Merge pull request #40526 from BerriAI/litellm_internal_staging

chore(ci): promote internal staging to main

basedpyright-code-budget.json

@@ -105,7 +105,7 @@
     "limit": 109
   },
   "reportUnknownMemberType": {
-    "limit": 38271
+    "limit": 38269
   },
   "reportUnknownParameterType": {
     "limit": 19584

enterprise/litellm_enterprise/proxy/hooks/managed_files.py

@@ -1801,7 +1801,16 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             # Remove conflicting keys from data to avoid duplicate keyword arguments
             filtered_data = {k: v for k, v in data.items() if k not in ("model", "file_id")}
             for model_id, model_file_id in specific_model_file_id_mapping.items():
-                delete_response = await llm_router.afile_delete(model=model_id, file_id=model_file_id, **filtered_data)  # type: ignore
+                credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_id)
+                delete_data = {
+                    **{k: v for k, v in filtered_data.items() if k != "_litellm_internal_model_credentials"},
+                    **(
+                        {"_litellm_internal_model_credentials": MappingProxyType(dict(credentials))}
+                        if credentials is not None
+                        else {}
+                    ),
+                }
+                delete_response = await llm_router.afile_delete(model=model_id, file_id=model_file_id, **delete_data)
 
         stored_file_object = await self.delete_unified_file_id(file_id, litellm_parent_otel_span)
 
@@ -1812,7 +1821,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 prom_logger.record_managed_file_deleted(result="success")
 
         if stored_file_object:
-            return stored_file_object
+            return OpenAIFileObject.model_validate(stored_file_object).model_copy(update={"id": file_id})
         elif delete_response:
             delete_response.id = file_id
             return delete_response

litellm-proxy-extras/litellm_proxy_extras/migrations/20260717000000_cascade_delete_jwt_key_mapping_on_token_delete/migration.sql

@@ -0,0 +1,15 @@
+-- DropForeignKey
+DO $$
+BEGIN
+    IF EXISTS (SELECT 1 FROM pg_constraint WHERE conname = 'LiteLLM_JWTKeyMapping_token_fkey') THEN
+        ALTER TABLE "LiteLLM_JWTKeyMapping" DROP CONSTRAINT "LiteLLM_JWTKeyMapping_token_fkey";
+    END IF;
+END $$;
+
+-- AddForeignKey
+DO $$
+BEGIN
+    IF NOT EXISTS (SELECT 1 FROM pg_constraint WHERE conname = 'LiteLLM_JWTKeyMapping_token_fkey') THEN
+        ALTER TABLE "LiteLLM_JWTKeyMapping" ADD CONSTRAINT "LiteLLM_JWTKeyMapping_token_fkey" FOREIGN KEY ("token") REFERENCES "LiteLLM_VerificationToken"("token") ON DELETE CASCADE ON UPDATE CASCADE;
+    END IF;
+END $$;

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -492,7 +492,7 @@ model LiteLLM_JWTKeyMapping {
     updated_at        DateTime @default(now()) @updatedAt
     updated_by        String?
 
-    litellm_verification_token LiteLLM_VerificationToken @relation(fields: [token], references: [token])
+    litellm_verification_token LiteLLM_VerificationToken @relation(fields: [token], references: [token], onDelete: Cascade)
 
     @@unique([jwt_claim_name, jwt_claim_value])
     @@index([jwt_claim_name, jwt_claim_value, is_active])

litellm/__init__.py

@@ -546,7 +546,7 @@ _key_management_system: Optional["KeyManagementSystem"] = None
 #### PII MASKING ####
 output_parse_pii: bool = False
 #############################################
-from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map
+from litellm.litellm_core_utils.get_model_cost_map import get_model_cost_map, mark_litellm_import_complete
 
 model_cost = get_model_cost_map(url=model_cost_map_url)
 cost_discount_config: Dict[str, float] = {}  # Provider-specific cost discounts {"vertex_ai": 0.05} = 5% discount
@@ -2405,3 +2405,5 @@ def __getattr__(name: str) -> Any:
 
 
 # ALL_LITELLM_RESPONSE_TYPES is lazy-loaded via __getattr__ to avoid loading utils at import time
+
+mark_litellm_import_complete()

litellm/_internal_context.py

@@ -6,9 +6,33 @@ be settable from user input. Context variables are scoped to the current
 asyncio task and cannot be injected via HTTP request bodies.
 """
 
+from collections.abc import Generator
+from contextlib import contextmanager
 from contextvars import ContextVar
+from datetime import datetime, timezone
 from typing import Final
 
 # When True, suppresses async logging and billing for internal sub-calls
 # (e.g., emulated file-search steps that make nested LLM calls).
 is_internal_call: Final[ContextVar[bool]] = ContextVar("is_internal_call", default=False)
+
+# One request prices its totals, its per-token-type lines and the rates it reports on
+# separate code paths. Each reads the clock for off-peak pricing, so without a pinned
+# moment they can land on either side of a window boundary and disagree with each other.
+_billing_time: Final[ContextVar[datetime | None]] = ContextVar("billing_time", default=None)
+
+
+@contextmanager
+def pinned_billing_time(moment: datetime) -> Generator[None]:
+    """Price every rate lookup inside this block at ``moment`` rather than at each one's own clock read."""
+    token: Final = _billing_time.set(moment)
+    try:
+        yield
+    finally:
+        _billing_time.reset(token)
+
+
+def current_billing_time() -> datetime:
+    """The pinned billing moment, or now in UTC outside a pinned block."""
+    pinned: Final = _billing_time.get()
+    return pinned if pinned is not None else datetime.now(timezone.utc)

litellm/a2a_protocol/providers/bedrock_agentcore/handler.py

@@ -13,6 +13,7 @@ from litellm._logging import verbose_logger
 from litellm.a2a_protocol.providers.bedrock_agentcore.transformation import (
     BedrockAgentCoreA2ATransformation,
 )
+from litellm.llms.bedrock.base_aws_llm import run_aws_signing
 from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
 from litellm.types.llms.custom_http import httpxSpecialProvider
 
@@ -45,7 +46,8 @@ class BedrockAgentCoreA2AHandler:
         Returns:
             A2A JSON-RPC response dict from the AgentCore agent
         """
-        url, headers, body = BedrockAgentCoreA2ATransformation.get_url_and_signed_request(
+        url, headers, body = await run_aws_signing(
+            BedrockAgentCoreA2ATransformation.get_url_and_signed_request,
             request_id=request_id,
             params=params,
             litellm_params=litellm_params,
@@ -91,7 +93,8 @@ class BedrockAgentCoreA2AHandler:
         Yields:
             A2A streaming response events from the AgentCore agent
         """
-        url, headers, body = BedrockAgentCoreA2ATransformation.get_url_and_signed_request(
+        url, headers, body = await run_aws_signing(
+            BedrockAgentCoreA2ATransformation.get_url_and_signed_request,
             request_id=request_id,
             params=params,
             litellm_params=litellm_params,

