response = litellm.embedding(
+            model=model,
+            input="a duck on water",
+            client=client,
+            aws_region_name="us-east-1",
+            api_key="test-bearer-token-12345",
+            **kwargs,
+        )
+
+    assert json.loads(mock_post.call_args.kwargs["data"]) == expected_body
+    assert mock_post.call_args.kwargs["url"].endswith(f"/model/{model.removeprefix('bedrock/').replace(':', '%3A')}/invoke")
+    assert len(response.data[0]["embedding"]) == 512
+    assert response.data[0]["embedding"][:2] == [0.0, 0.01]
+    assert response.usage.prompt_tokens == 0
+    assert response.usage.total_tokens == 0
+    assert response.usage.prompt_tokens_details.model_dump(exclude_none=True) == expected_usage_details
+
+
+def test_marengo_3_image_embedding_sends_the_media_under_the_image_key():
+    client = HTTPHandler()
+
+    with patch.object(client, "post") as mock_post:
+        mock_response = Mock()
+        mock_response.status_code = 200
+        mock_response.text = json.dumps(marengo_3_embedding_response)
+        mock_response.json = lambda: json.loads(mock_response.text)
+        mock_post.return_value = mock_response
+
+        response = litellm.embedding(
+            model="bedrock/us.twelvelabs.marengo-embed-3-0-v1:0",
+            input=MARENGO_3_DUCK,
+            client=client,
+            aws_region_name="us-east-1",
+            api_key="test-bearer-token-12345",
+            input_type="image",
+        )
+
+    assert json.loads(mock_post.call_args.kwargs["data"]) == {
+        "inputType": "image",
+        "image": {"mediaSource": {"base64String": "ZHVjaw=="}},
+    }
+    assert len(response.data[0]["embedding"]) == 512
+    assert response.data[0]["embedding"][:2] == [0.0, 0.01]
+    assert response.usage.prompt_tokens == 0
+    assert response.usage.prompt_tokens_details.model_dump(exclude_none=True) == {"image_count": 1}
+
+
+def test_marengo_2_7_embedding_keeps_the_flat_payload():
+    client = HTTPHandler()
+
+    with patch.object(client, "post") as mock_post:
+        mock_response = Mock()
+        mock_response.status_code = 200
+        mock_response.text = json.dumps(twelvelabs_embedding_response)
+        mock_response.json = lambda: json.loads(mock_response.text)
+        mock_post.return_value = mock_response
+
+        response = litellm.embedding(
+            model="bedrock/us.twelvelabs.marengo-embed-2-7-v1:0",
+            input="a duck on water",
+            client=client,
+            aws_region_name="us-east-1",
+            api_key="test-bearer-token-12345",
+            input_type="text",
+        )
+
+    assert json.loads(mock_post.call_args.kwargs["data"]) == {
+        "inputType": "text",
+        "inputText": "a duck on water",
+        "textTruncate": "end",
+    }
+    assert response.data[0]["embedding"] == [0.1, 0.2, 0.3]
+    assert response.usage.prompt_tokens == 0
+    assert response.usage.prompt_tokens_details.model_dump(exclude_none=True) == {"query_count": 1}
+
+
+def test_marengo_usage_counts_text_requests_and_images_across_a_batch():
+    duck = {"mediaType": "image", "base64String": "ZHVjaw=="}
+    response = TwelveLabsMarengoEmbeddingConfig()._transform_response(
+        response_list=[marengo_3_embedding_response, marengo_3_embedding_response, marengo_3_embedding_response],
+        model="us.twelvelabs.marengo-embed-3-0-v1:0",
+        batch_data=[
+            {"inputType": "text", "text": {"inputText": "a duck"}},
+            {"inputType": "image", "image": {"mediaSource": {"base64String": "ZHVjaw=="}}},
+            {"inputType": "multi_input", "multi_input": {"mediaSources": [{"name": "a", **duck}, {"name": "b", **duck}]}},
+        ],
+    )
+
+    assert [item["index"] for item in response.data] == [0, 1, 2]
+    assert response.usage.prompt_tokens == 0
+    assert response.usage.total_tokens == 0
+    assert response.usage.prompt_tokens_details.model_dump(exclude_none=True) == {"query_count": 1, "image_count": 3}
+
+
+def test_marengo_usage_without_request_data_bills_nothing():
+    response = TwelveLabsMarengoEmbeddingConfig()._transform_response(
+        response_list=[marengo_3_embedding_response], model="us.twelvelabs.marengo-embed-3-0-v1:0"
+    )
+
+    assert len(response.data[0]["embedding"]) == 512
+    assert response.usage.prompt_tokens == 0
+    assert response.usage.prompt_tokens_details is None
+
+
+def test_marengo_response_items_without_an_embedding_are_skipped():
+    response = TwelveLabsMarengoEmbeddingConfig()._transform_response(
+        response_list=[{"data": [{"embeddingOption": "visual-text", "startSec": 0.0}, {"embedding": [0.1, 0.2, 0.3]}]}],
+        model="us.twelvelabs.marengo-embed-3-0-v1:0",
+    )
+
+    assert [item["embedding"] for item in response.data] == [[0.1, 0.2, 0.3]]
+    assert response.data[0]["index"] == 0
+
+
+def test_marengo_3_text_image_without_media_source_is_a_bad_request():
+    with pytest.raises(litellm.BadRequestError, match=r"text_image.*media_source"):
+        litellm.embedding(
+            model="bedrock/us.twelvelabs.marengo-embed-3-0-v1:0",
+            input="a duck on water",
+            aws_region_name="us-east-1",
+            api_key="test-bearer-token-12345",
+            input_type="text_image",
+        )

tests/test_litellm/llms/bedrock/embed/test_twelvelabs_marengo_3_transformation.py

