fix(anthropic): drop the non-spec custom object from /v1/messages tool definitions (#48893)

* fix(anthropic): drop the non-spec custom object from /v1/messages tool definitions

Anthropic and Vertex AI reject a tool carrying a custom object with tools.N.custom.custom: Extra inputs are not permitted. Drop it on the shared Anthropic request path, keep the hoist of a boolean custom.defer_loading to top-level defer_loading, reuse the helper on Bedrock, and stop the chat adapter from merging custom into the tool's JSON schema

* fix(messages): hoist custom.defer_loading only on Claude routes and strip custom on count_tokens

Non-Claude Messages providers such as DeepSeek now get custom dropped without a defer_loading hoist, since they have no tool search. /v1/messages/count_tokens on Anthropic, Azure AI, and Vertex AI drops custom too, and the local token counter no longer counts it as schema.

* fix(anthropic): drop tool custom object only under drop_params and hoist defer_loading for Claude models

Without drop_params the tool is forwarded as sent, so the provider's 400 stands. A boolean custom.defer_loading still becomes the top-level flag for Claude models and on Bedrock Invoke and Mantle. Bedrock count_tokens now normalizes invokeModel tools the way Bedrock Messages sends them.

* fix(bedrock): drop tool custom object on the Mantle count_tokens fallback

Mantle Messages drops custom unconditionally, so its count_tokens call now counts the tools the same way.

* fix(messages): keep custom-tool Messages calls on Python and match Mantle count_tokens to Mantle Messages

* test(vertex): annotate count_tokens tool test locals with Final

* fix(messages): pass tools without custom through untouched so Bedrock fallbacks keep filled tool names

* fix(messages): keep Bedrock's in-place tool cleanup so fallbacks still work, and stop hoisting defer_loading on the OpenAI-like passthrough

---------

Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>

litellm/litellm_core_utils/token_counter.py

@@ -611,7 +611,7 @@ T_ParamSpec: Final = ParamSpec("T_ParamSpec")
 T_Retval = TypeVar("T_Retval")
 _COUNT_OFFLOAD_LIMITER: Final = anyio.lowlevel.RunVar[anyio.CapacityLimiter]("litellm_count_offload_limiter")
 _ANTHROPIC_TOOL_KEYS_MAPPED_BY_BRIDGE: Final = frozenset(
-    ("name", "input_schema", "description", "cache_control", "strict", "type", "eager_input_streaming")
+    ("name", "input_schema", "description", "cache_control", "strict", "type", "eager_input_streaming", "custom")
 )
 
 

litellm/llms/anthropic/common_utils.py

@@ -121,6 +121,54 @@ _DECODED_JSON: Final = TypeAdapter(object)
 _CLAUDE_CODE_USER_AGENT_PREFIXES: Final = ("claude-cli/", "claude-code/")
 
 
+def is_claude_tool_target(custom_llm_provider: str | None, model: str) -> bool:
+    if custom_llm_provider == "anthropic":
+        return True
+    model_lower: Final = model.lower()
+    if custom_llm_provider == "bedrock":
+        return "claude" in model_lower or ("arn:" in model_lower and ":bedrock:" in model_lower)
+    if custom_llm_provider == "azure_ai" or custom_llm_provider == "vertex_ai":
+        return "claude" in model_lower
+    return False
+
+
+def is_claude_model_target(custom_llm_provider: str | None, model: str) -> bool:
+    return "claude" in model.lower() and is_claude_tool_target(custom_llm_provider, model)
+
+
+def _hoistable_defer_loading(tool_dict: Mapping[object, object], custom: object) -> bool | None:
+    if not isinstance(custom, dict) or "defer_loading" in tool_dict:
+        return None
+    defer_loading: Final = OBJECT_MAPPING_ADAPTER.validate_python(custom).get("defer_loading")
+    return defer_loading if isinstance(defer_loading, bool) else None
+
+
+def _tool_without_custom_field(tool: object, hoist_defer_loading: bool, drop_custom: bool) -> object:
+    if not isinstance(tool, dict):
+        return tool
+
+    if "custom" not in tool:
+        return tool
+
+    tool_dict: Final = OBJECT_MAPPING_ADAPTER.validate_python(tool)
+
+    custom: Final = tool_dict["custom"]
+    without_custom: Final = {key: value for key, value in tool_dict.items() if key != "custom"}
+    defer_loading: Final = _hoistable_defer_loading(tool_dict, custom) if hoist_defer_loading else None
+    if defer_loading is None:
+        return without_custom if drop_custom else tool_dict
+
+    unhoisted: Final = {
+        key: value for key, value in OBJECT_MAPPING_ADAPTER.validate_python(custom).items() if key != "defer_loading"
+    }
+    kept_custom: Final = {"custom": unhoisted} if unhoisted and not drop_custom else {}
+    return {**without_custom, "defer_loading": defer_loading, **kept_custom}
+
+
+def tools_without_custom_field(tools: Sequence[object], hoist_defer_loading: bool, drop_custom: bool) -> list[object]:
+    return [_tool_without_custom_field(tool, hoist_defer_loading, drop_custom) for tool in tools]
+
+
 def is_anthropic_messages_url(url: str) -> bool:
     """Check whether a URL addresses Anthropic's Messages API."""
     parsed_url: Final = urlparse(url)

litellm/llms/anthropic/count_tokens/handler.py

@@ -39,6 +39,7 @@ class AnthropicCountTokensHandler(AnthropicCountTokensConfig):
         tools: list[dict[str, JsonValue]] | None = None,
         system: JsonValue = None,
         optional_params: Mapping[str, JsonValue] | None = None,
+        drop_params: bool = False,
     ) -> dict[str, JsonValue]:
         """
         Handle a CountTokens request using httpx.
@@ -69,6 +70,7 @@ class AnthropicCountTokensHandler(AnthropicCountTokensConfig):
                 tools=tools,
                 system=system,
                 optional_params=optional_params,
+                drop_params=drop_params,
             )
 
             verbose_logger.debug("Transformed request: %s", request_body)