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -25,7 +25,7 @@ import litellm
 from litellm import ModelResponse
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
-    responses_reasoning_item_from_thinking_blocks,
+    responses_reasoning_items_from_thinking_blocks,
 )
 from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
 from litellm.llms.base_llm.bridges.completion_transformation import (
@@ -129,8 +129,8 @@ def _reasoning_input_items(msg: "AllMessageValues") -> list[dict[str, object]]:
         return stored
     raw_blocks: Final = msg.get("thinking_blocks") or ()
     blocks: Final = cast("Iterable[ChatCompletionThinkingBlock]", raw_blocks)  # cast-ok: untyped client json
-    from_thinking: Final = responses_reasoning_item_from_thinking_blocks(blocks)
-    return [] if from_thinking is None else [dict(from_thinking)]  # mutable-ok: API message payload
+    replayed: Final = responses_reasoning_items_from_thinking_blocks(blocks)
+    return [dict(item) for item in replayed]  # mutable-ok: API message payload
 
 
 def _build_reasoning_item(
@@ -227,7 +227,7 @@ class _ChatToolCallDict(ChatCompletionToolCallChunk, total=False):
     provider_specific_fields: Mapping[str, object]
 
 
-def _tool_call_dict_from_output_item(item: Mapping[str, Any], index: int) -> _ChatToolCallDict:
+def tool_call_dict_from_output_item(item: Mapping[str, Any], index: int) -> _ChatToolCallDict:
     """Convert a ``function_call`` or ``custom_tool_call`` output item dict to a chat
     completions tool_call dict. Custom (grammar/freeform) tool calls carry their raw
     string payload in ``input`` rather than ``arguments``; both map to
@@ -755,7 +755,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
                     # Tool calls accumulate into the single trailing tool_calls choice
                     # like the typed branches above; a choice per call would hide every
                     # call after choices[0] from chat clients
-                    accumulated_tool_calls.append(_tool_call_dict_from_output_item(raw_item, tool_call_index))
+                    accumulated_tool_calls.append(tool_call_dict_from_output_item(raw_item, tool_call_index))
                     tool_call_index += 1
                 elif handle_raw_dict_callback is not None:
                     choice, index = handle_raw_dict_callback(item=raw_item, index=index)
@@ -1409,7 +1409,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
             # New output item added
             output_item = parsed_chunk.get("item", {})
             if output_item.get("type") in ("function_call", "custom_tool_call"):
-                converted: Final = _tool_call_dict_from_output_item(output_item, parsed_chunk.get("output_index", 0))
+                converted: Final = tool_call_dict_from_output_item(output_item, parsed_chunk.get("output_index", 0))
                 provider_specific_fields: Final = converted.get("provider_specific_fields")
 
                 function_chunk: Final = ChatCompletionToolCallFunctionChunk(
@@ -1484,7 +1484,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                                 index=0,
                                 delta=Delta(
                                     tool_calls=(
-                                        _tool_call_dict_from_output_item(
+                                        tool_call_dict_from_output_item(
                                             output_item, parsed_chunk.get("output_index", 0)
                                         ),
                                     )

litellm/constants.py

@@ -143,6 +143,7 @@ DEFAULT_MCP_SEMANTIC_FILTER_SIMILARITY_THRESHOLD: Final = float(
     os.getenv("DEFAULT_MCP_SEMANTIC_FILTER_SIMILARITY_THRESHOLD", 0.3)
 )
 MAX_MCP_SEMANTIC_FILTER_TOOLS_HEADER_LENGTH: Final = int(os.getenv("MAX_MCP_SEMANTIC_FILTER_TOOLS_HEADER_LENGTH", 150))
+MAX_GUARDRAIL_SCAN_METADATA_HEADER_LENGTH: Final = 2048
 
 DEFAULT_AUTO_ROUTER_MAX_INPUT_CHARS: Final = 2000
 
@@ -197,6 +198,7 @@ LITELLM_UI_ALLOW_HEADERS: Final = [
     "x-litellm-adaptive-router-model",
     "x-litellm-applied-guardrails",
     "x-litellm-guardrail-scan-id",
+    "x-litellm-guardrail-scan-metadata",
     "x-litellm-cache-key",
 ]
 
@@ -333,6 +335,7 @@ DEFAULT_SSL_CIPHERS: Final = os.getenv(
 
 ########### v2 Architecture constants for managing writing updates to the database ###########
 REDIS_UPDATE_BUFFER_KEY: Final = "litellm_spend_update_buffer"
+REDIS_GATEWAY_REQUESTS_BUFFER_KEY: Final = "litellm_gateway_requests_buffer"
 REDIS_DAILY_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_spend_update_buffer"
 REDIS_DAILY_TEAM_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_team_spend_update_buffer"
 REDIS_DAILY_ORG_SPEND_UPDATE_BUFFER_KEY: Final = "litellm_daily_org_spend_update_buffer"
@@ -395,6 +398,18 @@ TIKTOKEN_ENCODE_CHUNK_SIZE_CHARS: Final = get_env_int_in_range(
     minimum=1,
     maximum=TIKTOKEN_ENCODE_MAX_CHUNK_SIZE_CHARS,
 )
+TOKEN_COUNTER_MAX_EXACT_CHARS: Final = get_env_int_in_range(
+    "TOKEN_COUNTER_MAX_EXACT_CHARS",
+    default=4_000_000,
+    minimum=1,
+    maximum=1_000_000_000,
+)
+TOKEN_COUNTER_MAX_CONCURRENT_COUNTS: Final = get_env_int_in_range(
+    "TOKEN_COUNTER_MAX_CONCURRENT_COUNTS",
+    default=4,
+    minimum=1,
+    maximum=256,
+)
 MAX_TILE_WIDTH: Final = int(os.getenv("MAX_TILE_WIDTH", 512))
 MAX_TILE_HEIGHT: Final = int(os.getenv("MAX_TILE_HEIGHT", 512))
 OPENAI_FILE_SEARCH_COST_PER_1K_CALLS: Final = float(os.getenv("OPENAI_FILE_SEARCH_COST_PER_1K_CALLS", 2.5 / 1000))
@@ -567,6 +582,7 @@ LOGGING_WORKER_AGGRESSIVE_CLEAR_COOLDOWN_SECONDS: Final = float(
 LOGGING_EXECUTOR_MAX_THREADS: Final = get_env_int("LOGGING_EXECUTOR_MAX_THREADS", 100)
 LOGGING_EXECUTOR_MAX_PENDING_TASKS: Final = get_env_int("LOGGING_EXECUTOR_MAX_PENDING_TASKS", 10_000)
 LOGGING_EXECUTOR_DROPPED_TASK_LOG_INTERVAL_SECONDS: Final = 30.0
+AWS_SIGNING_MAX_THREADS: Final = 16
 DD_TRACER_STREAMING_CHUNK_YIELD_RESOURCE: Final = os.getenv(
     "DD_TRACER_STREAMING_CHUNK_YIELD_RESOURCE", "streaming.chunk.yield"
 )
@@ -1767,6 +1783,10 @@ LITELLM_SETTINGS_SAFE_DB_OVERRIDES: Final = [
 SPECIAL_LITELLM_AUTH_TOKEN: Final = ["ui-token"]
 DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL = int(os.getenv("DEFAULT_MANAGEMENT_OBJECT_IN_MEMORY_CACHE_TTL", 60))
 DEFAULT_ACCESS_GROUP_CACHE_TTL: Final = int(os.getenv("DEFAULT_ACCESS_GROUP_CACHE_TTL", 600))
+SPEND_LOG_KEY_METADATA_CACHE_TTL: Final = 600
+SPEND_LOG_KEY_METADATA_MISS_CACHE_TTL: Final = 30
+SPEND_LOG_KEY_METADATA_CACHE_MAX_ITEMS: Final = 10000
+SPEND_LOG_KEY_METADATA_QUERY_TIMEOUT_MS: Final = 5000
 # Short TTL for negative MCP access-group existence lookups. Keeps unauthenticated
 # callers from forcing a DB query per request for unknown names, while bounding
 # staleness so a transient DB error (which surfaces as an empty list) cannot