@@ -0,0 +1,416 @@
+import json
+from unittest.mock import Mock, patch
+
+import pytest
+
+import litellm
+from litellm.llms.bedrock.common_utils import BedrockError
+from litellm.llms.custom_httpx.http_handler import HTTPHandler
+from litellm.llms.bedrock.embed.twelvelabs_marengo_3_transformation import (
+    MARENGO_2_7_ONLY_PARAMS,
+    build_marengo_3_request,
+    is_marengo_3_model,
+)
+from litellm.llms.bedrock.embed.twelvelabs_marengo_transformation import (
+    TwelveLabsMarengoEmbeddingConfig,
+    drop_params_enabled,
+)
+
+MARENGO_3_BASE = "twelvelabs.marengo-embed-3-0-v1:0"
+MARENGO_3_US = "us.twelvelabs.marengo-embed-3-0-v1:0"
+MARENGO_27_US = "us.twelvelabs.marengo-embed-2-7-v1:0"
+DUCK_DATA_URL = "data:image/png;base64,ZHVjaw=="
+OUTPUT_S3_URI = "s3://out-bucket/marengo/"
+
+
+@pytest.mark.parametrize(
+    "model,expected",
+    [
+        (MARENGO_3_BASE, True),
+        (MARENGO_3_US, True),
+        ("eu.twelvelabs.marengo-embed-3-0-v1:0", True),
+        ("async_invoke/twelvelabs.marengo-embed-3-0-v1:0", True),
+        (MARENGO_27_US, False),
+        ("twelvelabs.marengo-embed-2-7-v1:0", False),
+        ("twelvelabs.marengo-embed-30-v1:0", False),
+        (None, False),
+    ],
+)
+def test_is_marengo_3_model(model, expected):
+    assert is_marengo_3_model(model) is expected
+
+
+def wire(request: object) -> object:
+    return json.loads(json.dumps(request))
+
+
+def test_text_request_nests_input_text_under_text():
+    assert build_marengo_3_request("a dog on the beach", {"input_type": "text"}) == {
+        "inputType": "text",
+        "text": {"inputText": "a dog on the beach"},
+    }
+
+
+def test_missing_input_type_defaults_to_text():
+    assert build_marengo_3_request("hello", {})["inputType"] == "text"
+
+
+def test_camel_case_input_type_wins_over_snake_case():
+    request = build_marengo_3_request(DUCK_DATA_URL, {"inputType": "image", "input_type": "text"})
+    assert request["inputType"] == "image"
+
+
+def test_image_request_strips_data_url_prefix():
+    assert build_marengo_3_request(DUCK_DATA_URL, {"input_type": "image"}) == {
+        "inputType": "image",
+        "image": {"mediaSource": {"base64String": "ZHVjaw=="}},
+    }
+
+
+def test_image_request_from_s3_carries_bucket_owner():
+    request = build_marengo_3_request("s3://media/duck.png", {"input_type": "image", "bucketOwner": "123456789012"})
+    assert request == {
+        "inputType": "image",
+        "image": {"mediaSource": {"s3Location": {"uri": "s3://media/duck.png", "bucketOwner": "123456789012"}}},
+    }
+
+
+@pytest.mark.parametrize(
+    "input_media,params",
+    [
+        ("s3://media/duck.png", {"input_type": "image"}),
+        ("s3://media/clip.mp4", {"input_type": "video"}),
+        ("a duck", {"input_type": "text_image", "media_source": "s3://media/duck.png"}),
+        ("a duck", {"input_type": "multi_input", "media_sources": {"img1": "s3://media/duck.png"}}),
+    ],
+)
+def test_s3_media_without_bucket_owner_is_rejected_naming_it(input_media, params):
+    with pytest.raises(BedrockError) as excinfo:
+        build_marengo_3_request(input_media, params)
+    assert excinfo.value.status_code == 400
+    assert excinfo.value.message == (
+        "s3:// media requires the 'bucketOwner' parameter, the account id that owns the bucket"
+    )
+
+
+def test_text_image_request_pairs_text_with_media_source():
+    request = build_marengo_3_request(
+        "a duck", {"input_type": "text_image", "media_source": DUCK_DATA_URL, "output_s3_uri": OUTPUT_S3_URI}
+    )
+    assert request == {
+        "inputType": "text_image",
+        "text_image": {"inputText": "a duck", "mediaSource": {"base64String": "ZHVjaw=="}},
+    }
+
+
+def test_text_image_request_requires_media_source():
+    with pytest.raises(BedrockError, match=r"text_image.*media_source") as excinfo:
+        build_marengo_3_request("a duck", {"input_type": "text_image"})
+    assert excinfo.value.status_code == 400
+
+
+def test_multi_input_request_names_each_media_source():
+    request = build_marengo_3_request(
+        "a photo of <@bird> next to <@dog>",
+        {
+            "input_type": "multi_input",
+            "media_sources": {"bird": DUCK_DATA_URL, "dog": "s3://media/dog.png"},
+            "bucketOwner": "123456789012",
+        },
+    )
+    assert wire(request) == {
+        "inputType": "multi_input",
+        "multi_input": {
+            "inputText": "a photo of <@bird> next to <@dog>",
+            "mediaSources": [
+                {"name": "bird", "mediaType": "image", "base64String": "ZHVjaw=="},
+                {
+                    "name": "dog",
+                    "mediaType": "image",
+                    "s3Location": {"uri": "s3://media/dog.png", "bucketOwner": "123456789012"},
+                },
+            ],
+        },
+    }
+
+
+def test_multi_input_without_text_omits_input_text():
+    request = build_marengo_3_request("", {"input_type": "multi_input", "media_sources": {"bird": DUCK_DATA_URL}})
+    assert "inputText" not in request["multi_input"]
+    assert request["multi_input"]["mediaSources"][0]["name"] == "bird"
+
+
+@pytest.mark.parametrize("params", [{"input_type": "multi_input"}, {"input_type": "multi_input", "media_sources": {}}])
+def test_multi_input_request_requires_media_sources(params):
+    with pytest.raises(BedrockError, match=r"multi_input.*media_sources") as excinfo:
+        build_marengo_3_request("<@bird>", params)
+    assert excinfo.value.status_code == 400
+
+
+@pytest.mark.parametrize("input_type", ["video", "audio"])
+def test_timed_media_request_nests_every_option_under_the_media_key(input_type):
+    request = build_marengo_3_request(
+        "s3://media/clip.mp4",
+        {
+            "input_type": input_type,
+            "startSec": 2,
+            "endSec": 12.5,
+            "segmentation": {"method": "dynamic", "dynamic": {"minDurationSec": 4}},
+            "embeddingOption": ["visual", "audio"],
+            "embeddingType": ["fused_embedding"],
+            "embeddingScope": ["clip", "asset"],
+            "inferenceId": "req-42",
+            "bucketOwner": "123456789012",
+        },
+    )
+    assert wire(request) == {
+        "inputType": input_type,
+        input_type: {
+            "mediaSource": {"s3Location": {"uri": "s3://media/clip.mp4", "bucketOwner": "123456789012"}},
+            "startSec": 2.0,
+            "endSec": 12.5,
+            "segmentation": {"method": "dynamic", "dynamic": {"minDurationSec": 4}},
+            "embeddingOption": ["visual", "audio"],
+            "embeddingType": ["fused_embedding"],
+            "embeddingScope": ["clip", "asset"],
+        },
+        "inferenceId": "req-42",
+    }
+
+
+def test_timed_media_request_without_options_carries_only_the_media_source():
+    request = build_marengo_3_request("s3://media/clip.mp4", {"input_type": "video", "bucketOwner": "123456789012"})
+    assert request["video"] == {
+        "mediaSource": {"s3Location": {"uri": "s3://media/clip.mp4", "bucketOwner": "123456789012"}}
+    }
+
+
+@pytest.mark.parametrize(
+    "params",
+    [
+        {"input_type": "clip"},
+        {"input_type": "video", "embeddingOption": ["visual-text"]},
+        {"input_type": "video", "segmentation": {"method": "fixed", "dynamic": {"minDurationSec": 4}}},
+        {"input_type": "multi_input", "media_sources": ["not", "a", "mapping"]},
+    ],
+)
+def test_invalid_marengo_3_params_are_rejected_before_the_request_is_sent(params):
+    with pytest.raises(BedrockError, match=r"Invalid Marengo 3\.0 parameters") as excinfo:
+        build_marengo_3_request("s3://media/clip.mp4", params)
+    assert excinfo.value.status_code == 400
+
+
+def test_config_sends_the_nested_payload_for_marengo_3_and_the_flat_one_for_2_7():
+    nested = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US)._transform_request(
+        input="hello", inference_params={"input_type": "text"}
+    )
+    flat = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_27_US)._transform_request(
+        input="hello", inference_params={"input_type": "text"}
+    )
+    assert nested == {"inputType": "text", "text": {"inputText": "hello"}}
+    assert flat == {"inputType": "text", "inputText": "hello", "textTruncate": "end"}
+
+
+def test_config_without_a_model_keeps_the_2_7_payload():
+    request = TwelveLabsMarengoEmbeddingConfig()._transform_request(input="hello", inference_params={})
+    assert request == {"inputType": "text", "inputText": "hello", "textTruncate": "end"}
+
+
+@pytest.mark.parametrize("input_type", ["video", "audio"])
+def test_marengo_3_video_and_audio_still_require_the_async_route(input_type):
+    with pytest.raises(ValueError, match=f"Input type '{input_type}' requires async_invoke route"):
+        TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_BASE)._transform_request(
+            input="s3://media/clip.mp4", inference_params={"input_type": input_type}
+        )
+
+
+def test_marengo_3_async_invoke_wraps_the_nested_payload_with_the_base_model_id():
+    request = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_BASE)._transform_request(
+        input="s3://media/clip.mp4",
+        inference_params={
+            "input_type": "video",
+            "embeddingOption": ["visual"],
+            "bucketOwner": "123456789012",
+            "output_s3_uri": OUTPUT_S3_URI,
+        },
+        async_invoke_route=True,
+        model_id="async_invoke%2Ftwelvelabs.marengo-embed-3-0-v1%3A0",
+        output_s3_uri=OUTPUT_S3_URI,
+    )
+    assert wire(request) == {
+        "modelId": MARENGO_3_BASE,
+        "modelInput": {
+            "inputType": "video",
+            "video": {
+                "mediaSource": {"s3Location": {"uri": "s3://media/clip.mp4", "bucketOwner": "123456789012"}},
+                "embeddingOption": ["visual"],
+            },
+        },
+        "outputDataConfig": {"s3OutputDataConfig": {"s3Uri": OUTPUT_S3_URI}},
+    }
+
+
+def test_marengo_3_async_invoke_requires_an_output_s3_uri():
+    with pytest.raises(ValueError, match="output_s3_uri cannot be empty"):
+        TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_BASE)._transform_request(
+            input="hello",
+            inference_params={"input_type": "text"},
+            async_invoke_route=True,
+            model_id=MARENGO_3_BASE,
+            output_s3_uri="",
+        )
+
+
+def test_encoding_format_float_no_longer_injects_2_7_embedding_options_for_marengo_3():
+    marengo_3 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US).map_openai_params(
+        non_default_params={"encoding_format": "float"}, optional_params={}
+    )
+    marengo_27 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_27_US).map_openai_params(
+        non_default_params={"encoding_format": "float"}, optional_params={}
+    )
+    assert marengo_3 == {}
+    assert marengo_27 == {"embeddingOption": ["visual-text", "visual-image"]}
+
+
+def test_marengo_3_only_params_are_forwarded_by_map_openai_params():
+    mapped = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US).map_openai_params(
+        non_default_params={
+            "input_type": "text_image",
+            "media_source": DUCK_DATA_URL,
+            "media_sources": {"bird": DUCK_DATA_URL},
+            "endSec": 5,
+            "segmentation": {"method": "fixed", "fixed": {"durationSec": 6}},
+            "embeddingType": ["separate_embedding"],
+            "embeddingScope": ["clip"],
+            "inferenceId": "req-1",
+        },
+        optional_params={},
+    )
+    assert mapped == {
+        "inputType": "text_image",
+        "media_source": DUCK_DATA_URL,
+        "media_sources": {"bird": DUCK_DATA_URL},
+        "endSec": 5,
+        "segmentation": {"method": "fixed", "fixed": {"durationSec": 6}},
+        "embeddingType": ["separate_embedding"],
+        "embeddingScope": ["clip"],
+        "inferenceId": "req-1",
+    }
+
+
+@pytest.mark.parametrize(
+    "params,problem",
+    [
+        (
+            {"input_type": "clip"},
+            "input_type: Input should be 'text', 'image', 'video', 'audio', 'text_image' or 'multi_input'",
+        ),
+        ({"input_type": "video", "embeddingOption": "visual"}, "embeddingOption: Input should be a valid tuple"),
+        (
+            {"input_type": "multi_input", "media_sources": ["not", "a", "mapping"]},
+            "media_sources: Input should be a valid dictionary",
+        ),
+    ],
+)
+def test_invalid_marengo_3_params_name_the_field_and_the_reason(params, problem):
+    with pytest.raises(BedrockError) as excinfo:
+        build_marengo_3_request("s3://media/clip.mp4", params)
+    assert excinfo.value.message == f"Invalid Marengo 3.0 parameters: {problem}"
+
+
+MARENGO_2_7_ONLY_VALUES = {"textTruncate": "end", "lengthSec": 5, "useFixedLengthSec": True, "minClipSec": 2}
+
+
+@pytest.mark.parametrize("name", MARENGO_2_7_ONLY_PARAMS)
+def test_marengo_2_7_only_params_are_rejected_on_3_0_unless_dropped(name):
+    params = {"input_type": "text", name: MARENGO_2_7_ONLY_VALUES[name]}
+    with pytest.raises(BedrockError) as excinfo:
+        build_marengo_3_request("hello", params)
+    assert excinfo.value.status_code == 400
+    assert excinfo.value.message == (
+        f"Marengo 3.0 does not accept the Marengo 2.7 parameters {name}; set drop_params to drop them"
+    )
+    assert build_marengo_3_request("hello", params, drop_params=True) == {
+        "inputType": "text",
+        "text": {"inputText": "hello"},
+    }
+
+
+def test_marengo_2_7_only_params_are_advertised_only_for_2_7():
+    marengo_3 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US).get_supported_openai_params()
+    marengo_27 = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_27_US).get_supported_openai_params()
+    assert set(MARENGO_2_7_ONLY_PARAMS).isdisjoint(marengo_3)
+    assert set(MARENGO_2_7_ONLY_PARAMS) <= set(marengo_27)
+    assert set(marengo_3) <= set(marengo_27)
+
+
+def test_drop_params_comes_from_the_call_or_the_global(monkeypatch):
+    monkeypatch.setattr(litellm, "drop_params", False)
+    assert drop_params_enabled({}) is False
+    assert drop_params_enabled({"drop_params": True}) is True
+    monkeypatch.setattr(litellm, "drop_params", True)
+    assert drop_params_enabled({}) is True
+
+
+def test_config_drops_marengo_2_7_only_params_only_when_asked():
+    config = TwelveLabsMarengoEmbeddingConfig(model=MARENGO_3_US)
+    with pytest.raises(BedrockError, match=r"Marengo 2\.7 parameters textTruncate"):
+        config._transform_request("hello", {"textTruncate": "end"})
+    assert config._transform_request("hello", {"textTruncate": "end"}, drop_params=True) == {
+        "inputType": "text",
+        "text": {"inputText": "hello"},
+    }
+
+
+@pytest.mark.parametrize(
+    "params",
+    [
+        {"input_type": "text"},
+        {"input_type": "image"},
+        {"input_type": "text_image", "media_source": DUCK_DATA_URL},
+        {"input_type": "multi_input", "media_sources": {"bird": DUCK_DATA_URL}},
+    ],
+)
+def test_timed_media_options_are_rejected_on_untimed_input_types_unless_dropped(params):
+    timed = {**params, "startSec": 0, "embeddingOption": ["visual"]}
+    with pytest.raises(BedrockError) as excinfo:
+        build_marengo_3_request(DUCK_DATA_URL, timed)
+    assert excinfo.value.status_code == 400
+    assert excinfo.value.message == (
+        f"Input type '{params['input_type']}' does not accept startSec, embeddingOption; set drop_params to drop them"
+    )
+    assert build_marengo_3_request(DUCK_DATA_URL, timed, drop_params=True) == build_marengo_3_request(
+        DUCK_DATA_URL, params
+    )
+
+
+def _embed_marengo_3_us(client: HTTPHandler, **params: object):
+    return litellm.embedding(
+        model=f"bedrock/{MARENGO_3_US}",
+        input="hello",
+        client=client,
+        aws_region_name="us-east-1",
+        aws_bedrock_runtime_endpoint="https://bedrock-runtime.us-east-1.amazonaws.com",
+        api_key="test-bearer-token",
+        **params,
+    )
+
+
+def test_per_request_drop_params_reaches_the_marengo_3_builder(monkeypatch):
+    monkeypatch.setattr(litellm, "drop_params", False)
+    client = HTTPHandler()
+    with patch.object(client, "post") as mock_post:
+        mock_response = Mock()
+        mock_response.status_code = 200
+        mock_response.text = json.dumps({"data": [{"embedding": [0.1, 0.2]}]})
+        mock_response.json = lambda: json.loads(mock_response.text)
+        mock_post.return_value = mock_response
+
+        with pytest.raises(litellm.BadRequestError, match=r"Marengo 2\.7 parameters textTruncate"):
+            _embed_marengo_3_us(client, textTruncate="end")
+        assert mock_post.call_count == 0
+
+        response = _embed_marengo_3_us(client, textTruncate="end", drop_params=True)
+
+    assert response.data[0]["embedding"] == [0.1, 0.2]
+    assert json.loads(mock_post.call_args.kwargs["data"]) == {"inputType": "text", "text": {"inputText": "hello"}}