litellm/llms/anthropic/count_tokens/token_counter.py

@@ -11,6 +11,7 @@ from pydantic import JsonValue
 import litellm
 from litellm._logging import verbose_logger
 from litellm.exceptions import AuthenticationError
+from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.llms.anthropic.count_tokens.handler import AnthropicCountTokensHandler
 from litellm.llms.anthropic.count_tokens.transformation import (
     AnthropicCountFields,
@@ -131,6 +132,7 @@ class AnthropicTokenCounter(BaseTokenCounter):
                 tools=tools,
                 system=system,
                 optional_params=count_options,
+                drop_params=drops_unsupported_params(litellm_params.get("drop_params")),
             )
 
             if result is not None:

litellm/llms/anthropic/count_tokens/transformation.py

@@ -26,7 +26,11 @@ from litellm.litellm_core_utils.prompt_templates.mid_conversation_system import
     split_leading_system_run,
 )
 from litellm.litellm_core_utils.url_utils import join_api_base
-from litellm.llms.anthropic.common_utils import merge_anthropic_beta_headers
+from litellm.llms.anthropic.common_utils import (
+    is_claude_model_target,
+    merge_anthropic_beta_headers,
+    tools_without_custom_field,
+)
 from litellm.llms.anthropic.wif import resolve_anthropic_base
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionImageObject, ValidUserMessageContentTypes
 from litellm.types.utils import TokenCountMessagesFormat, TokenCountResponse
@@ -372,6 +376,7 @@ class AnthropicCountTokensConfig:
         tools: list[dict[str, JsonValue]] | None = None,
         system: JsonValue = None,
         optional_params: Mapping[str, JsonValue] | None = None,
+        drop_params: bool = False,
     ) -> dict[str, JsonValue]:  # mutable-ok: provider transport requires JSON dictionaries
         """
         Transform request to Anthropic CountTokens format.
@@ -391,7 +396,19 @@ class AnthropicCountTokensConfig:
                     **MappingProxyType(
                         {
                             key: value
-                            for key, value in (("system", counted_system), ("tools", tools))
+                            for key, value in (
+                                ("system", counted_system),
+                                (
+                                    "tools",
+                                    None
+                                    if tools is None
+                                    else tools_without_custom_field(
+                                        tools,
+                                        hoist_defer_loading=is_claude_model_target("anthropic", model),
+                                        drop_custom=drop_params,
+                                    ),
+                                ),
+                            )
                             if value is not None
                         }
                     ),

litellm/llms/anthropic/messages/adapters/transformation.py

@@ -817,6 +817,7 @@ class LiteLLMAnthropicMessagesAdapter:
             "name",
             "input_schema",
             "description",
+            "custom",
             "cache_control",
             "strict",
             "type",

litellm/llms/anthropic/messages/native/transformation.py

@@ -29,13 +29,16 @@ from litellm.types.llms.anthropic_tool_search import get_tool_search_beta_header
 from litellm.types.router import GenericLiteLLMParams
 
 from ...common_utils import (
+    OBJECT_LIST_ADAPTER,
     AnthropicError,
     AnthropicModelInfo,
+    is_claude_model_target,
     merge_anthropic_beta_headers,
     optionally_handle_anthropic_oauth,
     requires_native_compaction_beta,
     strip_advisor_blocks_from_messages,
     strip_encrypted_reasoning_blocks_from_anthropic_messages,
+    tools_without_custom_field,
 )
 from .mid_conversation_system import (
     as_system_content_blocks,
@@ -181,6 +184,16 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
         """
         return False
 
+    def hoists_wrapped_defer_loading(self, model: str) -> bool:
+        return is_claude_model_target(self._resolved_provider, model)
+
+    def tools_for_upstream(self, tools: Sequence[object], model: str, drop_params: object) -> Sequence[object]:
+        return tools_without_custom_field(
+            tools,
+            hoist_defer_loading=self.hoists_wrapped_defer_loading(model),
+            drop_custom=drops_unsupported_params(drop_params),
+        )
+
     @staticmethod
     def _filter_billing_headers_from_system(system_param):
         """
@@ -700,7 +713,8 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
 
         # Auto-strip advisor blocks from history if advisor tool is absent.
         # Prevents Anthropic 400: advisor_tool_result in history requires advisor tool.
-        _tools: Final = anthropic_messages_optional_request_params.get("tools") or []
+        tools: Final = anthropic_messages_optional_request_params.get("tools")
+        _tools: Final = tools or []
         _has_advisor: Final = any(isinstance(t, dict) and t.get("type") == ANTHROPIC_ADVISOR_TOOL_TYPE for t in _tools)
         if not _has_advisor:
             messages = strip_advisor_blocks_from_messages(messages)
@@ -713,7 +727,14 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             model=model,
             **anthropic_messages_optional_request_params,
         )
-        return dict(anthropic_messages_request)
+        if not isinstance(tools, list):
+            return dict(anthropic_messages_request)
+        return {
+            **anthropic_messages_request,
+            "tools": self.tools_for_upstream(
+                OBJECT_LIST_ADAPTER.validate_python(tools), model, litellm_params.drop_params
+            ),
+        }
 
     def transform_anthropic_messages_response(
         self,

litellm/llms/azure_ai/anthropic/count_tokens/handler.py

@@ -13,6 +13,7 @@ from pydantic import JsonValue
 import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.asyncify import asyncify
+from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.llms.anthropic.common_utils import AnthropicError
 from litellm.llms.azure_ai.anthropic.count_tokens.transformation import (
     AzureAIAnthropicCountTokensConfig,
@@ -69,6 +70,7 @@ class AzureAIAnthropicCountTokensHandler(AzureAIAnthropicCountTokensConfig):
                 tools=tools,
                 system=system,
                 optional_params=optional_params,
+                drop_params=drops_unsupported_params((litellm_params or {}).get("drop_params")),
             )
 
             verbose_logger.debug("Transformed request: %s", request_body)

litellm/llms/bedrock/common_utils.py

@@ -15,7 +15,11 @@ from dataclasses import dataclass
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias
 