litellm/cost_calculator.py

@@ -25,6 +25,7 @@ from litellm.litellm_core_utils.llm_cost_calc.usage_object_transformation import
     TranscriptionUsageObjectTransformation,
 )
 from litellm.litellm_core_utils.llm_cost_calc.utils import (
+    BilledTokenRates,
     CostCalculatorUtils,
     _generic_cost_per_character,
     _get_regional_uplift_multiplier,
@@ -45,6 +46,9 @@ from litellm.llms.azure.cost_calculation import (
 from litellm.llms.azure_ai.cost_calculator import (
     cost_per_token as azure_ai_cost_per_token,
 )
+from litellm.llms.azure_ai.cost_calculator import (
+    is_azure_model_router as azure_ai_is_model_router_name,
+)
 from litellm.llms.base_llm.search.transformation import SearchResponse
 from litellm.llms.bedrock.cost_calculation import (
     cost_per_token as bedrock_cost_per_token,
@@ -1122,6 +1126,7 @@ def _store_cost_breakdown_in_logging_obj(
     service_tier: str | None = None,
     data_residency: str | None = None,
     vertex_location: str | None = None,
+    billed_token_rates: BilledTokenRates | None = None,
 ) -> None:
     """
     Helper function to store cost breakdown in the logging object.
@@ -1166,6 +1171,7 @@ def _store_cost_breakdown_in_logging_obj(
             service_tier=service_tier,
             data_residency=data_residency,
             vertex_location=vertex_location,
+            billed_token_rates=billed_token_rates,
         )
 
     except Exception as breakdown_error:
@@ -1659,11 +1665,10 @@ def completion_cost(
                     data_residency=data_residency,
                     vertex_location=vertex_location,
                     response=completion_response,
-                    request_model=request_model_for_cost,
                 )
 
                 # Get additional costs from provider (e.g., routing fees, infrastructure costs)
-                if custom_llm_provider == "azure_ai":
+                if custom_llm_provider == "azure_ai" and not azure_ai_is_model_router_name(model):
                     model_for_additional_costs = request_model_for_cost
                     if completion_response is not None:
                         hidden_params = getattr(completion_response, "_hidden_params", None) or {}
@@ -1735,6 +1740,7 @@ def completion_cost(
                     _reasoning_cost: float | None = None
                     _cache_read_cost: float | None = None
                     _cache_creation_cost: float | None = None
+                    _billed_token_rates: BilledTokenRates | None = None
                     if cost_per_token_usage_object is not None and model:
                         _breakdown_provider: str | None = (
                             custom_llm_provider if isinstance(custom_llm_provider, str) else None
@@ -1746,10 +1752,12 @@ def completion_cost(
                             service_tier=service_tier,
                             data_residency=data_residency,
                             vertex_location=vertex_location,
+                            custom_cost_per_token=custom_cost_per_token,
                         )
                         _reasoning_cost = _token_type_breakdown.reasoning_cost
                         _cache_read_cost = _token_type_breakdown.cache_read_cost
                         _cache_creation_cost = _token_type_breakdown.cache_creation_cost
+                        _billed_token_rates = _token_type_breakdown.rates
                     _store_cost_breakdown_in_logging_obj(
                         litellm_logging_obj=litellm_logging_obj,
                         prompt_tokens_cost_usd_dollar=prompt_tokens_cost_usd_dollar,
@@ -1769,6 +1777,7 @@ def completion_cost(
                         service_tier=service_tier,
                         data_residency=data_residency,
                         vertex_location=vertex_location,
+                        billed_token_rates=_billed_token_rates,
                     )
 
                 return _final_cost

litellm/experimental_mcp_client/client.py

@@ -18,6 +18,7 @@ from mcp import ClientSession, McpError, ReadResourceResult, Resource, StdioServ
 from mcp.client.sse import sse_client
 from mcp.client.stdio import stdio_client
 from mcp.shared.message import SessionMessage
+from mcp.shared.session import RequestResponder
 from typing_extensions import Unpack
 
 _TransportStreams: TypeAlias = tuple[
@@ -56,10 +57,13 @@ def missing_streamable_http_client_error() -> ImportError:
 from mcp.types import CallToolRequestParams as MCPCallToolRequestParams
 from mcp.types import CallToolResult as MCPCallToolResult
 from mcp.types import (
+    ClientResult,
     GetPromptRequestParams,
     GetPromptResult,
     Prompt,
     ResourceTemplate,
+    ServerNotification,
+    ServerRequest,
     TextContent,
 )
 from mcp.types import Tool as MCPTool
@@ -146,8 +150,8 @@ _SDK_READ_TIMEOUT_CODE: Final = int(httpx.codes.REQUEST_TIMEOUT)
 otherwise carries JSON-RPC error codes."""
 