tests/test_litellm/llms/ollama/test_ollama_completion_transformation.py

@@ -2,9 +2,11 @@ import json
 from litellm._uuid import uuid
 from unittest.mock import MagicMock, patch
 
+import httpx
 import pytest
 
-
+import litellm
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.llms.ollama.completion.transformation import (
     OllamaConfig,
     OllamaTextCompletionResponseIterator,
@@ -502,3 +504,43 @@ class TestOllamaTextCompletionResponseIterator:
         assert result["usage"]["prompt_tokens"] == 10
         assert result["usage"]["completion_tokens"] == 5
         assert result["usage"]["total_tokens"] == 15
+
+
+async def test_ollama_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch):
+    image_url = f"https://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "model": "llava",
+                "response": "Green",
+                "done": True,
+                "prompt_eval_count": 1,
+                "eval_count": 1,
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="ollama/llava",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        api_base="http://ollama.example:11434",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    assert captured["body"]["images"] == [async_only_image_fetch.base64_png]

tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/anthropic/test_vertex_ai_anthropic_image_url_handling.py

@@ -6,11 +6,15 @@ Vertex AI Anthropic models don't support URL sources for images.
 LiteLLM should convert image URLs to base64 when using Vertex AI Anthropic.
 """
 
+import json
+import sys
 from unittest.mock import patch, MagicMock
 
+import httpx
 import pytest
 
-
+import litellm
+from litellm._uuid import uuid
 from litellm.litellm_core_utils.prompt_templates.factory import (
     anthropic_messages_pt,
     convert_to_anthropic_tool_result,
@@ -371,3 +375,59 @@ class TestToolMessageImageURLHandling:
                                 assert item["source"]["type"] == "url"
                                 return
         pytest.fail("Could not find image in tool result")
+
+
+async def test_vertex_ai_anthropic_async_completion_inlines_https_images_off_the_event_loop(async_only_image_fetch):
+    from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
+
+    image_url = f"https://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-sonnet-4-6",
+                "content": [{"type": "text", "text": "Green"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+    vertexai = MagicMock()
+    vertexai.preview.language_models = MagicMock()
+
+    with (
+        patch.dict(sys.modules, {"vertexai": vertexai}),
+        patch.object(  # test-quality-ok: litellm.acompletion has no seam for Vertex token minting
+            litellm.main.vertex_partner_models_chat_completion,
+            "_ensure_access_token",
+            return_value=("token", "test-project"),
+        ),
+    ):
+        response = await litellm.acompletion(
+            model="vertex_ai/claude-sonnet-4-6",
+            messages=[
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "What colour is this?"},
+                        {"type": "image_url", "image_url": {"url": image_url}},
+                    ],
+                }
+            ],
+            vertex_project="test-project",
+            vertex_location="us-east5",
+            client=client,
+        )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    sources = [part["source"] for part in captured["body"]["messages"][0]["content"] if part["type"] == "image"]
+    assert sources == [{"type": "base64", "media_type": "image/png", "data": async_only_image_fetch.base64_png}]

tests/test_litellm/llms/watsonx/test_watsonx.py

@@ -356,6 +356,74 @@ async def test_watsonx_gpt_oss_uses_async_http_handler():
         assert result["status"] == "success", "Should return success status"
 
 
+@pytest.mark.parametrize("tokenizer_config_cached", [False, True], ids=["tokenizer_config", "cached_config_jinja"])
+async def test_watsonx_text_gpt_oss_async_completion_fetches_hf_template_off_the_event_loop(
+    monkeypatch, tokenizer_config_cached
+):
+    import httpx
+
+    from litellm._uuid import uuid
+    from litellm.litellm_core_utils.prompt_templates import huggingface_template_handler
+    from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
+
+    hf_model = f"openai/gpt-oss-{uuid.uuid4()}"
+    chat_template = "{% for m in messages %}<|{{ m['role'] }}|>{{ m['content'] }}{% endfor %}"
+    if tokenizer_config_cached:
+        cached_config = {"status": "success", "tokenizer": {"bos_token": None, "eos_token": None}}
+        monkeypatch.setattr(litellm, "known_tokenizer_config", {hf_model: cached_config})
+        expected_fetch = f"https://huggingface.co/{hf_model}/raw/main/chat_template.jinja"
+    else:
+        monkeypatch.setattr(litellm, "known_tokenizer_config", {})
+        expected_fetch = f"https://huggingface.co/{hf_model}/raw/main/tokenizer_config.json"
+    hf_fetched = []
+    captured = {}
+
+    def forbid_sync_client():
+        raise AssertionError("sync HuggingFace fetch ran on the request path")
+
+    async def serve_hf_file(url, **kwargs):
+        hf_fetched.append(url)
+        if url.endswith(".jinja"):
+            return httpx.Response(200, content=chat_template.encode())
+        return httpx.Response(200, json={"chat_template": chat_template, "bos_token": None, "eos_token": None})
+
+    monkeypatch.setattr(huggingface_template_handler, "_get_httpx_client", forbid_sync_client)
+    monkeypatch.setattr(huggingface_template_handler, "get_async_httpx_client", lambda **kwargs: Mock(get=serve_hf_file))
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "model_id": hf_model,
+                "results": [
+                    {
+                        "generated_text": "Hi",
+                        "generated_token_count": 1,
+                        "input_token_count": 1,
+                        "stop_reason": "eos_token",
+                    }
+                ],
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model=f"watsonx_text/{hf_model}",
+        messages=[{"role": "user", "content": "Hi there"}],
+        api_base="https://test-api.watsonx.ai",
+        project_id="test-project-id",
+        token="test-token",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Hi"
+    assert hf_fetched == [expected_fetch]
+    assert captured["body"]["input"] == "<|user|>Hi there"
+
+
 def test_watsonx_chat_completion_with_reasoning_effort(monkeypatch):
     """
     Test that 'reasoning_effort' is correctly passed through to the WatsonX API payload.

tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py

@@ -7638,6 +7638,75 @@ class TestMCPMetaTraceCarrier:
         assert _mcp_meta_trace_carrier(SimpleNamespace(meta=only_progress)) is None
 
 
+@pytest.mark.asyncio
+async def test_stateful_mcp_tool_call_uses_current_requests_otel_destinations() -> None:
+    from types import SimpleNamespace
+
+    from mcp.server.lowlevel.server import request_ctx
+    from mcp.shared.context import RequestContext
+
+    from litellm.integrations.otel.model.destination import OtelDestination
+    from litellm.integrations.otel.plumbing.context import (
+        request_destinations,
+        reset_request_destinations,
+        set_request_destinations,
+    )
+    from litellm.proxy._experimental.mcp_server.mcp_server_manager import (
+        global_mcp_server_manager,
+    )
+    from litellm.proxy._experimental.mcp_server.server import (
+        _MCP_DESTINATIONS_SCOPE_KEY,
+        mcp_server_tool_call,
+        set_auth_context,
+    )
+    from litellm.proxy._experimental.mcp_server.tool_registry import (
+        global_mcp_tool_registry,
+    )
+
+    initialized_destination = OtelDestination(endpoint="https://initialize.example", callback_name="langfuse_otel")
+    current_destination = OtelDestination(endpoint="https://current.example", callback_name="arize")
+    server = MCPServer(
+        server_id="otel-context-test",
+        name="otelcontext",
+        transport=MCPTransport.http,
+        allow_all_keys=True,
+    )
+
+    async def observe_destinations() -> str:
+        assert request_destinations() == (current_destination,)
+        return "ok"
+
+    global_mcp_server_manager.registry[server.server_id] = server
+    global_mcp_server_manager.tool_name_to_mcp_server_name_mapping["otelcontext-observe"] = server.name
+    global_mcp_tool_registry.register_tool(
+        name="otelcontext-observe",
+        description="Observe request destinations",
+        input_schema={"type": "object"},
+        handler=observe_destinations,
+    )
+    set_auth_context(None, raw_headers={})
+    destinations_token = set_request_destinations((initialized_destination,))
+    scope = {_MCP_DESTINATIONS_SCOPE_KEY: (current_destination,)}
+    current_request_context = RequestContext(
+        request_id=1,
+        meta=None,
+        session=SimpleNamespace(),
+        lifespan_context=None,
+        request=SimpleNamespace(scope=scope),
+    )
+    request_token = request_ctx.set(current_request_context)
+    try:
+        result = await mcp_server_tool_call("otelcontext-observe", {})
+        assert result.isError is False
+        assert request_destinations() == (initialized_destination,)
+    finally:
+        request_ctx.reset(request_token)
+        reset_request_destinations(destinations_token)
+        global_mcp_tool_registry.tools.pop("otelcontext-observe", None)
+        global_mcp_server_manager.registry.pop(server.server_id, None)
+        global_mcp_server_manager.tool_name_to_mcp_server_name_mapping.pop("otelcontext-observe", None)
+
+
 @pytest.mark.asyncio
 async def test_get_allowed_mcp_servers_includes_active_servers_submitted_by_user():
     """BYOM submitters can see approved servers they submitted without allow_all_keys."""