-from typing_extensions import ReadOnly, TypedDict
+from typing_extensions import (
+    ReadOnly,
+    TypedDict,
+    TypeIs,  # noqa: TID251  # narrows the caller's tools without copying them
+)
 
 if TYPE_CHECKING:
     from botocore.eventstream import EventStreamMessage
@@ -573,6 +577,26 @@ def strip_unsupported_bedrock_invoke_output_config_keys(
         request_body["output_config"] = {"format": preserved_format}  # rebind-ok: out-param
 
 
+_TOOL_ITEMS_ADAPTER: Final = TypeAdapter(tuple[object, ...])
+
+
+def _is_str_keyed_dict(
+    value: object,
+) -> TypeIs[dict[str, object]]:  # guard-ok: JSON tool objects  # mutable-ok: fallbacks reuse tools edited in place
+    return isinstance(value, dict)
+
+
+def _normalize_custom_field_on_tool(tool: object) -> None:
+    if not _is_str_keyed_dict(tool):
+        return
+    custom: Final = tool.pop("custom", None)
+    if not _is_str_keyed_dict(custom) or "defer_loading" in tool:
+        return
+    deferred: Final = custom.get("defer_loading")
+    if isinstance(deferred, bool):
+        tool["defer_loading"] = deferred  # rebind-ok: a fallback reuses the caller's tool, as on main
+
+
 def normalize_custom_field_on_tools(request_body: dict) -> None:
     """
     Drop the ``custom`` field from each tool, first hoisting a boolean
@@ -590,15 +614,8 @@ def normalize_custom_field_on_tools(request_body: dict) -> None:
     tools: Final = request_body.get("tools")
     if not tools or not isinstance(tools, list):
         return
-    for tool in tools:
-        if not isinstance(tool, dict):
-            continue
-        custom: dict[str, object] | None = tool.pop("custom", None)
-        if not isinstance(custom, dict) or "defer_loading" in tool:
-            continue
-        deferred: object = custom.get("defer_loading")
-        if isinstance(deferred, bool):
-            tool["defer_loading"] = deferred
+    for tool in _TOOL_ITEMS_ADAPTER.validate_python(tools):
+        _normalize_custom_field_on_tool(tool)
 
 
 _TOOL_DICTS_ADAPTER: Final = TypeAdapter(tuple[Mapping[str, object], ...])

litellm/llms/bedrock/count_tokens/mantle_handler.py

@@ -8,6 +8,7 @@ from typing_extensions import NotRequired, ReadOnly, TypedDict
 import litellm
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.url_utils import join_api_base
+from litellm.llms.anthropic.common_utils import tools_without_custom_field
 from litellm.llms.anthropic.count_tokens.transformation import COUNT_TOKEN_OPTION_NAMES, AnthropicCountTokensConfig
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, run_aws_signing
 from litellm.llms.bedrock.common_utils import BedrockError, build_mantle_messages_url
@@ -26,6 +27,7 @@ class MantleCountTokensRequest(TypedDict):
 _COUNT_REQUEST: Final = TypeAdapter(MantleCountTokensRequest)
 _COUNT_RESPONSE: Final = TypeAdapter(dict[str, JsonValue])
 _COUNT_OPTIONS: Final[TypeAdapter[Mapping[str, JsonValue]]] = TypeAdapter(Mapping[str, JsonValue])
+_TOOLS: Final = TypeAdapter(list[dict[str, JsonValue]])
 
 
 class BedrockMantleCountTokensHandler(BaseAWSLLM):
@@ -59,10 +61,15 @@ class BedrockMantleCountTokensHandler(BaseAWSLLM):
                 self.get_aws_region_name, optional_params=litellm_params, model=resolved_model, model_id=None
             )
             endpoint_url: Final = await run_aws_signing(self.get_mantle_count_tokens_endpoint, aws_region_name)
+            tools: Final = request.get("tools")
             body: Final = self._anthropic_config.transform_request_to_count_tokens(
                 model=resolved_model,
                 messages=request["messages"],
-                tools=request.get("tools"),
+                tools=None
+                if tools is None
+                else _TOOLS.validate_python(
+                    tools_without_custom_field(tools, hoist_defer_loading=True, drop_custom=True)
+                ),
                 system=request.get("system"),
                 optional_params=_COUNT_OPTIONS.validate_python(
                     {name: value for name, value in request_data.items() if name in COUNT_TOKEN_OPTION_NAMES}

litellm/llms/bedrock/count_tokens/transformation.py

@@ -12,6 +12,7 @@ from typing import Final, Literal
 from pydantic import JsonValue
 
 from litellm.litellm_core_utils.url_utils import join_api_base
+from litellm.llms.anthropic.common_utils import tools_without_custom_field
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import get_bedrock_base_model
 
@@ -194,7 +195,15 @@ class BedrockCountTokensConfig(BaseAWSLLM):
 
         # For InvokeModel, we need to provide the raw body that would be sent to the model
         # Remove the 'model' field from the body as it's not part of the model input
-        body_data: Final = {k: v for k, v in request_data.items() if k != "model"}
+        tools: Final = request_data.get("tools")
+        body_data: Final = {
+            **{k: v for k, v in request_data.items() if k != "model"},
+            **(
+                {"tools": tools_without_custom_field(tools, hoist_defer_loading=True, drop_custom=True)}
+                if isinstance(tools, list)
+                else {}
+            ),
+        }
 
         if "messages" in body_data:
             # Bedrock validates the body against the model's InvokeModel schema;

litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude_transformation.py

@@ -102,6 +102,9 @@ class AmazonAnthropicClaudeMessagesConfig(
     def should_strip_billing_metadata(self) -> bool:
         return True
 
+    def tools_for_upstream(self, tools: Sequence[object], model: str, drop_params: object) -> Sequence[object]:
+        return tools
+
     def drops_dangerous_tool_use_beta_without_safeguards(self) -> bool:
         return True
 

litellm/llms/openai_like/messages/transformation.py

@@ -37,6 +37,9 @@ class OpenAILikeAnthropicMessagesConfig(AnthropicMessagesConfig):
     def is_native_messages_route(self) -> bool:
         return False
 