 
-def _as_read_timeout(exc: BaseException) -> TimeoutError | None:
-    """The session read timeout elapsing, re-expressed as a ``TimeoutError``, or ``None``.
+def as_mcp_read_timeout(exc: BaseException) -> TimeoutError | None:
+    """Normalize an MCP SDK read timeout for client and gateway diagnostics, or return ``None``.
 
     The SDK reports its own elapsed read timeout as ``McpError`` carrying an HTTP status code in a
     field that otherwise holds JSON-RPC error codes, and it relays an upstream's JSON-RPC error
@@ -442,6 +446,18 @@ class MCPClient:
         in_flight_error: BaseException | None = None
         try:
             read_stream, write_stream = transport[0], transport[1]
+            stream_error: Final[asyncio.Future[Exception]] = asyncio.get_running_loop().create_future()
+
+            async def receive_message(
+                message: RequestResponder[ServerRequest, ClientResult] | ServerNotification | Exception,
+            ) -> None:
+                if not isinstance(message, (ValueError, httpx.RequestError, OSError)):
+                    return
+                if not stream_error.done():
+                    stream_error.set_result(message)
+                # The SDK closes pending requests when its message handler raises.
+                raise RuntimeError("MCP response stream failed")
+
             # Build session kwargs with optional callbacks
             session_kwargs: Final[dict[str, Any]] = {}
             if self._sampling_callback is not None:
@@ -456,6 +472,7 @@ class MCPClient:
                 read_stream,
                 write_stream,
                 read_timeout_seconds=timedelta(seconds=self.timeout),
+                message_handler=receive_message,
                 **session_kwargs,
             )
             session: Final = await session_ctx.__aenter__()
@@ -467,6 +484,10 @@ class MCPClient:
                     if isinstance(ins, str) and ins.strip():
                         self._last_initialize_instructions = ins.strip()
                 return await operation(session)
+            except McpError:
+                if stream_error.done():
+                    raise stream_error.result()
+                raise
             finally:
                 try:
                     await session_ctx.__aexit__(None, None, None)
@@ -501,11 +522,10 @@ class MCPClient:
             transport_ctx, http_client = self._create_transport_context()
             return await self._execute_session_operation(transport_ctx, operation)
         except Exception as e:
-            read_timeout: Final = _as_read_timeout(e)
+            read_timeout: Final = as_mcp_read_timeout(e)
             if read_timeout is not None:
                 verbose_logger.warning(
-                    "MCP client timed out after %ss waiting for %s to answer; the server accepted the "
-                    "request and ended its response stream without a JSON-RPC reply",
+                    "MCP client timed out after %ss waiting for a valid MCP response from %s",
                     self.timeout,
                     self.server_url or "stdio",
                 )

litellm/files/main.py

@@ -31,7 +31,7 @@ FileCreateProvider = Literal[
 FileRetrieveProvider = Literal[
     "openai", "azure", "gemini", "vertex_ai", "hosted_vllm", "litellm_proxy", "manus", "anthropic"
 ]
-FileDeleteProvider = Literal["openai", "azure", "gemini", "litellm_proxy", "manus", "anthropic"]
+FileDeleteProvider = Literal["openai", "azure", "gemini", "bedrock", "litellm_proxy", "manus", "anthropic"]
 FileListProvider = Literal["openai", "azure", "litellm_proxy", "manus", "anthropic"]
 import litellm
 from litellm import get_secret_str

litellm/integrations/azure_sentinel/azure_sentinel.py

@@ -21,6 +21,8 @@ from types import MappingProxyType
 from typing import Final, TypeVar
 from urllib.parse import urlparse
 
+import httpx
+
 from litellm._logging import verbose_logger
 from litellm.integrations.batch_utils import (
     BatchSendCancelled,
@@ -418,7 +420,7 @@ class AzureSentinelLogger(CustomBatchLogger):
             "Content-Type": "application/json",
         }
 
-        async def _send_batch(batch: Sequence[_QueuedPayload]):
+        async def _send_batch(batch: Sequence[_QueuedPayload]) -> httpx.Response:
             body: Final = safe_dumps(batch)
             return await self.async_httpx_client.post(
                 url=api_endpoint,

litellm/integrations/custom_guardrail.py

@@ -850,20 +850,24 @@ class CustomGuardrail(CustomLogger):
         if self.should_run_guardrail(data=request_data, event_type=GuardrailEventHooks.post_call) is not True:
             return None
 
-        # CHECK IF GUARDRAIL REJECTS THE REQUEST
         target: Final = self._deployment_hook_target()
-        hook_request_data: Final = {**request_data, "guardrail_to_apply": self} if target is not self else request_data
-        result: Final = await target.async_post_call_success_hook(
-            user_api_key_dict=UserAPIKeyAuth(
-                user_id=request_data.get("user_api_key_user_id"),
-                team_id=request_data.get("user_api_key_team_id"),
-                end_user_id=request_data.get("user_api_key_end_user_id"),
-                api_key=request_data.get("user_api_key_hash"),
-                request_route=request_data.get("user_api_key_request_route"),
-            ),
-            data=hook_request_data,
-            response=response,
-        )
+        try:
+            if target is not self:
+                request_data["guardrail_to_apply"] = self  # rebind-ok: dispatch consumes this key
+            result: Final = await target.async_post_call_success_hook(
+                user_api_key_dict=UserAPIKeyAuth(
+                    user_id=request_data.get("user_api_key_user_id"),
+                    team_id=request_data.get("user_api_key_team_id"),
+                    end_user_id=request_data.get("user_api_key_end_user_id"),
+                    api_key=request_data.get("user_api_key_hash"),
+                    request_route=request_data.get("user_api_key_request_route"),
+                ),
+                data=request_data,
+                response=response,
+            )
+        finally:
+            if target is not self:
+                request_data.pop("guardrail_to_apply", None)
 
         if not self._is_valid_response_type(result):
             return None

litellm/integrations/langfuse/langfuse.py

@@ -2,6 +2,7 @@
 #    On success, logs events to Langfuse
 import inspect
 import os
+import re
 import traceback
 from collections.abc import Callable, Iterable, Mapping
 from datetime import datetime
@@ -63,6 +64,44 @@ def _object_mapping(value: object) -> Mapping[str, object] | None:
     return value if isinstance(value, dict) else None
 