tests/test_litellm/proxy/agent_endpoints/test_a2a_endpoints.py

@@ -7,11 +7,24 @@ Tests that invoke_agent_a2a properly integrates with add_litellm_data_to_request
 import json
 import socket
 import sys
-from contextlib import ExitStack
+from collections.abc import Awaitable, Callable, Mapping
+from contextlib import AbstractContextManager, ExitStack
+from dataclasses import dataclass
+from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
+from litellm.proxy._types import UserAPIKeyAuth
+
+AddLiteLLMData = Callable[..., Awaitable[dict[str, object]]]
+
+
+@dataclass(frozen=True, slots=True)
+class CapturedAgentCall:
+    request_id: object
+    agent_extra_headers: dict[str, str] | None
+
 
 @pytest.mark.asyncio
 async def test_invoke_agent_a2a_adds_litellm_data():
@@ -364,7 +377,7 @@ def _make_agent_mock(url: str = "http://backend-agent:10001") -> MagicMock:
 
 
 def _make_request_mock(
-    method: str, params: dict, request_id: object = "req-1"
+    method: str, params: Mapping[str, object], request_id: object = "req-1"
 ) -> MagicMock:
     req = MagicMock()
     req.headers = {}
@@ -379,7 +392,9 @@ def _make_request_mock(
     return req
 
 
-def _base_patches(agent: MagicMock):
+def _base_patches(
+    agent: MagicMock, add_litellm_data: AddLiteLLMData | None = None
+) -> list[AbstractContextManager[object]]:
     return [
         patch(
             "litellm.proxy.agent_endpoints.a2a_endpoints._get_agent",
@@ -391,7 +406,7 @@ def _base_patches(agent: MagicMock):
         ),
         patch(
             "litellm.proxy.common_request_processing.add_litellm_data_to_request",
-            new=AsyncMock(side_effect=_add_proxy_data),
+            new=AsyncMock(side_effect=add_litellm_data or _add_proxy_data),
         ),
         patch("litellm.proxy.proxy_server.general_settings", {}),
         patch("litellm.proxy.proxy_server.proxy_config", MagicMock()),
@@ -399,84 +414,67 @@ def _base_patches(agent: MagicMock):
     ]
 
 
-async def _add_proxy_data(data, **kwargs):
-    data["proxy_server_request"] = {
-        "url": "http://localhost:4000",
-        "method": "POST",
-        "headers": {},
-        "body": {},
+async def _add_proxy_data(data: dict[str, object], **kwargs: object) -> dict[str, object]:
+    return {
+        **data,
+        "proxy_server_request": {"url": "http://localhost:4000", "method": "POST", "headers": {}, "body": {}},
+        "metadata": data.get("metadata", {}),
     }
-    data.setdefault("metadata", {})
-    return data
 
 
-@pytest.mark.asyncio
-@pytest.mark.parametrize("method", ["message/send", "message/stream"])
-async def test_message_methods_preserve_numeric_zero_request_id(method: str):
+_HELLO_MESSAGE_PARAMS = {
+    "message": {
+        "role": "user",
+        "parts": [{"kind": "text", "text": "Hello"}],
+        "messageId": "msg-123",
+    }
+}
+
+
+async def _invoke_message_method(
+    method: str,
+    mock_request: MagicMock,
+    user_api_key_dict: UserAPIKeyAuth,
+    add_litellm_data: AddLiteLLMData | None = None,
+) -> CapturedAgentCall:
     from fastapi.responses import JSONResponse
-    from litellm.proxy._types import UserAPIKeyAuth
 
     class MessageSendParams:
-        def __init__(self, **kwargs):
+        def __init__(self, **kwargs: object) -> None:
             self.__dict__.update(kwargs)
 
     class SendMessageRequest:
-        def __init__(self, **kwargs):
+        def __init__(self, **kwargs: object) -> None:
             self.__dict__.update(kwargs)
 
-    agent = _make_agent_mock()
-    params = {
-        "message": {
-            "role": "user",
-            "parts": [{"kind": "text", "text": "Hello"}],
-            "messageId": "msg-123",
-        }
-    }
-    mock_request = _make_request_mock(method, params, request_id=0)
-    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="u1", team_id="t1")
-    captured = {}
-
-    async def capture_asend_message(request, **kwargs):
-        captured["request_id"] = request.id
-        response = MagicMock()
+    async def fake_asend_message(request: SendMessageRequest, **kwargs: object) -> MagicMock:
+        response: Final = MagicMock()
         response.model_dump.return_value = {
             "jsonrpc": "2.0",
-            "id": request.id,
+            "id": request.__dict__["id"],
             "result": {"status": "success"},
         }
         return response
 
-    async def capture_stream_message(**kwargs):
-        captured["request_id"] = kwargs["request_id"]
-        return JSONResponse({"jsonrpc": "2.0", "id": kwargs["request_id"]})
+    async def fake_stream_message(request_id: object, **kwargs: object) -> JSONResponse:
+        return JSONResponse({"jsonrpc": "2.0", "id": request_id})
 
-    mock_a2a_types = MagicMock()
+    mock_a2a_types: Final = MagicMock()
     mock_a2a_types.MessageSendParams = MessageSendParams
     mock_a2a_types.SendMessageRequest = SendMessageRequest
+    is_send: Final = method == "message/send"
+    downstream: Final = AsyncMock(side_effect=fake_asend_message if is_send else fake_stream_message)
 
     with ExitStack() as stack:
-        for p in _base_patches(agent):
+        for p in _base_patches(_make_agent_mock(), add_litellm_data):
             stack.enter_context(p)
         stack.enter_context(patch("litellm.a2a_protocol.main.A2A_SDK_AVAILABLE", True))
-        if method == "message/send":
-            stack.enter_context(
-                patch.dict(
-                    sys.modules,
-                    {"a2a": MagicMock(), "a2a.types": mock_a2a_types},
-                )
-            )
-            stack.enter_context(
-                patch(
-                    "litellm.a2a_protocol.asend_message",
-                    new=AsyncMock(side_effect=capture_asend_message),
-                )
-            )
+        if is_send:
+            stack.enter_context(patch.dict(sys.modules, {"a2a": MagicMock(), "a2a.types": mock_a2a_types}))
+            stack.enter_context(patch("litellm.a2a_protocol.asend_message", new=downstream))
         else:
             stack.enter_context(
-                patch(
-                    "litellm.proxy.agent_endpoints.a2a_endpoints._handle_stream_message",
-                    new=AsyncMock(side_effect=capture_stream_message),
-                )
+                patch("litellm.proxy.agent_endpoints.a2a_endpoints._handle_stream_message", new=downstream)
             )
 
         from litellm.proxy.agent_endpoints.a2a_endpoints import invoke_agent_a2a
@@ -488,7 +486,82 @@ async def test_message_methods_preserve_numeric_zero_request_id(method: str):
             user_api_key_dict=user_api_key_dict,
         )
 
-    assert captured["request_id"] == 0
+    kwargs: Final = downstream.call_args.kwargs
+    request_id: Final = kwargs["request"].__dict__["id"] if is_send else kwargs["request_id"]
+    return CapturedAgentCall(request_id=request_id, agent_extra_headers=kwargs.get("agent_extra_headers"))
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["message/send", "message/stream"])
+async def test_message_methods_preserve_numeric_zero_request_id(method: str):
+    mock_request = _make_request_mock(method, _HELLO_MESSAGE_PARAMS, request_id=0)
+    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="u1", team_id="t1")
+
+    captured = await _invoke_message_method(method, mock_request, user_api_key_dict)
+
+    assert captured.request_id == 0
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["message/send", "message/stream"])
+async def test_message_methods_forward_caller_identity_headers(method: str):
+    mock_request = _make_request_mock(method, _HELLO_MESSAGE_PARAMS)
+    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="user-abc", team_id="team-xyz")
+
+    captured = await _invoke_message_method(method, mock_request, user_api_key_dict)
+
+    forwarded_headers = captured.agent_extra_headers or {}
+    assert forwarded_headers.get("X-LiteLLM-User-Id") == "user-abc"
+    assert forwarded_headers.get("X-LiteLLM-Team-Id") == "team-xyz"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["message/send", "message/stream"])
+async def test_message_methods_caller_identity_headers_cannot_be_spoofed(method: str):
+    mock_request = _make_request_mock(method, _HELLO_MESSAGE_PARAMS)
+    mock_request.headers = {
+        "x-a2a-test-agent-x-litellm-user-id": "attacker-user",
+        "x-a2a-test-agent-x-litellm-team-id": "attacker-team",
+    }
+    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="real-user", team_id="real-team")
+
+    captured = await _invoke_message_method(method, mock_request, user_api_key_dict)
+
+    forwarded_headers = captured.agent_extra_headers or {}
+    assert (
+        forwarded_headers.get("X-LiteLLM-User-Id") == "real-user"
+    ), "authenticated user id must not be overridden by forwarded client headers"
+    assert (
+        forwarded_headers.get("X-LiteLLM-Team-Id") == "real-team"
+    ), "authenticated team id must not be overridden by forwarded client headers"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["message/send", "message/stream"])
+async def test_message_methods_forward_key_bound_identity_not_pre_call_rewrite(method: str):
+    from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
+
+    mock_request = _make_request_mock(method, _HELLO_MESSAGE_PARAMS)
+    mock_request.headers = {"X-OpenWebUI-User-Id": "header-mapped-user"}
+    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="key-user", team_id="key-team")
+    general_settings: Final = {
+        "user_header_mappings": [{"header_name": "X-OpenWebUI-User-Id", "litellm_user_role": "internal_user"}]
+    }
+
+    async def apply_user_header_mapping(data: dict[str, object], **kwargs: object) -> dict[str, object]:
+        LiteLLMProxyRequestSetup.add_internal_user_from_user_mapping(
+            general_settings, user_api_key_dict, dict(mock_request.headers)
+        )
+        return await _add_proxy_data(data, **kwargs)
+
+    captured = await _invoke_message_method(
+        method, mock_request, user_api_key_dict, add_litellm_data=apply_user_header_mapping
+    )
+
+    assert user_api_key_dict.user_id == "header-mapped-user", "precondition: pre-call rewrite ran"
+    forwarded_headers = captured.agent_extra_headers or {}
+    assert forwarded_headers.get("X-LiteLLM-User-Id") == "key-user"
+    assert forwarded_headers.get("X-LiteLLM-Team-Id") == "key-team"
 
 
 @pytest.mark.asyncio

tests/test_litellm/proxy/agent_endpoints/test_agent_headers.py

@@ -223,7 +223,7 @@ async def test_static_overrides_dynamic():
 
 @pytest.mark.asyncio
 async def test_no_headers():
-    """When no headers are configured, agent_extra_headers is None and behaviour is unchanged."""
+    """When no headers are configured, only the caller identity is forwarded."""
     mock_agent = _make_mock_agent()  # no static_headers or extra_headers
     mock_request = _make_mock_request()
 
@@ -231,7 +231,7 @@ async def test_no_headers():
 
     call_kwargs = mock_asend.call_args.kwargs
     headers = call_kwargs.get("agent_extra_headers")
-    assert headers is None
+    assert headers == {"X-LiteLLM-User-Id": "u1"}
 