+    def hoists_wrapped_defer_loading(self, model: str) -> bool:
+        return False
+
     def validate_anthropic_messages_environment(
         self,
         headers: dict[str, str],

litellm/llms/vertex_ai/vertex_ai_partner_models/count_tokens/handler.py

@@ -6,9 +6,12 @@ Unlike Gemini models which use Google's token counting API, partner models use
 their respective publisher-specific count-tokens endpoints.
 """
 
+from collections.abc import Mapping
 from typing import Any, Final
 
+from litellm.litellm_core_utils.core_helpers import drops_unsupported_params
 from litellm.litellm_core_utils.url_utils import join_api_base
+from litellm.llms.anthropic.common_utils import OBJECT_LIST_ADAPTER, is_claude_model_target, tools_without_custom_field
 from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
 from litellm.llms.vertex_ai.common_utils import get_vertex_base_url
 from litellm.llms.vertex_ai.vertex_llm_base import VertexBase
@@ -20,6 +23,22 @@ _COUNT_TOKENS_PUBLISHERS: Final = (
 )
 
 
+def count_tokens_request_body(
+    request_data: Mapping[str, object], model: str, drop_params: object
+) -> Mapping[str, object]:
+    tools: Final = request_data.get("tools")
+    if not isinstance(tools, list):
+        return request_data
+    return {
+        **request_data,
+        "tools": tools_without_custom_field(
+            OBJECT_LIST_ADAPTER.validate_python(tools),
+            hoist_defer_loading=is_claude_model_target("vertex_ai", model),
+            drop_custom=drops_unsupported_params(drop_params),
+        ),
+    }
+
+
 def count_tokens_publisher(model: str) -> str | None:
     return next((publisher for marker, publisher in _COUNT_TOKENS_PUBLISHERS if marker in model), None)
 
@@ -169,7 +188,7 @@ class VertexAIPartnerModelsTokenCounter(VertexBase):
         response: Final = await async_client.post(
             endpoint_url,
             headers=headers,
-            json=request_data,
+            json=dict(count_tokens_request_body(request_data, model, litellm_params.get("drop_params"))),
             timeout=30.0,
         )
 

litellm/main.py

@@ -234,6 +234,7 @@ from .litellm_core_utils.streaming_chunk_builder_utils import (
     ChunkProcessor as ChunkProcessor,  # noqa: PLC0414  # public re-export
 )
 from .llms.anthropic.chat import AnthropicChatCompletion
+from .llms.anthropic.common_utils import is_claude_tool_target
 from .llms.azure.audio_transcriptions import AzureAudioTranscription
 from .llms.azure.azure import AzureChatCompletion, check_dynamic_azure_params
 from .llms.azure.chat.o_series_handler import AzureOpenAIO1ChatCompletion
@@ -1135,17 +1136,6 @@ from litellm.responses.bridge_check import (  # noqa: E402  # re-export at the o
 _ANTHROPIC_ONLY_TOOL_KEYS: Final = frozenset({"input_examples", "eager_input_streaming"})
 
 
-def _is_claude_tool_target(custom_llm_provider: str | None, model: str) -> bool:
-    if custom_llm_provider == "anthropic":
-        return True
-    model_lower: Final = model.lower()
-    if custom_llm_provider == "bedrock":
-        return "claude" in model_lower or ("arn:" in model_lower and ":bedrock:" in model_lower)
-    if custom_llm_provider == "azure_ai" or custom_llm_provider == "vertex_ai":
-        return "claude" in model_lower
-    return False
-
-
 def _history_without_unanswered_tool_search_calls(
     messages: Sequence[object], custom_llm_provider: str | None, provider_config: BaseConfig
 ) -> tuple[object, ...] | None:
@@ -5752,7 +5742,7 @@ def completion(
             }
         )
 
-        if not _is_claude_tool_target(custom_llm_provider=custom_llm_provider, model=model):
+        if not is_claude_tool_target(custom_llm_provider=custom_llm_provider, model=model):
             tools = _drop_anthropic_only_tool_keys(tools=tools)
 
         if provider_specific_header is not None:

litellm/messages/dispatch.py

@@ -111,6 +111,7 @@ _FEATURES: Final[detectors.Detectors] = MappingProxyType(
         Feature.CACHE_CONTROL_INJECTION: detectors.cache_control_injection,
         Feature.STREAM_COST_INJECTION: route_detectors.stream_cost_injection,
         Feature.THINKING_SIGNATURE_RETRY: route_detectors.thinking_signature_retry,
+        Feature.TOOL_CUSTOM_FIELD: route_detectors.tool_custom_field,
     }
 )
 

litellm/rust_bridge/catalog.py

@@ -96,6 +96,8 @@ class Feature(str, Enum):
     """litellm.include_cost_in_streaming_usage, where Python adds cost to streamed usage."""
     THINKING_SIGNATURE_RETRY = "thinking_signature_retry"
     """Replayed thinking blocks, which Python strips and retries once if Anthropic rejects their signature."""
+    TOOL_CUSTOM_FIELD = "tool_custom_field"
+    """Tools carrying a non-spec custom object, which Python drops or hoists for the provider."""
     CLIENT_HEADERS = "client_headers"
     """Proxy-forwarded headers that Python merges into extra_headers."""
     DEPLOYMENT_HOOK = "deployment_hook"
@@ -176,6 +178,7 @@ MESSAGES_PYTHON_FEATURES: Final = COMMON_PYTHON_FEATURES | {
     Feature.CACHE_CONTROL_INJECTION,
     Feature.STREAM_COST_INJECTION,
     Feature.THINKING_SIGNATURE_RETRY,
+    Feature.TOOL_CUSTOM_FIELD,
 }
 
 RESPONSES_PYTHON_FEATURES: Final = COMMON_PYTHON_FEATURES | {

litellm/rust_bridge/messages/detectors.py

@@ -43,6 +43,11 @@ def thinking_signature_retry(request: NativeCall) -> bool:
     return any(_replays_thinking(message) for message in optional_sequence(request.resolved.get("messages")) or ())
 