 
+def _widened_items(mapping: Mapping[str, object]) -> Iterable[tuple[object, object]]:
+    """Header pairs with the key type widened back to what a caller-supplied dict can actually hold."""
+    return mapping.items()
+
+
+def _is_session_header_trace(trace_id: object, session_id: object, proxy_server_request: object) -> bool:
+    if not isinstance(trace_id, str) or not isinstance(session_id, str):
+        return False
+    request: Final = _object_mapping(proxy_server_request)
+    raw_headers: Final = _object_mapping(request.get("headers")) if request is not None else None
+    if raw_headers is None:
+        return False
+    headers: Final = MappingProxyType(
+        {key.lower(): value for key, value in _widened_items(raw_headers) if isinstance(key, str)}
+    )
+    if headers.get("x-litellm-trace-id"):
+        return False
+    if headers.get("langfuse_trace_id") is not None:
+        return False
+    if trace_id != session_id and headers.get("langfuse_session_id") != session_id:
+        return False
+    if headers.get("x-litellm-session-id") == trace_id:
+        return True
+    if re.fullmatch(r"[a-zA-Z0-9_\-]{8,}", trace_id) is None:
+        return False
+    user_agent: Final = headers.get("user-agent")
+    codex: Final = isinstance(user_agent, str) and re.match(r"^codex[-_ /]", user_agent, re.IGNORECASE) is not None
+    return any(
+        value == trace_id
+        and (
+            key == "x-session-id"
+            or re.fullmatch(r"x-.+-session-id", key) is not None
+            or (codex and key in ("session-id", "session_id", "thread-id", "conversation_id"))
+        )
+        for key, value in headers.items()
+    )
+
+
 class _UsageObject(Protocol):
     """Token-count surface the Langfuse logger reads off a response usage payload."""
 
@@ -609,6 +648,18 @@ class LangFuseLogger:
             # This allows continuing an existing trace while still returning the correct trace_id
             if existing_trace_id is not None:
                 trace_id = existing_trace_id
+            resolved_trace_id: Final = (
+                litellm_call_id or trace_id
+                if existing_trace_id is None
+                and _is_session_header_trace(trace_id, session_id, litellm_params.get("proxy_server_request"))
+                else trace_id
+            )
+            if resolved_trace_id != trace_id:
+                verbose_logger.debug(
+                    "Langfuse: trace_id %s came from a session header; using call id %s so each call gets its own trace",
+                    trace_id,
+                    resolved_trace_id,
+                )
             requested_trace_keys: Final = _as_steering_key_sequence(clean_metadata.pop("update_trace_keys", ()))
             update_trace_keys: Final = (
                 requested_trace_keys if _as_steering_flag(litellm.langfuse_enable_update_trace_keys) else ()
@@ -663,7 +714,7 @@ class LangFuseLogger:
                     trace_params["output"] = masked_output if not mask_output else "redacted-by-litellm"
             else:  # don't overwrite an existing trace
                 trace_params = {
-                    "id": trace_id,
+                    "id": resolved_trace_id,
                     "name": trace_name,
                     "session_id": session_id,
                     "input": masked_input if not mask_input else "redacted-by-litellm",
@@ -845,13 +896,13 @@ class LangFuseLogger:
             # Verify langfuse accepted our trace_id; if it differs, log a warning but still return our intended value
             # to match expected test behavior
             if hasattr(generation_client, "trace_id") and generation_client.trace_id:
-                if generation_client.trace_id != trace_id:
+                if generation_client.trace_id != resolved_trace_id:
                     verbose_logger.warning(
                         "Langfuse trace_id mismatch: set %s, but langfuse returned %s. Using our intended trace_id for consistency.",
-                        trace_id,
+                        resolved_trace_id,
                         generation_client.trace_id,
                     )
-            return trace_id, generation_id
+            return resolved_trace_id, generation_id
         except Exception:
             verbose_logger.error("Langfuse Layer Error - %s", traceback.format_exc())
             return None, None

litellm/integrations/s3_v2.py

@@ -24,7 +24,7 @@ from litellm.integrations.s3 import (
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker
-from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
+from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, run_aws_signing
 from litellm.llms.custom_httpx.http_handler import (
     _get_httpx_client,
     get_async_httpx_client,
@@ -366,7 +366,7 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
             # Sign the request
             aws_request: Final = AWSRequest(method="PUT", url=url, data=json_string, headers=headers)
             aws_region_name: Final = self.get_aws_region_name_for_non_llm_api_calls(aws_region_name=self.s3_region_name)
-            S3SigV4Auth(credentials, "s3", aws_region_name).add_auth(aws_request)
+            await run_aws_signing(S3SigV4Auth(credentials, "s3", aws_region_name).add_auth, aws_request)
 
             # Prepare the signed headers
             signed_headers: Final = dict(aws_request.headers.items())
@@ -597,7 +597,7 @@ class S3Logger(CustomBatchLogger, BaseAWSLLM):
 
             # Sign the request
             aws_request: Final = AWSRequest(method="GET", url=url, headers=headers)
-            S3SigV4Auth(credentials, "s3", self.s3_region_name).add_auth(aws_request)
+            await run_aws_signing(S3SigV4Auth(credentials, "s3", self.s3_region_name).add_auth, aws_request)
 
             # Prepare the signed headers
             signed_headers: Final = dict(aws_request.headers.items())

litellm/integrations/sqs.py

@@ -22,7 +22,7 @@ from litellm.constants import (
     SQS_SEND_MESSAGE_ACTION,
 )
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
-from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
+from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, run_aws_signing
 from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
@@ -295,7 +295,7 @@ class SQSLogger(CustomBatchLogger, BaseAWSLLM):
                 data=prepped.body,
                 headers=prepped.headers,
             )
-            SigV4Auth(credentials, "sqs", self.sqs_region_name).add_auth(aws_request)
+            await run_aws_signing(SigV4Auth(credentials, "sqs", self.sqs_region_name).add_auth, aws_request)
 
             signed_headers: Final = dict(aws_request.headers.items())
 

litellm/litellm_core_utils/get_model_cost_map.py

@@ -2,6 +2,7 @@
 Pulls the cost + context window + provider route for known models from https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json
 