 
 # ---------------------------------------------------------------------------
@@ -303,7 +303,7 @@ async def test_convention_unrelated_prefix_not_forwarded():
     mock_asend = await _invoke(mock_agent, mock_request, None)
 
     headers = mock_asend.call_args.kwargs.get("agent_extra_headers")
-    assert headers is None
+    assert headers == {"X-LiteLLM-User-Id": "u1"}
 
 
 # ---------------------------------------------------------------------------
@@ -393,7 +393,7 @@ async def test_non_databricks_agent_skips_oauth_resolution():
 
     mock_resolve.assert_not_called()
     headers = mock_asend.call_args.kwargs.get("agent_extra_headers")
-    assert headers == {"x-custom": "v"}
+    assert headers == {"x-custom": "v", "X-LiteLLM-User-Id": "u1"}
     assert "Authorization" not in headers
 
 
@@ -477,7 +477,7 @@ async def test_convention_header_blocked_by_case_variant_static():
 
     headers = mock_asend.call_args.kwargs.get("agent_extra_headers")
     assert headers is not None
-    assert headers == {"Authorization": "Bearer admin-token"}
+    assert headers == {"Authorization": "Bearer admin-token", "X-LiteLLM-User-Id": "u1"}
     assert "authorization" not in headers
 
 

tests/test_litellm/proxy/auth/test_route_checks.py

@@ -3638,3 +3638,191 @@ def test_agent_registry_route_gate_open_to_non_admin_roles(user_role, method, ro
         valid_token=valid_token,
         request_data={},
     )
+TEAM_CALLBACK_ROUTES = (
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/callback",
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/callback/langfuse",
+    # the routes register team_id with the :path converter, so a team id may
+    # contain a slash
+    "/team/tenant/06bda574/callback",
+    "/team/tenant/06bda574/callback/langfuse",
+    # team_id is a free-form string, so it may also contain a colon
+    "/team/tenant:06bda574/callback",
+    "/team/tenant:06bda574/callback/langfuse",
+    # or both, which is the shape neither a "[^:]+" nor a "[^/]+" expansion
+    # of the placeholder reaches on its own
+    "/team/tenant:acme/prod/callback",
+    "/team/tenant:acme/prod/callback/langfuse",
+)
+
+
+def _gate(route, role) -> str:
+    """Drive the real route gate for a non-proxy-admin caller.
+
+    Reports "allowed" when the gate lets the request through to its handler, and
+    the denial message otherwise, so a caller asserts the verdict as a value
+    instead of on whether an exception escaped.
+    """
+    user_obj = LiteLLM_UserTable(
+        user_id="team_admin_user",
+        user_email="team-admin@example.com",
+        user_role=role,
+    )
+    request = MagicMock(spec=Request)
+    request.query_params = {}
+    try:
+        RouteChecks.non_proxy_admin_allowed_routes_check(
+            user_obj=user_obj,
+            _user_role=role,
+            route=route,
+            request=request,
+            valid_token=UserAPIKeyAuth(user_id="team_admin_user", user_role=role),
+            request_data={},
+        )
+    except Exception as exc:
+        return f"denied: {exc}"
+    return "allowed"
+
+
+def test_team_callback_routes_are_self_managed():
+    """The grant has to come from self_managed_routes specifically.
+
+    That list is the one whose entries carry no role predicate, so the handler
+    decides. Granting the same paths through internal_user_routes instead would
+    look identical for an internal_user while silently denying the org admins and
+    view-only roles that list does not cover.
+    """
+    for template in (
+        "/team/{team_id:path}/callback",
+        "/team/{team_id:path}/callback/{callback_name}",
+    ):
+        assert template in LiteLLMRoutes.self_managed_routes.value
+
+
+@pytest.mark.parametrize("route", TEAM_CALLBACK_ROUTES)
+@pytest.mark.parametrize(
+    "role",
+    [
+        LitellmUserRoles.INTERNAL_USER.value,
+        LitellmUserRoles.INTERNAL_USER_VIEW_ONLY.value,
+        LitellmUserRoles.ORG_ADMIN.value,
+    ],
+)
+def test_team_callback_routes_reach_their_handler_for_non_admins(route, role):
+    """A team admin manages their own team's logging callbacks, so the route gate
+    must let a non-proxy-admin through to the handler.
+
+    The handler is what authorizes: every team callback endpoint calls
+    _verify_team_access, which admits only a proxy admin, an org admin for the
+    team, or an admin of that team, and 403s everyone else. Before this, the gate
+    rejected the team admin with a 401 naming proxy admin, so the handler's own
+    check was unreachable for them.
+    """
+    assert _gate(route, role) == "allowed"
+
+
+@pytest.mark.parametrize(
+    "pattern, route, matches",
+    [
+        # a :path placeholder takes what the router's path converter takes
+        ("/team/{team_id:path}/callback", "/team/plain/callback", True),
+        ("/team/{team_id:path}/callback", "/team/tenant/acme/callback", True),
+        ("/team/{team_id:path}/callback", "/team/tenant:acme/callback", True),
+        ("/team/{team_id:path}/callback", "/team/tenant:acme/prod/callback", True),
+        # and still has to reach the template's own suffix
+        ("/team/{team_id:path}/callback", "/team/tenant:acme/disable_logging", False),
+        # a template with a ":" literal after the placeholder keeps the suffix
+        (
+            "/v1beta/models/{model_name:path}:generateContent",
+            "/v1beta/models/gemini-2.5-flash:generateContent",
+            True,
+        ),
+        (
+            "/v1beta/models/{model_name:path}:generateContent",
+            "/v1beta/models/publishers/google/gemini-2.5-flash:generateContent",
+            True,
+        ),
+        # the value must not swallow that suffix and match a different verb
+        (
+            "/v1beta/models/{model_name:path}:generateContent",
+            "/v1beta/models/gemini-2.5-flash:countTokens",
+            False,
+        ),
+        # a %0A in the value reaches the handler through the path converter, so
+        # the gate has to see it too or DISABLE_ADMIN_ENDPOINTS is bypassable
+        ("/v1/mcp/server/{path:path}", "/v1/mcp/server/abc\ndef", True),
+        ("/team/{team_id:path}/callback", "/team/ten\nant/callback", True),
+        ("/v1beta/models/{model_name:path}:generateContent", "/v1beta/models/gem\nini:generateContent", True),
+        # an ordinary placeholder stays one segment
+        ("/team/{team_id}/members/me", "/team/abc/members/me", True),
+        ("/team/{team_id}/members/me", "/team/tenant/abc/members/me", False),
+        ("/team/{team_id}/members/me", "/team/ab\nc/members/me", True),
+    ],
+)
+def test_path_placeholder_matches_what_the_router_accepts(pattern, route, matches):
+    """The gate's placeholder expansion has to agree with the router's.
+
+    A team id may carry a slash, a colon, or both, and the router mounted these
+    paths with the same :path converter, so an id the router routes must not be
+    an id the gate fails to recognize. The one narrowing that stays is a template
+    whose own suffix begins with a colon: there the value stops before it, or
+    ":generateContent" would also match a ":countTokens" request.
+    """
+    assert RouteChecks._route_matches_pattern(route=route, pattern=pattern) is matches
+
+
+# Every other route the proxy mounts under /team/{team_id}, spelled the way it
+# is registered. None of them takes a path converter, so none can be reached by
+# a URL that ends in the callback suffix.
+PROTECTED_TEAM_ROUTES = (
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112",
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/disable_logging",
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/members/me",
+    "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/member/u-1/reset_spend",
+    # the same routes with the callback suffix spliced in, which is the shape a
+    # caller would craft to make a protected route look self-managed
+    "/team/06bda574/callback/disable_logging/x",
+    "/team/06bda574/callback/member/u-1/reset_spend",
+    "/team/06bda574/callback/members/me",
+)
+
+
+@pytest.mark.parametrize("route", PROTECTED_TEAM_ROUTES)
+def test_the_callback_grant_does_not_reach_another_team_route(route):
+    """Widening the callback templates must not hand out any neighbouring route.
+
+    The grant is two templates ending in the callback suffix. Every other team
+    route registers an ordinary single-segment placeholder, so no URL the router
+    sends to one of them can end in "/callback" or "/callback/<name>" -- and the
+    gate must agree, or a crafted team id would carry a caller into a handler
+    the grant never covered.
+    """
+    for template in (
+        "/team/{team_id:path}/callback",
+        "/team/{team_id:path}/callback/{callback_name}",
+    ):
+        assert RouteChecks._route_matches_pattern(route=route, pattern=template) is False
+
+
+def test_team_disable_logging_stays_proxy_admin_only():
+    """disable_logging was left out of the grant, so it must still be rejected at
+    the gate. It is the one team callback route a team admin cannot reach."""
+    verdict = _gate(
+        "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/disable_logging",
+        LitellmUserRoles.INTERNAL_USER.value,
+    )
+
+    assert "Only proxy admin" in verdict
+    assert "disable_logging" in verdict
+
+
+@pytest.mark.parametrize(
+    "route",
+    [
+        "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112",
+        "/team/update",
+        "/team/06bda574-5ca9-43d3-beb8-3b23c2f17112/model/add",
+    ],
+)
+def test_neighbouring_team_routes_stay_closed(route):
+    """The grant is the callback paths and nothing else on the team namespace."""
+    assert "Only proxy admin" in _gate(route, LitellmUserRoles.INTERNAL_USER.value)