 
+def tool_custom_field(request: NativeCall) -> bool:
+    tools: Final = optional_sequence(request.resolved.get("tools")) or ()
+    return any("custom" in (optional_mapping(tool) or {}) for tool in tools)
+
+
 def stream_cost_injection(request: NativeCall) -> bool:
     """Python adds response cost to streaming usage when the global setting is truthy."""
     return streaming(request) and bool(litellm.include_cost_in_streaming_usage)

tests/test_litellm_rust/messages/test_dispatch.py

@@ -473,6 +473,12 @@ ADVISOR_TOOL: Final = {"type": ANTHROPIC_ADVISOR_TOOL_TYPE, "name": "advisor", "
             {"messages": [{"role": "assistant", "content": [{"type": "redacted_thinking", "data": "x"}]}]},
             {},
         ),
+        (
+            Feature.TOOL_CUSTOM_FIELD,
+            "anthropic/claude-sonnet-4-5",
+            {"tools": [{"name": "a", "input_schema": {"type": "object"}, "custom": {"eager_input_streaming": True}}]},
+            {},
+        ),
         (
             Feature.STREAM_COST_INJECTION,
             "anthropic/claude-sonnet-4-5",

tests/unit/litellm_core_utils/test_token_counter.py

@@ -700,6 +700,7 @@ def test_token_counter_counts_anthropic_tool_choices_like_openai_choices(
         _anthropic_tool(_WEATHER_FUNCTION),
         {**_anthropic_tool(_WEATHER_FUNCTION), "strict": True},
         {**_anthropic_tool(_WEATHER_FUNCTION), "defer_loading": True, "input_examples": [{"location": "Paris"}]},
+        {**_anthropic_tool(_WEATHER_FUNCTION), "custom": {"eager_input_streaming": True}},
         _anthropic_tool(_WEB_SEARCH_FUNCTION),
     ],
 )

tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py

@@ -4238,6 +4238,21 @@ def test_translate_anthropic_tools_to_openai_preserves_parameters_type():
     assert new_tools[0]["type"] == "function"
 
 
+def test_translate_anthropic_custom_field_does_not_change_openai_parameters():
+    input_schema: Final = {"type": "object", "properties": {"city": {"type": "string"}}}
+    tools: Final = [
+        {
+            "name": "get_weather",
+            "input_schema": input_schema,
+            "custom": {"eager_input_streaming": True},
+        }
+    ]
+
+    new_tools, _ = LiteLLMAnthropicMessagesAdapter().translate_anthropic_tools_to_openai(tools=tools)
+
+    assert new_tools[0]["function"]["parameters"] == input_schema
+
+
 def test_translate_anthropic_tools_to_openai_maps_strict_onto_function_not_parameters():
     """A tool-level `strict` lands on the OpenAI function, leaving the caller's `input_schema` untouched."""
     adapter = LiteLLMAnthropicMessagesAdapter()

tests/unit/llms/anthropic/messages/native/test_anthropic_messages_transformation.py

@@ -14,9 +14,13 @@ from litellm.llms.anthropic.messages.native.transformation import (
     AnthropicMessagesConfig,
 )
 from litellm.llms.anthropic.wif import get_anthropic_wif_token
+from litellm.llms.deepseek.messages.transformation import DeepSeekAnthropicMessagesConfig
 from litellm.llms.minimax.messages.transformation import MinimaxMessagesConfig
 from litellm.llms.openai_like.messages.transformation import OpenAILikeAnthropicMessagesConfig
 from litellm.llms.tencent.messages.transformation import TencentAnthropicMessagesConfig
+from litellm.llms.vertex_ai.vertex_ai_partner_models.anthropic.experimental_pass_through.transformation import (
+    VertexAIPartnerModelsAnthropicMessagesConfig,
+)
 from litellm.types.llms.anthropic import AnthropicFinishReason
 from litellm.types.router import GenericLiteLLMParams
 from tests.unit.llms.anthropic.test_anthropic_wif import (
@@ -79,6 +83,84 @@ def test_native_response_defaults_stop_sequence_without_changing_other_fields(
     }
 