 This can be disabled by setting the LITELLM_LOCAL_MODEL_COST_MAP environment variable to True.
+The ``lite`` and ``litellm-proxy`` CLI entry points also use the bundled map without fetching.
 
 ```
 export LITELLM_LOCAL_MODEL_COST_MAP=True
@@ -13,11 +14,14 @@ import hashlib
 import json
 import os
 import random
+import sys
+import threading
 import time
 from collections.abc import Awaitable, Callable
 from dataclasses import dataclass, replace
 from datetime import datetime, timezone
 from importlib.resources import files
+from pathlib import Path
 from typing import Final, Protocol
 
 import httpx
@@ -33,6 +37,12 @@ from litellm.litellm_core_utils.fallback_generalizations import (
 )
 
 FALLBACK_GENERALIZATIONS_KEY: Final = "fallback_generalizations"
+_CLI_ENTRYPOINT_NAMES: Final = frozenset({"lite", "litellm-proxy"})
+
+
+def _is_cli_process() -> bool:
+    return Path(sys.argv[0]).stem in _CLI_ENTRYPOINT_NAMES
+
 
 # Reserved top-level keys that are not model entries. They must be excluded
 # from the model-count integrity check so a real upstream shrink can't be masked.
@@ -176,6 +186,11 @@ class GetModelCostMap:
 RETRYABLE_FETCH_STATUS_CODES: Final = frozenset({429, 500, 502, 503, 504})
 MODEL_COST_MAP_FETCH_MAX_ATTEMPTS: Final = 3
 MODEL_COST_MAP_FETCH_MAX_WAIT_SECONDS: Final = 30.0
+_litellm_import_complete = threading.Event()
+
+
+def mark_litellm_import_complete() -> None:
+    _litellm_import_complete.set()
 
 
 @dataclass(frozen=True, slots=True)
@@ -314,12 +329,13 @@ async def _fetch_remote_model_cost_map_with_retry(
 def _fetch_remote_model_cost_map_with_retry_sync(
     url: str,
     timeout: int,
-    max_attempts: int,
+    attempts: range,
     sleep: Callable[[float], None],
     rng: random.Random,
     client: _SyncGetClient,
 ) -> ModelCostMapReloadResult:
-    for attempt in range(1, max_attempts + 1):
+    max_attempts: Final = attempts.stop - 1
+    for attempt in attempts:
         outcome = _attempt_fetch_sync(client=client, url=url, timeout=timeout)
         if not isinstance(outcome, _FetchAttemptRetryable):
             return outcome
@@ -520,6 +536,68 @@ def _finalize_loaded_model_cost_map(loaded: ModelCostMapReloaded) -> ModelCostMa
     return replace(loaded, model_cost_map=_finalize_model_cost_map(loaded.model_cost_map))
 
 
+def adopt_model_cost_map(
+    new_model_cost_map: dict,  # mutable-ok: public API preserves the mutable cost-map contract
+) -> int:
+    import litellm
+    from litellm import utils
+
+    litellm.model_cost = new_model_cost_map
+    utils._invalidate_model_cost_lowercase_map()  # pyright: ignore[reportPrivateUsage]  # required cache invalidation
+    litellm.add_known_models(model_cost_map=new_model_cost_map)
+    fetched_model_count: Final = len(new_model_cost_map) if new_model_cost_map else 0
+    utils.reapply_runtime_model_cost_registrations()
+    return fetched_model_count
+
+
+def _retry_remote_fetch_in_background(
+    url: str,
+    timeout: int,
+    max_attempts: int,
+    sleep: Callable[[float], None],
+    rng: random.Random,
+    client: _SyncGetClient,
+    first_outcome: _FetchAttemptRetryable,
+) -> None:
+    try:
+        first_wait: Final = _next_retry_wait(outcome=first_outcome, attempt=1, max_attempts=max_attempts, rng=rng)
+        if isinstance(first_wait, ModelCostMapReloadUnavailable):
+            return
+        sleep(first_wait)
+        result: Final = _fetch_remote_model_cost_map_with_retry_sync(
+            url=url,
+            timeout=timeout,
+            attempts=range(2, max_attempts + 1),
+            sleep=sleep,
+            rng=rng,
+            client=client,
+        )
+        if isinstance(result, ModelCostMapReloadUnavailable):
+            verbose_logger.warning(
+                "LiteLLM: Failed to fetch remote model cost map from %s after %d attempts; keeping local backup",
+                url,
+                max_attempts,
+            )
+            return
+        _litellm_import_complete.wait()
+        if not GetModelCostMap.validate_model_cost_map(
+            fetched_map=result.model_cost_map,
+            backup_model_count=GetModelCostMap._get_backup_model_count(),  # pyright: ignore[reportPrivateUsage]  # integrity cache
+        ):
+            verbose_logger.warning(
+                "LiteLLM: Fetched model cost map failed integrity check. Using local backup instead. url=%s",
+                url,
+            )
+            return
+        finalized: Final = _finalize_loaded_model_cost_map(result).model_cost_map
+        _cost_map_source_info.source = "remote"
+        _cost_map_source_info.fallback_reason = None
+        _cost_map_source_info.loaded_at = datetime.now(timezone.utc)
+        adopt_model_cost_map(finalized)
+    except Exception as e:  # noqa: BLE001  # a failed background retry must not kill the task; the backup stays
+        verbose_logger.warning("LiteLLM: Background model cost map retry failed: %s", e)
+
+
 def get_model_cost_map(
     url: str,
     timeout: int = 5,
@@ -531,10 +609,12 @@ def get_model_cost_map(
     """
     Public entry point — returns the model cost map dict.
 
-    1. If ``LITELLM_LOCAL_MODEL_COST_MAP`` is set, uses the local backup only.
+    1. If ``LITELLM_LOCAL_MODEL_COST_MAP`` is set or this is a ``lite`` /
+       ``litellm-proxy`` CLI process, uses the local backup only.
     2. Otherwise fetches from ``url``, retrying transient HTTP errors
-       (429/5xx/transport) with Retry-After-aware backoff, validates
-       integrity, and falls back to the local backup on any failure.
+       (429/5xx/transport) with Retry-After-aware backoff in a background
+       thread, validates integrity, and falls back to the local backup on any
+       failure.
 
     Only the backup model count is cached (a single int) for validation.
     The full backup dict is only parsed when it must be *returned* as a
@@ -543,7 +623,7 @@ def get_model_cost_map(
     _cost_map_source_info.loaded_at = datetime.now(timezone.utc)
     # Note: can't use get_secret_bool here — this runs during litellm.__init__
     # before litellm._key_management_settings is set.
-    if os.getenv("LITELLM_LOCAL_MODEL_COST_MAP", "").lower() == "true":
+    if os.getenv("LITELLM_LOCAL_MODEL_COST_MAP", "").lower() == "true" or _is_cli_process():
         _cost_map_source_info.source = "local"
         _cost_map_source_info.url = None
         _cost_map_source_info.is_env_forced = True
@@ -553,24 +633,34 @@ def get_model_cost_map(
     _cost_map_source_info.url = url
     _cost_map_source_info.is_env_forced = False
 
-    result: Final = _fetch_remote_model_cost_map_with_retry_sync(
-        url=url,
-        timeout=timeout,
-        max_attempts=max_attempts,
-        sleep=sleep,
-        rng=rng if rng is not None else random.Random(),
-        client=client if client is not None else httpx,