tests/test_litellm/proxy/common_utils/test_openai_error_payload.py

@@ -0,0 +1,145 @@
+import json
+
+import pytest
+from fastapi import HTTPException
+
+from litellm.proxy._types import ProxyErrorTypes, ProxyException
+from litellm.proxy.common_utils.openai_error_payload import (
+    error_status_code,
+    openai_error_param,
+    openai_error_type,
+)
+
+
+@pytest.mark.parametrize(
+    "status_code, expected_type",
+    [
+        (400, "invalid_request_error"),
+        (401, "authentication_error"),
+        (403, "permission_error"),
+        (404, "invalid_request_error"),
+        (408, "invalid_request_error"),
+        (422, "invalid_request_error"),
+        (429, "rate_limit_error"),
+        (499, "invalid_request_error"),
+        (500, "internal_server_error"),
+        (502, "internal_server_error"),
+        (503, "internal_server_error"),
+    ],
+)
+def test_status_code_decides_the_type_when_the_exception_carries_none(status_code: int, expected_type: str):
+    """A route that raises a bare HTTPException carries no error type, so the status it
+    answered with is the only thing left to name the OpenAI type from."""
+    assert openai_error_type(HTTPException(status_code=status_code, detail="boom"), status_code) == expected_type
+
+
+def test_a_carried_type_wins_over_the_one_the_status_would_imply():
+    """A ProxyException raised mid-request already names its own type, and relabelling a
+    402 budget_exceeded as the status map's guess would lose what the client branches on."""
+    carried = ProxyException(
+        message="Budget has been exceeded",
+        type=ProxyErrorTypes.budget_exceeded.value,
+        param=None,
+        code=400,
+    )
+
+    assert openai_error_type(carried, 400) == ProxyErrorTypes.budget_exceeded.value
+
+
+@pytest.mark.parametrize("carried_type", [None, 400, {"type": "invalid_request_error"}, ["invalid_request_error"]])
+def test_a_non_string_carried_type_falls_back_to_the_status(carried_type: object):
+    """OpenAI types error.type as a string, so anything else on the exception is not one and
+    must not reach the wire the way the literal "None" used to."""
+
+    class _Carrier(Exception):
+        type = carried_type
+
+    assert openai_error_type(_Carrier("boom"), 401) == "authentication_error"
+
+
+def test_the_type_is_never_the_string_none_after_a_json_round_trip():
+    """The bug this module exists for: json.dumps of a "None" default is indistinguishable
+    from a real type to a client's error handler."""
+    payload = json.loads(
+        json.dumps(
+            {
+                "type": openai_error_type(HTTPException(status_code=400, detail="boom"), 400),
+                "param": openai_error_param(HTTPException(status_code=400, detail="boom")),
+            }
+        )
+    )
+
+    assert payload == {"type": "invalid_request_error", "param": None}
+
+
+def test_a_carried_param_names_the_offending_field():
+    carried = ProxyException(message="Invalid purpose", type="invalid_request_error", param="purpose", code=400)
+
+    assert openai_error_param(carried) == "purpose"
+
+
+@pytest.mark.parametrize("exc", [HTTPException(status_code=400, detail="boom"), ValueError("boom"), None])
+def test_param_is_json_null_when_the_exception_names_no_field(exc: Exception | None):
+    assert openai_error_param(exc) is None
+
+
+def test_a_non_string_carried_param_is_json_null():
+    class _Carrier(Exception):
+        param = 42
+
+    assert openai_error_param(_Carrier("boom")) is None
+
+
+def test_a_carried_status_code_wins_over_the_default():
+    assert error_status_code(HTTPException(status_code=429, detail="slow down"), 400) == 429
+
+
+@pytest.mark.parametrize("default", [400, 500])
+def test_the_default_status_stands_when_the_exception_carries_none(default: int):
+    assert error_status_code(ValueError("boom"), default) == default
+
+
+@pytest.mark.parametrize("carried_status", [True, False, "429", None, 429.0])
+def test_a_non_int_carried_status_falls_back_to_the_default(carried_status: object):
+    """True is an int in Python but not an HTTP status, and a stringified one would break
+    every caller that compares the code numerically."""
+
+    class _Carrier(Exception):
+        status_code = carried_status
+
+    assert error_status_code(_Carrier("boom"), 500) == 500
+
+
+def test_a_proxy_exception_keeps_the_status_it_was_raised_with():
+    """ProxyException stores its status as the string ``code`` rather than ``status_code``,
+    so a route tail that rewraps one used to answer a 4xx rejection as a 500."""
+    rejection = ProxyException(message="session_id is required", type="bad_request_error", param="session_id", code=400)
+
+    assert error_status_code(rejection, 500) == 400
+
+
+@pytest.mark.parametrize("carried_code", [None, "None", "", "rate_limited", "4xx", 404])
+def test_a_code_that_is_not_a_decimal_string_falls_back_to_the_default(carried_code: object):
+    """Only ProxyException's stringified status is a status; ``code`` on anything else
+    (OpenAI's ``invalid_api_key``, a stray int) says nothing about the HTTP answer."""
+
+    class _Carrier(Exception):
+        code = carried_code
+
+    assert error_status_code(_Carrier("boom"), 500) == 500
+
+
+def test_a_status_code_wins_over_a_stringified_code():
+    class _Carrier(Exception):
+        status_code = 429
+        code = "400"
+
+    assert error_status_code(_Carrier("boom"), 500) == 429
+
+
+def test_a_status_carried_by_an_exception_drives_the_type_it_reports():
+    """The two helpers compose at every call site: the status the exception carries is what
+    names its type, not the default the route would have used."""
+    exc = HTTPException(status_code=403, detail="blocked by policy")
+
+    assert openai_error_type(exc, error_status_code(exc, 400)) == "permission_error"

tests/test_litellm/proxy/db/test_db_spend_update_writer.py

@@ -2944,11 +2944,9 @@ async def test_insert_spend_log_asks_for_an_immediate_flush_on_rows_other_worker
     A `previous_response_id` chained straight off the previous turn reads the DB, so a
     Responses row cannot sit in this worker's queue until the monitor's next poll.
     """
-    from litellm.proxy.utils import PrismaClient
-
     db_writer = DBSpendUpdateWriter()
     prisma = _tool_usage_prisma()
-    PrismaClient.spend_log_flush_requested.clear()
+    prisma.spend_log_flush_requested = asyncio.Event()
 
     await db_writer._insert_spend_log_to_db(
         payload={"request_id": "req-1", "call_type": call_type},
@@ -2956,8 +2954,7 @@ async def test_insert_spend_log_asks_for_an_immediate_flush_on_rows_other_worker
     )
 
     assert prisma.spend_log_transactions == [{"request_id": "req-1", "call_type": call_type}]
-    assert PrismaClient.spend_log_flush_requested.is_set() is expects_flush
-    PrismaClient.spend_log_flush_requested.clear()
+    assert prisma.spend_log_flush_requested.is_set() is expects_flush
 
 
 def _batch_cost_payload() -> dict:

tests/test_litellm/proxy/db/test_db_url_settings.py

@@ -35,6 +35,7 @@ _MANAGED_DB_ENV_VARS = (
     "IAM_TOKEN_DB_AUTH",
     "AZURE_POSTGRESQL_AUTH",
     "DATABASE_DISABLE_PREPARED_STATEMENTS",
+    "DATABASE_MAX_IDLE_CONNECTION_LIFETIME",
     "DATABASE_URL",
     "DIRECT_URL",
     "DATABASE_URL_READ_REPLICA",
@@ -111,7 +112,7 @@ def test_assembles_writer_url_when_iam_enabled(monkeypatch):
 
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://litellm:WRITER_TOKEN@writer.example.com:5432/litellm_db"
+        == "postgresql://litellm:WRITER_TOKEN@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
     # Reader was never configured, so it must not have been set.
     assert "DATABASE_URL_READ_REPLICA" not in os.environ
@@ -130,7 +131,9 @@ def test_a_pre_encoded_iam_user_survives_url_assembly(monkeypatch):
     with _stub_iam_token("WRITER_TOKEN"):
         assert _apply() is True
 
-    assert os.environ["DATABASE_URL"] == "postgresql://svc%40corp:WRITER_TOKEN@writer.example.com:5432/litellm_db"
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://svc%40corp:WRITER_TOKEN@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
+    )
 
 
 def test_an_unreadable_toggle_fails_the_settings_model(monkeypatch):
@@ -168,7 +171,7 @@ def test_reader_url_assembled_when_host_set_and_url_unset(monkeypatch):
 
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://litellm:READER_TOKEN@reader.example.com:5432/litellm_db"
+        == "postgresql://litellm:READER_TOKEN@reader.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -191,7 +194,7 @@ def test_reader_url_not_clobbered_when_already_set(monkeypatch):
 
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://app:secret@reader.example.com:5432/litellm_db"
+        == "postgresql://app:secret@reader.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -222,7 +225,8 @@ def test_reader_field_fallbacks_default_to_writer_values(monkeypatch):
 
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://litellm:READER_TOKEN@reader.example.com:5432/litellm_db?schema=public"
+        == "postgresql://litellm:READER_TOKEN@reader.example.com:5432/litellm_db"
+        "?schema=public&max_idle_connection_lifetime=60"
     )
 
 
@@ -242,7 +246,7 @@ def test_assembles_writer_url_when_azure_entra_enabled(monkeypatch):
 
     assert os.environ["DATABASE_URL"] == (
         "postgresql://litellm%40contoso.onmicrosoft.com:ENTRA_TOKEN"
-        "@writer.postgres.database.azure.com:5432/litellm_db"
+        "@writer.postgres.database.azure.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
     assert os.environ["AZURE_POSTGRESQL_AUTH"] == "True"
     assert "IAM_TOKEN_DB_AUTH" not in os.environ
@@ -261,7 +265,7 @@ def test_azure_reader_url_assembled_from_writer_fallbacks(monkeypatch):
 
     assert os.environ["DATABASE_URL_READ_REPLICA"] == (
         "postgresql://litellm%40contoso.onmicrosoft.com:ENTRA_TOKEN"
-        "@reader.postgres.database.azure.com:5432/litellm_db?schema=public"
+        "@reader.postgres.database.azure.com:5432/litellm_db?schema=public&max_idle_connection_lifetime=60"
     )
 
 
@@ -357,7 +361,7 @@ def test_assembles_writer_url_from_password(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db"
+        == "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -370,7 +374,7 @@ def test_writer_password_is_percent_encoded(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://litellm:p%40ss%2Fw%3Ard@writer.example.com:5432/litellm_db"
+        == "postgresql://litellm:p%40ss%2Fw%3Ard@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -388,7 +392,7 @@ def test_writer_url_not_clobbered_when_already_set(monkeypatch):
     assert _apply() is False
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://pinned:url@db.example.com:5432/litellm_db"
+        == "postgresql://pinned:url@db.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -400,7 +404,7 @@ def test_writer_url_passwordless(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://litellm@writer.example.com:5432/litellm_db"
+        == "postgresql://litellm@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -415,7 +419,7 @@ def test_database_username_alias(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL"]
-        == "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db"
+        == "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -429,7 +433,7 @@ def test_password_reader_falls_back_to_writer_password(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://litellm:s3cr3t@reader.example.com:5432/litellm_db"
+        == "postgresql://litellm:s3cr3t@reader.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -445,7 +449,7 @@ def test_password_reader_uses_own_credentials(monkeypatch):
     assert _apply() is True
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://litellm_ro:ro_pw@reader.example.com:5432/litellm_db"
+        == "postgresql://litellm_ro:ro_pw@reader.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
     )
 
 
@@ -641,19 +645,19 @@ def test_reader_keeps_its_own_options_when_writer_params_are_appended(monkeypatc
     assert query["connection_limit"] == ["3"]
 
 
-def test_reader_url_left_alone_when_writer_has_no_params(monkeypatch):
+def test_reader_url_left_alone_when_nothing_is_missing(monkeypatch):
     """No params to inherit must mean the reader URL is not rewritten at all."""
     monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@writer.example.com:5432/db")
     monkeypatch.setenv(
         "DATABASE_URL_READ_REPLICA",
-        "postgresql://u:p@reader.example.com:5432/db?options=-c%20search_path%3Dapp",
+        "postgresql://u:p@reader.example.com:5432/db?options=-c%20search_path%3Dapp&max_idle_connection_lifetime=45",
     )
 