 
+_EMPTY_SCHEMA: Final[dict[str, object]] = {"type": "object", "properties": {}}
+
+
+@pytest.mark.parametrize(
+    "config_class",
+    [AnthropicMessagesConfig, VertexAIPartnerModelsAnthropicMessagesConfig],
+)
+@pytest.mark.parametrize(
+    ("drop_params", "expected_eager_tool"),
+    [
+        (True, {"name": "A", "input_schema": _EMPTY_SCHEMA}),
+        (False, {"name": "A", "input_schema": _EMPTY_SCHEMA, "custom": {"eager_input_streaming": True}}),
+    ],
+)
+def test_messages_request_hoists_wrapped_defer_loading_and_drops_custom_only_under_drop_params(
+    config_class: type[AnthropicMessagesConfig], drop_params: bool, expected_eager_tool: dict[str, object]
+) -> None:
+    tools: Final[list[dict[str, object]]] = [
+        {"name": "A", "input_schema": _EMPTY_SCHEMA, "custom": {"eager_input_streaming": True}},
+        {"name": "B", "input_schema": _EMPTY_SCHEMA, "custom": {"defer_loading": True}},
+        {"name": "C", "input_schema": _EMPTY_SCHEMA, "custom": {"defer_loading": True, "eager_input_streaming": True}},
+        {"name": "D", "input_schema": _EMPTY_SCHEMA},
+    ]
+    optional_params: Final[dict[str, object]] = {"max_tokens": 50, "tools": tools}
+
+    request: Final = config_class().transform_anthropic_messages_request(
+        model="claude-haiku-5-5",
+        messages=[{"role": "user", "content": "Hi"}],
+        anthropic_messages_optional_request_params=optional_params,
+        litellm_params=GenericLiteLLMParams(drop_params=drop_params),
+        headers={},
+    )
+
+    assert request["tools"] == [
+        expected_eager_tool,
+        {"name": "B", "input_schema": _EMPTY_SCHEMA, "defer_loading": True},
+        {
+            "name": "C",
+            "input_schema": _EMPTY_SCHEMA,
+            "defer_loading": True,
+            **({} if drop_params else {"custom": {"eager_input_streaming": True}}),
+        },
+        {"name": "D", "input_schema": _EMPTY_SCHEMA},
+    ]
+    assert tools[0]["custom"] == {"eager_input_streaming": True}
+    assert optional_params["tools"] is tools
+
+
+@pytest.mark.parametrize(
+    ("config", "model"),
+    [
+        (DeepSeekAnthropicMessagesConfig(), "deepseek-chat"),
+        (MinimaxMessagesConfig(), "MiniMax-M2"),
+        (AnthropicMessagesConfig(), "deepseek-chat"),
+        (OpenAILikeAnthropicMessagesConfig(), "anthropic/claude-haiku-4.5"),
+    ],
+)
+@pytest.mark.parametrize("drop_params", [True, False])
+def test_messages_request_never_defers_tools_off_claude(
+    config: AnthropicMessagesConfig, model: str, drop_params: bool
+) -> None:
+    wrapped_tool: Final[dict[str, object]] = {
+        "name": "A",
+        "input_schema": _EMPTY_SCHEMA,
+        "custom": {"defer_loading": True},
+    }
+
+    request: Final = config.transform_anthropic_messages_request(
+        model=model,
+        messages=[{"role": "user", "content": "Hi"}],
+        anthropic_messages_optional_request_params={"max_tokens": 50, "tools": [wrapped_tool]},
+        litellm_params=GenericLiteLLMParams(drop_params=drop_params),
+        headers={},
+    )
+
+    assert request["tools"] == [{"name": "A", "input_schema": _EMPTY_SCHEMA} if drop_params else wrapped_tool]
+
+
 def test_workload_identity_allowed_for_anthropic() -> None:
     assert AnthropicMessagesConfig()._allows_workload_identity is True
 

tests/unit/llms/anthropic/test_anthropic_common_utils.py

@@ -25,6 +25,7 @@ import pytest
 import respx
 
 import litellm
+from litellm.llms.anthropic.common_utils import is_claude_tool_target
 from litellm.llms.custom_httpx.http_handler import HTTPHandler
 
 sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../../..")))
@@ -4522,3 +4523,22 @@ def test_shared_legacy_thinking_translation_drops_a_non_numeric_budget_under_dro
     )
 
     assert optional_params == {"thinking": {"type": "adaptive"}}
+
+
+@pytest.mark.parametrize(
+    ("custom_llm_provider", "model", "expected"),
+    [
+        ("anthropic", "claude-sonnet-5", True),
+        ("bedrock", "us.anthropic.claude-sonnet-5-20260501-v1:0", True),
+        ("bedrock", "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123", True),
+        ("bedrock", "us.amazon.nova-2-lite-v1:0", False),
+        ("vertex_ai", "claude-sonnet-5", True),
+        ("vertex_ai", "gemini-3.8-flash", False),
+        ("azure_ai", "claude-sonnet-4-6", True),
+        ("azure_ai", "gpt-5.6", False),
+        ("openai", "gpt-5.6", False),
+        ("gemini", "gemini-3.8-flash", False),
+    ],
+)
+def test_is_claude_tool_target(custom_llm_provider: str, model: str, expected: bool):
+    assert is_claude_tool_target(custom_llm_provider=custom_llm_provider, model=model) is expected

tests/unit/llms/anthropic/test_anthropic_count_tokens_transformation.py

@@ -171,6 +171,54 @@ def test_transform_includes_tools():
     assert result["tools"] == tools
 
 
+_PATH_SCHEMA: Final[dict[str, JsonValue]] = {"type": "object", "properties": {"path": {"type": "string"}}}
+
+
+@pytest.mark.parametrize(
+    ("model", "drop_params", "expected_tools"),
+    [
+        (
+            "claude-haiku-5-5",
+            True,
+            [
+                {"name": "read_file", "input_schema": _PATH_SCHEMA},
+                {"name": "list_files", "input_schema": _PATH_SCHEMA, "defer_loading": True},
+            ],
+        ),
+        (
+            "claude-haiku-5-5",
+            False,
+            [
+                {"name": "read_file", "input_schema": _PATH_SCHEMA, "custom": {"eager_input_streaming": True}},
+                {"name": "list_files", "input_schema": _PATH_SCHEMA, "defer_loading": True},
+            ],
+        ),
+        (
+            "deepseek-chat",
+            True,
+            [
+                {"name": "read_file", "input_schema": _PATH_SCHEMA},
+                {"name": "list_files", "input_schema": _PATH_SCHEMA},
+            ],
+        ),
+    ],
+)
+def test_transform_counts_tools_the_way_messages_sends_them(
+    model: str, drop_params: bool, expected_tools: list[dict[str, JsonValue]]
+):
+    result: Final = AnthropicCountTokensConfig().transform_request_to_count_tokens(
+        model=model,
+        messages=[{"role": "user", "content": "Hello"}],
+        tools=[
+            {"name": "read_file", "input_schema": _PATH_SCHEMA, "custom": {"eager_input_streaming": True}},
+            {"name": "list_files", "input_schema": _PATH_SCHEMA, "custom": {"defer_loading": True}},
+        ],
+        drop_params=drop_params,
+    )
+
+    assert result["tools"] == expected_tools
+
+
 def test_transform_includes_system_and_tools():
     """Test that both system and tools are included together."""
     config = AnthropicCountTokensConfig()

tests/unit/llms/bedrock/count_tokens/test_bedrock_count_tokens_transformation.py

@@ -73,6 +73,27 @@ def test_transform_to_invoke_model_format_base64_encodes_body():
     assert body["max_tokens"] == DEFAULT_ANTHROPIC_INVOKE_MODEL_MAX_TOKENS
 