-    )
-    if isinstance(result, ModelCostMapReloadUnavailable):
+    fetch_client: Final = client if client is not None else httpx
+    fetch_rng: Final = rng if rng is not None else random.Random()
+    outcome: Final = _attempt_fetch_sync(client=fetch_client, url=url, timeout=timeout)
+    if isinstance(outcome, _FetchAttemptRetryable) and max_attempts > 1:
+        threading.Thread(
+            target=_retry_remote_fetch_in_background,
+            kwargs={  # mutable-ok: threading requires a mutable keyword-arguments mapping
+                "url": url,
+                "timeout": timeout,
+                "max_attempts": max_attempts,
+                "sleep": sleep,
+                "rng": fetch_rng,
+                "client": fetch_client,
+                "first_outcome": outcome,
+            },
+            name="litellm-model-cost-map-retry",
+            daemon=True,
+        ).start()
+    if not isinstance(outcome, ModelCostMapReloaded):
         verbose_logger.warning(
             "LiteLLM: Failed to fetch remote model cost map from %s: %s. Falling back to local backup.",
             url,
-            result.reason,
+            outcome.reason,
         )
         _cost_map_source_info.source = "local"
-        _cost_map_source_info.fallback_reason = f"Remote fetch failed: {result.reason}"
+        _cost_map_source_info.fallback_reason = f"Remote fetch failed: {outcome.reason}"
         return _finalize_loaded_model_cost_map(GetModelCostMap.load_local_model_cost_map_with_revision()).model_cost_map
-    content: Final = result.model_cost_map
+    content: Final = outcome.model_cost_map
 
     # Validate using cached count (cheap int comparison, no file I/O)
     if not GetModelCostMap.validate_model_cost_map(
@@ -587,4 +677,4 @@ def get_model_cost_map(
 
     _cost_map_source_info.source = "remote"
     _cost_map_source_info.fallback_reason = None
-    return _finalize_loaded_model_cost_map(result).model_cost_map
+    return _finalize_loaded_model_cost_map(outcome).model_cost_map

litellm/litellm_core_utils/litellm_logging.py

@@ -203,6 +203,7 @@ if TYPE_CHECKING:
 
     from litellm.integrations.otel.logger import OpenTelemetryV2
     from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+    from litellm.litellm_core_utils.llm_cost_calc.utils import BilledTokenRates
     from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
 try:
     from litellm_enterprise.enterprise_callbacks.callback_controls import (
@@ -590,6 +591,7 @@ class Logging(LiteLLMLoggingBaseClass):
 
         # Initialize cost breakdown field
         self.cost_breakdown: CostBreakdown | None = None
+        self.billed_token_rates: BilledTokenRates | None = None
 
         # Init Caching related details
         self.caching_details: CachingDetails | None = None
@@ -1587,6 +1589,7 @@ class Logging(LiteLLMLoggingBaseClass):
         service_tier: str | None = None,
         data_residency: str | None = None,
         vertex_location: str | None = None,
+        billed_token_rates: "BilledTokenRates | None" = None,
     ) -> None:
         """
         Helper method to store cost breakdown in the logging object.
@@ -1606,8 +1609,10 @@ class Logging(LiteLLMLoggingBaseClass):
             service_tier: Tier the costs above were priced on, already resolved
             data_residency: Region uplift the costs above were priced on, already resolved
             vertex_location: Vertex AI location the costs above were priced on, already resolved
+            billed_token_rates: Per-token rates the costs above were billed at, already resolved
         """
 
+        self.billed_token_rates = billed_token_rates
         self.cost_breakdown = CostBreakdown(
             input_cost=input_cost,
             output_cost=output_cost,

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -10,6 +10,7 @@ from typing import Any, Final, Literal, TypedDict, cast
 from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
 
 import litellm
+from litellm._internal_context import current_billing_time
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.llm_cost_calc.tiered_pricing import (
     select_tier_for_input,
@@ -19,6 +20,7 @@ from litellm.types.utils import (
     CacheCreationTokenDetails,
     CallTypes,
     CompletionTokensDetailsWrapper,
+    CostPerToken,
     DataResidency,
     ImageResponse,
     ModelInfo,
@@ -305,7 +307,7 @@ def _is_within_off_peak_window(off_peak_hours_utc: str | Sequence[str], current_
     than being localised, so callers must pass datetime.now(timezone.utc), never datetime.now(),
     or every window shifts by the host's offset.
     """
-    reference: Final = current_time if current_time is not None else datetime.now(timezone.utc)
+    reference: Final = current_time if current_time is not None else current_billing_time()
     now: Final = (reference.astimezone(timezone.utc) if reference.tzinfo is not None else reference).time()
     windows: Final = (off_peak_hours_utc,) if isinstance(off_peak_hours_utc, str) else off_peak_hours_utc
     for window in windows:
@@ -392,7 +394,7 @@ def _is_off_peak(off_peak: Mapping[str, object], current_time: datetime | None =
     rules: the flat hours_utc windows, which apply every day, or any entry in windows, whose
     hours apply only on its weekdays.
     """
-    reference: Final = current_time if current_time is not None else datetime.now(timezone.utc)
+    reference: Final = current_time if current_time is not None else current_billing_time()
     reference_utc: Final = (
         reference.astimezone(timezone.utc) if reference.tzinfo is not None else reference.replace(tzinfo=timezone.utc)
     )
@@ -1195,7 +1197,7 @@ def generic_cost_per_token(
             usage.prompt_tokens - cache_hit - audio_tokens - cache_creation - image_tokens - video_tokens, 0
         )
 
-    billing_time: Final = current_time if current_time is not None else datetime.now(timezone.utc)
+    billing_time: Final = current_time if current_time is not None else current_billing_time()
     (
         prompt_base_cost,
         completion_base_cost,
@@ -1309,42 +1311,90 @@ def _coerce_token_count(value: object) -> int:
     return value if isinstance(value, int) and value > 0 else 0
 
 
+@dataclass(frozen=True, slots=True)
+class BilledTokenRates:
+    """Per-token rates one request's usage bills at, after token tiers, off-peak windows and the
+    regional multipliers the totals apply, so each cost line equals its token count times its rate."""
+
+    input_cost_per_token: float
+    output_cost_per_token: float
+    cache_read_input_token_cost: float
+    cache_creation_input_token_cost: float
+    cache_creation_input_token_cost_above_1hr: float
+    output_cost_per_reasoning_token: float
+
+    def scaled(self, multiplier: float) -> "BilledTokenRates":