     _apply()
 
     assert (
         os.environ["DATABASE_URL_READ_REPLICA"]
-        == "postgresql://u:p@reader.example.com:5432/db?options=-c%20search_path%3Dapp"
+        == "postgresql://u:p@reader.example.com:5432/db?options=-c%20search_path%3Dapp&max_idle_connection_lifetime=45"
     )
 
 
@@ -671,7 +675,7 @@ def test_disable_prepared_statements_appends_pgbouncer_to_assembled_writer(monke
 
     assert _apply() is True
     assert os.environ["DATABASE_URL"] == (
-        "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db?pgbouncer=true"
+        "postgresql://litellm:s3cr3t@writer.example.com:5432/litellm_db?pgbouncer=true&max_idle_connection_lifetime=60"
     )
     assert "DIRECT_URL" not in os.environ
 
@@ -685,7 +689,9 @@ def test_disable_prepared_statements_appends_pgbouncer_to_pinned_writer(monkeypa
     monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@db.example.com:5432/litellm_db")
 
     assert _apply() is False
-    assert os.environ["DATABASE_URL"] == "postgresql://u:p@db.example.com:5432/litellm_db?pgbouncer=true"
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?pgbouncer=true&max_idle_connection_lifetime=60"
+    )
 
 
 def test_disable_prepared_statements_respects_a_pinned_pgbouncer_value(monkeypatch):
@@ -694,7 +700,9 @@ def test_disable_prepared_statements_respects_a_pinned_pgbouncer_value(monkeypat
 
     _apply()
 
-    assert os.environ["DATABASE_URL"] == "postgresql://u:p@db.example.com:5432/litellm_db?pgbouncer=false"
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?pgbouncer=false&max_idle_connection_lifetime=60"
+    )
 
 
 def test_disable_prepared_statements_applies_to_direct_url(monkeypatch):
@@ -704,7 +712,9 @@ def test_disable_prepared_statements_applies_to_direct_url(monkeypatch):
 
     _apply()
 
-    assert os.environ["DIRECT_URL"] == "postgresql://u:p@direct.example.com:5432/litellm_db?pgbouncer=true"
+    assert os.environ["DIRECT_URL"] == (
+        "postgresql://u:p@direct.example.com:5432/litellm_db?pgbouncer=true&max_idle_connection_lifetime=60"
+    )
 
 
 def test_reader_inherits_pgbouncer_from_disable_prepared_statements(monkeypatch):
@@ -724,7 +734,9 @@ def test_disable_prepared_statements_off_leaves_urls_alone(monkeypatch):
 
     _apply()
 
-    assert os.environ["DATABASE_URL"] == "postgresql://u:p@db.example.com:5432/litellm_db"
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
+    )
 
 
 def test_disable_prepared_statements_rejects_an_unreadable_value(monkeypatch):
@@ -760,6 +772,7 @@ def test_libpq_verify_full_and_sslrootcert_become_prisma_strict_sslcert(monkeypa
         "sslmode": ["require"],
         "sslcert": ["/certs/rds-bundle.pem"],
         "sslaccept": ["strict"],
+        "max_idle_connection_lifetime": ["60"],
     }
 
 
@@ -768,7 +781,11 @@ def test_libpq_verify_ca_becomes_prisma_strict(monkeypatch):
 
     _apply()
 
-    assert _query(os.environ["DATABASE_URL"]) == {"sslmode": ["require"], "sslaccept": ["strict"]}
+    assert _query(os.environ["DATABASE_URL"]) == {
+        "sslmode": ["require"],
+        "sslaccept": ["strict"],
+        "max_idle_connection_lifetime": ["60"],
+    }
 
 
 def test_sslrootcert_alone_turns_on_strict_verification(monkeypatch):
@@ -784,6 +801,7 @@ def test_sslrootcert_alone_turns_on_strict_verification(monkeypatch):
         "sslmode": ["require"],
         "sslcert": ["/certs/ca.pem"],
         "sslaccept": ["strict"],
+        "max_idle_connection_lifetime": ["60"],
     }
 
 
@@ -800,11 +818,15 @@ def test_pinned_prisma_ssl_params_win_over_libpq_translation(monkeypatch):
         "sslmode": ["require"],
         "sslcert": ["/pinned.pem"],
         "sslaccept": ["accept_invalid_certs"],
+        "max_idle_connection_lifetime": ["60"],
     }
 
 
 def test_prisma_native_ssl_url_is_left_untouched(monkeypatch):
-    url = "postgresql://u:p@db.example.com:5432/litellm_db?sslmode=require&sslcert=/certs/ca.pem&sslaccept=strict"
+    url = (
+        "postgresql://u:p@db.example.com:5432/litellm_db"
+        "?sslmode=require&sslcert=/certs/ca.pem&sslaccept=strict&max_idle_connection_lifetime=60"
+    )
     monkeypatch.setenv("DATABASE_URL", url)
 
     _apply()
@@ -820,4 +842,84 @@ def test_libpq_ssl_translation_covers_direct_url_and_read_replica(monkeypatch):
     _apply()
 
     for env_var in ("DATABASE_URL", "DIRECT_URL", "DATABASE_URL_READ_REPLICA"):
-        assert _query(os.environ[env_var]) == {"sslmode": ["require"], "sslaccept": ["strict"]}, env_var
+        assert _query(os.environ[env_var]) == {
+            "sslmode": ["require"],
+            "sslaccept": ["strict"],
+            "max_idle_connection_lifetime": ["60"],
+        }, env_var
+
+
+def test_default_idle_lifetime_applied_to_pinned_writer_and_direct_url(monkeypatch):
+    monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@db.example.com:5432/litellm_db")
+    monkeypatch.setenv("DIRECT_URL", "postgresql://u:p@direct.example.com:5432/litellm_db")
+
+    assert _apply() is False
+
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
+    )
+    assert os.environ["DIRECT_URL"] == (
+        "postgresql://u:p@direct.example.com:5432/litellm_db?max_idle_connection_lifetime=60"
+    )
+
+
+def test_url_pinned_idle_lifetime_wins_over_default_and_env_knob(monkeypatch):
+    monkeypatch.setenv("DATABASE_MAX_IDLE_CONNECTION_LIFETIME", "45")
+    monkeypatch.setenv(
+        "DATABASE_URL", "postgresql://u:p@db.example.com:5432/litellm_db?max_idle_connection_lifetime=300"
+    )
+
+    _apply()
+
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?max_idle_connection_lifetime=300"
+    )
+
+
+def test_env_knob_overrides_default_idle_lifetime(monkeypatch):
+    monkeypatch.setenv("DATABASE_MAX_IDLE_CONNECTION_LIFETIME", "45")
+    monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@db.example.com:5432/litellm_db")
+    monkeypatch.setenv("DIRECT_URL", "postgresql://u:p@direct.example.com:5432/litellm_db")
+
+    _apply()
+
+    assert os.environ["DATABASE_URL"] == (
+        "postgresql://u:p@db.example.com:5432/litellm_db?max_idle_connection_lifetime=45"
+    )
+    assert os.environ["DIRECT_URL"] == (
+        "postgresql://u:p@direct.example.com:5432/litellm_db?max_idle_connection_lifetime=45"
+    )
+
+
+def test_env_knob_rejects_a_non_integer_value(monkeypatch):
+    monkeypatch.setenv("DATABASE_MAX_IDLE_CONNECTION_LIFETIME", "soon")
+
+    with pytest.raises(ValidationError, match="DATABASE_MAX_IDLE_CONNECTION_LIFETIME"):
+        DatabaseURLSettings.from_env()
+
+
+@pytest.mark.parametrize(("knob", "expected"), [(None, "60"), ("45", "45")])
+def test_reader_inherits_the_writer_idle_lifetime(monkeypatch, knob, expected):
+    if knob is not None:
+        monkeypatch.setenv("DATABASE_MAX_IDLE_CONNECTION_LIFETIME", knob)
+    monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@writer.example.com:5432/db")
+    monkeypatch.setenv("DATABASE_URL_READ_REPLICA", "postgresql://u:p@reader.example.com:5432/db")
+
+    _apply()
+
+    assert os.environ["DATABASE_URL_READ_REPLICA"] == (
+        f"postgresql://u:p@reader.example.com:5432/db?max_idle_connection_lifetime={expected}"
+    )
+
+
+def test_reader_keeps_its_own_pinned_idle_lifetime(monkeypatch):
+    monkeypatch.setenv("DATABASE_URL", "postgresql://u:p@writer.example.com:5432/db")
+    monkeypatch.setenv(
+        "DATABASE_URL_READ_REPLICA", "postgresql://u:p@reader.example.com:5432/db?max_idle_connection_lifetime=120"
+    )
+
+    _apply()
+
+    assert os.environ["DATABASE_URL_READ_REPLICA"] == (
+        "postgresql://u:p@reader.example.com:5432/db?max_idle_connection_lifetime=120"
+    )

tests/test_litellm/proxy/db/test_prisma_client.py

@@ -291,6 +291,56 @@ def test_azure_entra_mint_writes_an_encoded_url_into_the_db_url_env_var(azure_en
     assert os.environ["DATABASE_URL"] == db_url
 
 
+@pytest.mark.parametrize(
+    ("previous_query", "expected_query"),
+    [
+        ("max_idle_connection_lifetime=60", {"max_idle_connection_lifetime": ["60"]}),
+        (
+            "connection_limit=20&pgbouncer=true&max_idle_connection_lifetime=45",
+            {"connection_limit": ["20"], "pgbouncer": ["true"], "max_idle_connection_lifetime": ["45"]},
+        ),
+    ],
+)
+def test_token_refresh_keeps_the_connection_params_of_the_url_it_replaces(
+    azure_env, monkeypatch, previous_query, expected_query
+):
+    old_token = _entra_jwt(60)
+    monkeypatch.setenv(
+        "DATABASE_URL",
+        f"postgresql://litellm%40contoso.onmicrosoft.com:{urllib.parse.quote(old_token, safe='')}"
+        f"@pg.postgres.database.azure.com:5432/litellm_db?{previous_query}",
+    )
+    new_token = _entra_jwt(3600)
+
+    db_url = _azure_wrapper(new_token).get_rds_iam_token()
+
+    assert db_url is not None
+    assert os.environ["DATABASE_URL"] == db_url
+    assert urllib.parse.quote(new_token, safe="") in db_url
+    assert urllib.parse.parse_qs(urllib.parse.urlsplit(db_url).query) == expected_query
+
+
+def test_token_refresh_keeps_the_reader_url_params_separate_from_the_writer(azure_env, monkeypatch):
+    from litellm.proxy.db.token_auth import IAMEndpoint
+
+    monkeypatch.setenv("DATABASE_URL", "postgresql://w:t@pg:5432/litellm_db?max_idle_connection_lifetime=45")
+    monkeypatch.setenv(
+        "DATABASE_URL_READ_REPLICA", "postgresql://r:t@replica:5432/litellm_db?max_idle_connection_lifetime=60"
+    )
+    reader = _azure_wrapper(
+        _entra_jwt(3600),
+        db_url_env_var="DATABASE_URL_READ_REPLICA",
+        iam_endpoint=IAMEndpoint(host="replica", port="5432", user="r", name="litellm_db", schema=None),
+    )
+
+    reader_url = reader.get_rds_iam_token()
+
+    assert reader_url is not None
+    assert reader_url.startswith("postgresql://r:")
+    assert urllib.parse.parse_qs(urllib.parse.urlsplit(reader_url).query) == {"max_idle_connection_lifetime": ["60"]}
+    assert os.environ["DATABASE_URL"].endswith("?max_idle_connection_lifetime=45")
+
+
 def test_azure_entra_refresh_is_scheduled_off_the_jwt_expiry(azure_env):
     """Without reading `exp` this falls back to a fixed 600s interval, which silently
     outlives a token and breaks every reconnect after it lapses (issue #29661)."""