 
+def test_transform_to_invoke_model_format_counts_tools_the_way_bedrock_messages_sends_them():
+    schema = {"type": "object", "properties": {}}
+
+    result = BedrockCountTokensConfig().transform_anthropic_to_bedrock_count_tokens(
+        {
+            "model": "anthropic.claude-haiku-5-5",
+            "messages": [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}],
+            "tools": [
+                {"name": "read", "input_schema": schema, "custom": {"defer_loading": True}},
+                {"name": "write", "input_schema": schema, "custom": {"eager_input_streaming": True}},
+            ],
+        }
+    )
+
+    body = json.loads(base64.b64decode(result["input"]["invokeModel"]["body"]))
+    assert body["tools"] == [
+        {"name": "read", "input_schema": schema, "defer_loading": True},
+        {"name": "write", "input_schema": schema},
+    ]
+
+
 def test_transform_to_invoke_model_format_raw_body_unchanged():
     """Non-messages bodies (e.g. Titan inputText) must not get Anthropic fields."""
     config = BedrockCountTokensConfig()

tests/unit/llms/bedrock/count_tokens/test_bedrock_mantle_count_tokens_handler.py

@@ -74,6 +74,34 @@ async def test_counts_on_mantle_with_the_base_model_and_the_deployment_credentia
     }
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "resolved_model",
+    ["anthropic.claude-opus-4-8", "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123"],
+)
+async def test_counts_tools_the_way_mantle_messages_sends_them(resolved_model: str) -> None:
+    mantle: Final = _MantleEndpoint(200, {"input_tokens": 11})
+    schema: Final = {"type": "object", "properties": {}}
+
+    await BedrockMantleCountTokensHandler().handle_count_tokens_request(
+        request_data={
+            **REQUEST,
+            "tools": [
+                {"name": "eager", "input_schema": schema, "custom": {"eager_input_streaming": True}},
+                {"name": "deferred", "input_schema": schema, "custom": {"defer_loading": True}},
+            ],
+        },
+        litellm_params=dict(LITELLM_PARAMS),
+        resolved_model=resolved_model,
+        client=_client(mantle),
+    )
+
+    assert json.loads(mantle.only_request().content)["tools"] == [
+        {"name": "eager", "input_schema": schema},
+        {"name": "deferred", "input_schema": schema, "defer_loading": True},
+    ]
+
+
 @pytest.mark.asyncio
 async def test_bedrock_mantle_api_base_env_names_the_host(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("BEDROCK_MANTLE_API_BASE", "https://vpce-abc.bedrock-mantle.us-east-1.vpce.api.aws")

tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude_transformation.py

@@ -4198,3 +4198,110 @@ def test_bedrock_messages_without_guardrail_config_sends_no_guardrail_headers()
         headers={}, model="anthropic.claude-sonnet-5-v1:0", messages=[], optional_params={}, litellm_params={}
     )
     assert not any(name.startswith("X-Amzn-Bedrock-Guardrail") for name in headers)
+
+
+@dataclass(slots=True)
+class BedrockDownAnthropicUp:
+    anthropic_bodies: tuple[dict[str, object], ...] = ()
+
+    def __call__(self, request: httpx.Request) -> httpx.Response:
+        if request.url.host != "api.anthropic.com":
+            return httpx.Response(500, json={"message": "Bedrock is unavailable"})
+        self.anthropic_bodies = (*self.anthropic_bodies, REASONING_BODY_ADAPTER.validate_json(request.content))
+        return httpx.Response(200, json=_ANTHROPIC_MESSAGE)
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize(
+    ("unnamed_tool_extras", "sent_extras"),
+    [
+        ({}, {}),
+        ({"custom": {"eager_input_streaming": True}}, {}),
+        ({"custom": {"defer_loading": False}}, {"defer_loading": False}),
+    ],
+)
+async def test_bedrock_messages_fallback_to_anthropic_sends_the_tool_names_bedrock_filled_in(
+    respx_mock: respx.Router,
+    monkeypatch: pytest.MonkeyPatch,
+    unnamed_tool_extras: dict[str, object],
+    sent_extras: dict[str, object],
+) -> None:
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    upstream: Final = BedrockDownAnthropicUp()
+    respx_mock.post(url__regex=r"https://.*").mock(side_effect=upstream)
+    router: Final = litellm.Router(
+        model_list=[
+            {
+                "model_name": "bedrock-claude",
+                "litellm_params": {
+                    "model": "bedrock/invoke/us.anthropic.claude-opus-4-8",
+                    "aws_access_key_id": "test-access-key",
+                    "aws_secret_access_key": "test-secret-key",
+                    "aws_region_name": "us-east-1",
+                },
+            },
+            {"model_name": "anthropic-claude", "litellm_params": {"model": "anthropic/claude-opus-4-8", "api_key": "test"}},
+        ],
+        fallbacks=[{"bedrock-claude": ["anthropic-claude"]}],
+        num_retries=0,
+    )
+    schema: Final[Mapping[str, object]] = {"type": "object", "properties": {}}
+
+    await router.aanthropic_messages(
+        model="bedrock-claude",
+        messages=[{"role": "user", "content": "What is the weather?"}],
+        max_tokens=100,
+        tools=[{"input_schema": schema, **unnamed_tool_extras}, {"name": "get_weather", "input_schema": schema}],
+    )
+
+    (body,) = upstream.anthropic_bodies
+    assert body["tools"] == [
+        {"name": "litellm_unnamed_tool_0", "input_schema": schema, **sent_extras},
+        {"name": "get_weather", "input_schema": schema},
+    ]
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize(
+    ("custom", "sent_extras"),
+    [({"eager_input_streaming": True}, {}), ({"defer_loading": False}, {"defer_loading": False})],
+)
+async def test_bedrock_chat_fallback_to_anthropic_sends_the_tool_bedrock_cleaned(
+    respx_mock: respx.Router,
+    monkeypatch: pytest.MonkeyPatch,
+    custom: dict[str, object],
+    sent_extras: dict[str, object],
+) -> None:
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    upstream: Final = BedrockDownAnthropicUp()
+    respx_mock.post(url__regex=r"https://.*").mock(side_effect=upstream)
+    router: Final = litellm.Router(
+        model_list=[
+            {
+                "model_name": "bedrock-claude",
+                "litellm_params": {
+                    "model": "bedrock/invoke/us.anthropic.claude-opus-4-8",
+                    "aws_access_key_id": "test-access-key",
+                    "aws_secret_access_key": "test-secret-key",
+                    "aws_region_name": "us-east-1",
+                },
+            },
+            {
+                "model_name": "anthropic-claude",
+                "litellm_params": {"model": "anthropic/claude-opus-4-8", "api_key": "test", "drop_params": True},
+            },
+        ],
+        fallbacks=[{"bedrock-claude": ["anthropic-claude"]}],
+        num_retries=0,
+    )
+    schema: Final[Mapping[str, object]] = {"type": "object", "properties": {}}
+
+    await router.acompletion(
+        model="bedrock-claude",
+        messages=[{"role": "user", "content": "What is the weather?"}],
+        max_tokens=100,
+        tools=[{"name": "get_weather", "input_schema": schema, "custom": custom}],
+    )
+
+    (body,) = upstream.anthropic_bodies
+    assert body["tools"] == [{"name": "get_weather", "input_schema": schema, **sent_extras}]