tests/test_litellm/proxy/image_endpoints/test_endpoints.py

@@ -5,12 +5,12 @@ from typing import Any, Dict
 
 import orjson
 import pytest
-from fastapi import FastAPI
+from fastapi import FastAPI, HTTPException
 from fastapi.testclient import TestClient
 from starlette.requests import Request
 from starlette.responses import Response
 
-from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy._types import ProxyException, UserAPIKeyAuth
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.image_endpoints import endpoints
 
@@ -167,3 +167,47 @@ def test_image_edit_multipart_n_that_is_not_a_number_is_left_alone(monkeypatch):
 
     assert response.status_code == 200
     assert captured["n"] == "two"
+
+
+@pytest.mark.asyncio
+async def test_a_model_the_router_cannot_serve_answers_an_openai_typed_error(monkeypatch: pytest.MonkeyPatch):
+    """A bare HTTPException carries no type or param, so the tail used to ship the
+    literal string "None" in both fields."""
+
+    async def fake_add_litellm_data_to_request(**kwargs: object) -> object:
+        return kwargs["data"]
+
+    async def fake_pre_call_hook(*, user_api_key_dict: UserAPIKeyAuth, data: dict[str, object], call_type: str) -> dict[str, object]:
+        return data
+
+    async def fake_post_call_failure_hook(**_: object) -> None:
+        return None
+
+    async def failing_route_request(**_: object) -> None:
+        raise HTTPException(
+            status_code=404, detail={"error": "image_generation: Invalid model name passed in model=dall-e-3"}
+        )
+
+    monkeypatch.setattr("litellm.proxy.proxy_server.add_litellm_data_to_request", fake_add_litellm_data_to_request)
+    monkeypatch.setattr("litellm.proxy.proxy_server.general_settings", {})
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.proxy_config", {})
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.proxy_logging_obj",
+        SimpleNamespace(pre_call_hook=fake_pre_call_hook, post_call_failure_hook=fake_post_call_failure_hook),
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.user_model", None)
+    monkeypatch.setattr("litellm.proxy.proxy_server.version", "test-version")
+    monkeypatch.setattr("litellm.proxy.image_endpoints.endpoints.route_request", failing_route_request)
+
+    body = orjson.dumps({"model": "dall-e-3", "prompt": "a lighthouse at dusk"})
+
+    async def receive() -> dict[str, object]:
+        return {"type": "http.request", "body": body, "more_body": False}
+
+    request = Request({"type": "http", "method": "POST", "path": "/v1/images/generations", "headers": []}, receive)
+
+    with pytest.raises(ProxyException) as raised:
+        await endpoints.image_generation(request=request, fastapi_response=Response(), user_api_key_dict=UserAPIKeyAuth())
+
+    assert (raised.value.type, raised.value.param, raised.value.code) == ("invalid_request_error", None, "404")

tests/test_litellm/proxy/management_endpoints/test_team_callback_endpoints.py

@@ -19,6 +19,7 @@ from litellm.proxy._types import (
     LitellmUserRoles,
     UserAPIKeyAuth,
 )
+from litellm.proxy.common_utils.callback_config_validation import cross_entry_family_error
 from litellm.proxy.management_endpoints.team_callback_endpoints import (
     add_team_callbacks,
     delete_team_callback,
@@ -1443,3 +1444,118 @@ async def test_delete_team_callback_route_accepts_team_ids_containing_slashes():
     assert response.json()["data"]["success_callbacks"] == ["langsmith"]
     written = json.loads(mock_prisma.db.litellm_teamtable.update.await_args.kwargs["data"]["metadata"])
     assert [entry["callback_name"] for entry in written["logging"]] == ["langsmith"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "call_handler",
+    [
+        lambda caller: add_team_callbacks(
+            data=AddTeamCallback(
+                callback_name="langfuse",
+                callback_type="success",
+                callback_vars={"langfuse_public_key": "pk", "langfuse_secret_key": "sk"},
+            ),
+            http_request=Mock(spec=Request),
+            team_id="team-does-not-exist",
+            user_api_key_dict=caller,
+        ),
+        lambda caller: get_team_callbacks(
+            http_request=Mock(spec=Request),
+            team_id="team-does-not-exist",
+            user_api_key_dict=caller,
+        ),
+        lambda caller: delete_team_callback(
+            http_request=Mock(spec=Request),
+            team_id="team-does-not-exist",
+            callback_name="langfuse",
+            user_api_key_dict=caller,
+        ),
+    ],
+    ids=["add", "get", "delete"],
+)
+async def test_unknown_team_is_indistinguishable_from_no_access(call_handler, unauthorized_caller):
+    """An unauthorized caller must not learn whether a team id exists.
+
+    These routes are reachable by any authenticated caller so a team admin can get
+    as far as the access check, so a distinct "does not exist" would turn them into
+    a probe for valid team ids. The unknown-team response has to match the
+    no-access one exactly, status and body.
+    """
+    with patch("litellm.proxy.proxy_server.prisma_client") as mock_client:  # test-quality-ok: the handler imports prisma_client from proxy_server at call time, so there is no seam to inject through
+        mock_client.get_data = AsyncMock(return_value=None)
+        with pytest.raises(HTTPException) as unknown_team:
+            await call_handler(unauthorized_caller)
+
+    with patch("litellm.proxy.proxy_server.prisma_client") as mock_client:  # test-quality-ok: the handler imports prisma_client from proxy_server at call time, so there is no seam to inject through
+        mock_client.get_data = AsyncMock(return_value=_team_row())
+        mock_client.db.litellm_teamtable.update = AsyncMock()
+        with patch(  # test-quality-ok: _verify_team_access calls this module-level helper directly, so there is no seam to inject through
+            "litellm.proxy.management_endpoints.team_endpoints._is_user_org_admin_for_team",
+            new_callable=AsyncMock,
+            return_value=False,
+        ):
+            with pytest.raises(HTTPException) as no_access:
+                await call_handler(unauthorized_caller)
+
+    assert unknown_team.value.status_code == no_access.value.status_code == 403
+    assert unknown_team.value.detail == no_access.value.detail
+    assert "does not exist" not in str(unknown_team.value.detail)
+
+
+@pytest.mark.asyncio
+async def test_proxy_admin_still_told_the_team_is_unknown():
+    """The masking is only for callers who could not have managed the team; a proxy
+    admin keeps the diagnosable error."""
+    admin = UserAPIKeyAuth(user_role=LitellmUserRoles.PROXY_ADMIN, user_id="admin", api_key="sk-admin")
+    with patch("litellm.proxy.proxy_server.prisma_client") as mock_client:  # test-quality-ok: the handler imports prisma_client from proxy_server at call time, so there is no seam to inject through
+        mock_client.get_data = AsyncMock(return_value=None)
+        with pytest.raises(HTTPException) as exc:
+            await get_team_callbacks(
+                http_request=Mock(spec=Request),
+                team_id="team-does-not-exist",
+                user_api_key_dict=admin,
+            )
+
+    assert exc.value.status_code == 404
+    assert "does not exist" in str(exc.value.detail)
+
+
+@pytest.mark.parametrize(
+    "new_vars, stored, rejected",
+    [
+        # the redirect, in every carrier a caller could pick: an entry naming
+        # only a host, pairing with a key pair written on another entry
+        ({"langfuse_host": "http://attacker.invalid"}, [{"langfuse_public_key": "pk", "langfuse_secret_key": "sk"}], True),
+        # the sibling carrier -- langfuse and langfuse_otel are one account
+        ({"langfuse_host": "http://attacker.invalid"}, [{"langfuse_host": "https://us.cloud.langfuse.com", "langfuse_secret_key": "sk"}], True),
+        # a destination variable no integration registry lists
+        ({"dd_agent_host": "attacker.invalid"}, [{"dd_api_key": "k", "dd_site": "us5.datadoghq.com"}], True),
+        # one entry owning its family end to end is the feature
+        ({"langfuse_host": "https://eu.cloud.langfuse.com", "langfuse_public_key": "pk", "langfuse_secret_key": "sk"}, [], False),
+        # a different family alongside an existing one stays fine
+        ({"gcs_bucket_name": "bucket"}, [{"langfuse_public_key": "pk", "langfuse_secret_key": "sk"}], False),
+        ({"langsmith_api_key": "k"}, [{"dd_api_key": "k"}], False),
+        # variables that configure no backend carry nothing to redirect
+        ({"turn_off_message_logging": "true"}, [{"langfuse_secret_key": "sk"}], False),
+        # the same integration registered for a second event: identical values
+        # flatten to the identical dict, so there is nothing to redirect
+        ({"langfuse_host": "https://us.cloud.langfuse.com", "langfuse_public_key": "pk", "langfuse_secret_key": "sk"}, [{"langfuse_host": "https://us.cloud.langfuse.com", "langfuse_public_key": "pk", "langfuse_secret_key": "sk"}], False),
+        # the same credential under its other spelling is the same credential
+        ({"langfuse_secret": "sk"}, [{"langfuse_public_key": "pk", "langfuse_secret_key": "sk"}], False),
+        # a value the family already holds cannot be moved into another of its
+        # variables either; the exporter would address or authenticate with it
+        ({"langfuse_host": "pk"}, [{"langfuse_host": "https://us.cloud.langfuse.com", "langfuse_public_key": "pk"}], True),
+        # the same shape with one value moved is the redirect again
+        ({"langfuse_host": "http://attacker.invalid", "langfuse_public_key": "pk", "langfuse_secret_key": "sk"}, [{"langfuse_host": "https://us.cloud.langfuse.com", "langfuse_public_key": "pk", "langfuse_secret_key": "sk"}], True),
+    ],
+)
+def test_one_entry_owns_a_credential_family(new_vars, stored, rejected):
+    """A team admin must not be able to redirect a credential they cannot read.
+
+    The stored entries are flattened into one dict before a request reads them,
+    so an entry naming only a destination pairs with a key written elsewhere and
+    carries it to that destination.
+    """
+    error = cross_entry_family_error(new_vars, stored)
+    assert (error is not None) is rejected