tests/unit/llms/bedrock_mantle/test_bedrock_mantle_messages_transformation.py

@@ -224,6 +224,20 @@ class TestRequestBody:
         )
         assert "stream" not in body
 
+    def test_body_hoists_wrapped_defer_loading_and_drops_custom(self):
+        schema: Final = {"type": "object", "properties": {}}
+        body = BedrockMantleAnthropicMessagesConfig().transform_anthropic_messages_request(
+            model="anthropic.claude-sonnet-5",
+            messages=[{"role": "user", "content": "ping"}],
+            anthropic_messages_optional_request_params={
+                "max_tokens": 8,
+                "tools": [{"name": "Read", "input_schema": schema, "custom": {"defer_loading": True}}],
+            },
+            litellm_params=GenericLiteLLMParams(),
+            headers={},
+        )
+        assert body["tools"] == [{"name": "Read", "input_schema": schema, "defer_loading": True}]
+
 
 class TestAuth:
     def test_bearer_from_api_key_skips_aws_credentials(self):

tests/unit/llms/vertex_ai/vertex_ai_partner_models/count_tokens/test_count_tokens_location.py

@@ -4,10 +4,13 @@ Tests for Vertex AI partner models count_tokens location resolution.
 Ref: https://github.com/BerriAI/litellm/issues/23872
 """
 
+from typing import Final
+
 import pytest
 
 from litellm.llms.vertex_ai.vertex_ai_partner_models.count_tokens.handler import (
     VertexAIPartnerModelsTokenCounter,
+    count_tokens_request_body,
 )
 
 
@@ -294,3 +297,21 @@ class TestCountTokensVersionSuffixStripping:
 
         # The model name sent to the API must NOT have the @default suffix
         assert captured_json["model"] == "claude-sonnet-4-6"
+
+
+@pytest.mark.parametrize("drop_params", [True, False])
+def test_count_tokens_counts_tools_the_way_vertex_messages_sends_them(drop_params):
+    schema: Final = {"type": "object", "properties": {}}
+    eager_tool: Final = {"name": "write", "input_schema": schema, "custom": {"eager_input_streaming": True}}
+    request_data: Final = {
+        "messages": [{"role": "user", "content": "hi"}],
+        "tools": [{"name": "read", "input_schema": schema, "custom": {"defer_loading": True}}, eager_tool],
+    }
+
+    body: Final = count_tokens_request_body(request_data, "claude-sonnet-4-6", drop_params)
+
+    assert body["messages"] == request_data["messages"]
+    assert body["tools"] == [
+        {"name": "read", "input_schema": schema, "defer_loading": True},
+        {"name": "write", "input_schema": schema} if drop_params else eager_tool,
+    ]

tests/unit/rust_bridge/test_detectors.py

@@ -189,6 +189,9 @@ def _wrap_nested_level(value: object, depth: int) -> object:
             {"messages": [{"role": "user", "content": [{"type": "text"}]}]},
             False,
         ),
+        (messages_detectors.tool_custom_field, {"tools": [{"name": "a"}, {"name": "b", "custom": {}}]}, True),
+        (messages_detectors.tool_custom_field, {"tools": [{"name": "a", "defer_loading": True}]}, False),
+        (messages_detectors.tool_custom_field, {}, False),
         (detectors.drop_params, {"drop_params": True}, True),
         (detectors.drop_params, {"drop_params": False}, False),
         (detectors.drop_params, {"drop_params": " True "}, True),

tests/unit/test_main.py

@@ -503,25 +503,6 @@ def test_bedrock_latency_optimized_inference():
         assert json_data["performanceConfig"]["latency"] == "optimized"
 
 
-@pytest.mark.parametrize(
-    ("custom_llm_provider", "model", "expected"),
-    [
-        ("anthropic", "claude-sonnet-5", True),
-        ("bedrock", "us.anthropic.claude-sonnet-5-20260501-v1:0", True),
-        ("bedrock", "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123", True),
-        ("bedrock", "us.amazon.nova-2-lite-v1:0", False),
-        ("vertex_ai", "claude-sonnet-5", True),
-        ("vertex_ai", "gemini-3.8-flash", False),
-        ("azure_ai", "claude-sonnet-4-6", True),
-        ("azure_ai", "gpt-5.6", False),
-        ("openai", "gpt-5.6", False),
-        ("gemini", "gemini-3.8-flash", False),
-    ],
-)
-def test_is_claude_tool_target(custom_llm_provider: str, model: str, expected: bool):
-    assert litellm_main._is_claude_tool_target(custom_llm_provider=custom_llm_provider, model=model) is expected
-
-
 @pytest.mark.parametrize("key", ["input_examples", "eager_input_streaming"])
 def test_drop_anthropic_only_tool_keys_strips_tool_and_function_levels(key: str):
     tools = [