+        if multiplier == 1.0:
+            return self
+        return BilledTokenRates(
+            input_cost_per_token=self.input_cost_per_token * multiplier,
+            output_cost_per_token=self.output_cost_per_token * multiplier,
+            cache_read_input_token_cost=self.cache_read_input_token_cost * multiplier,
+            cache_creation_input_token_cost=self.cache_creation_input_token_cost * multiplier,
+            cache_creation_input_token_cost_above_1hr=self.cache_creation_input_token_cost_above_1hr * multiplier,
+            output_cost_per_reasoning_token=self.output_cost_per_reasoning_token * multiplier,
+        )
+
+
 @dataclass(frozen=True, slots=True)
 class TokenTypeCostBreakdown:
     reasoning_cost: float
     cache_read_cost: float
     cache_creation_cost: float
+    rates: BilledTokenRates | None = None
+    """Rates these lines were billed at, so a caller reporting both cannot resolve them a second,
+    differently-argued way. None when the model's pricing could not be resolved."""
 
 
-def get_token_type_cost_breakdown(
-    model: str,
-    custom_llm_provider: str | None,
+def _reasoning_token_count(usage: Usage) -> int:
+    parsed: Final = (
+        parse_completion_tokens_details(usage)["reasoning_tokens"] if usage.completion_tokens_details is not None else 0
+    )
+    return parsed or _coerce_token_count(getattr(usage, "reasoning_tokens", 0))
+
+
+def _cache_token_counts(usage: Usage) -> tuple[int, int, CacheCreationTokenDetails | None]:
+    """(cache read tokens, cache creation tokens, cache creation details): read from prompt_tokens_details
+    first, then the private top-level counters the Usage constructor mirrors cache tokens onto for
+    providers/callers that bypass the details."""
+    parsed: Final = parse_prompt_tokens_details(usage) if usage.prompt_tokens_details is not None else None
+    parsed_read: Final = parsed["cache_hit_tokens"] if parsed is not None else 0
+    parsed_creation: Final = parsed["cache_creation_tokens"] if parsed is not None else 0
+    return (
+        parsed_read or _coerce_token_count(getattr(usage, "_cache_read_input_tokens", 0)),
+        parsed_creation or _coerce_token_count(getattr(usage, "_cache_creation_input_tokens", 0)),
+        parsed["cache_creation_token_details"] if parsed is not None else None,
+    )
+
+
+def _custom_pricing_rates(custom_cost_per_token: CostPerToken) -> BilledTokenRates:
+    """Flat custom pricing has no tiers, uplifts or reasoning rate: cache tokens bill at the configured
+    cache rates (else the input rate) and reasoning at the output rate, as _cost_per_token_custom_pricing_helper does."""
+    input_rate: Final = custom_cost_per_token["input_cost_per_token"]
+    output_rate: Final = custom_cost_per_token["output_cost_per_token"]
+    cache_creation_rate: Final = custom_cost_per_token.get("cache_creation_input_token_cost", input_rate)
+    return BilledTokenRates(
+        input_cost_per_token=input_rate,
+        output_cost_per_token=output_rate,
+        cache_read_input_token_cost=custom_cost_per_token.get("cache_read_input_token_cost", input_rate),
+        cache_creation_input_token_cost=cache_creation_rate,
+        cache_creation_input_token_cost_above_1hr=cache_creation_rate,
+        output_cost_per_reasoning_token=output_rate,
+    )
+
+
+def _cost_map_billed_rates(
+    model_info: ModelInfo,
     usage: Usage,
-    service_tier: str | None = None,
-    data_residency: str | None = None,
-    vertex_location: str | None = None,
-    current_time: datetime | None = None,
-) -> TokenTypeCostBreakdown:
-    """
-    Provider-agnostic cost of reasoning and cache tokens, derived from the usage
-    object and model pricing alone.
-
-    This works for every provider, including Perplexity/Cerebras/Dashscope whose
-    cost calculators bypass ``generic_cost_per_token``, because cache tokens always
-    land on ``prompt_tokens_details`` (via the Usage constructor and provider
-    transformations) and reasoning tokens on ``completion_tokens_details``. It reuses
-    the same rate-resolution primitives as the total-cost path so the breakdown can
-    never drift from the totals. Returns zeros (never raises) when the model or its
-    pricing cannot be resolved.
-    """
-    try:
-        model_info: Final = get_model_info(model=model, custom_llm_provider=custom_llm_provider)
-    except Exception:
-        return TokenTypeCostBreakdown(0.0, 0.0, 0.0)
-
-    billing_time: Final = current_time if current_time is not None else datetime.now(timezone.utc)
+    custom_llm_provider: str | None,
+    service_tier: str | None,
+    data_residency: str | None,
+    vertex_location: str | None,
+    current_time: datetime | None,
+) -> BilledTokenRates:
+    billing_time: Final = current_time if current_time is not None else current_billing_time()
     (
-        _prompt_base_cost,
+        prompt_base_cost,
         completion_base_cost,
         cache_creation_cost_rate,
         cache_creation_cost_above_1hr_rate,
@@ -1356,13 +1406,6 @@ def get_token_type_cost_breakdown(
         current_time=billing_time,
         threshold_is_inclusive=_uses_inclusive_token_thresholds(custom_llm_provider),
     )
-
-    reasoning_tokens = (
-        parse_completion_tokens_details(usage)["reasoning_tokens"] if usage.completion_tokens_details is not None else 0
-    )
-    if not reasoning_tokens:
-        reasoning_tokens = _coerce_token_count(getattr(usage, "reasoning_tokens", 0))
-
     reasoning_rate: Final = _resolve_billed_reasoning_rate(
         model_info=model_info,
         usage=usage,
@@ -1370,57 +1413,103 @@ def get_token_type_cost_breakdown(
         completion_base_cost=completion_base_cost,
         current_time=billing_time,
     )
-    reasoning_cost = float(reasoning_tokens) * reasoning_rate
+    multiplier: Final = (
+        _get_regional_uplift_multiplier(model_info, data_residency)
+        * get_vertex_regional_endpoint_uplift(model_info, vertex_location)
+        * get_provider_specific_geo_multiplier(model_info=model_info, usage=usage)
+    )
+    return BilledTokenRates(
+        input_cost_per_token=prompt_base_cost,
+        output_cost_per_token=completion_base_cost,
+        cache_read_input_token_cost=cache_read_cost_rate,
+        cache_creation_input_token_cost=cache_creation_cost_rate,
+        cache_creation_input_token_cost_above_1hr=cache_creation_cost_above_1hr_rate,
+        output_cost_per_reasoning_token=reasoning_rate,
+    ).scaled(multiplier)
 
-    cache_read_tokens = 0
-    cache_creation_tokens = 0
-    cache_creation_token_details: CacheCreationTokenDetails | None = None
-    if usage.prompt_tokens_details is not None:
-        prompt_tokens_details: Final = parse_prompt_tokens_details(usage)
-        cache_read_tokens = prompt_tokens_details["cache_hit_tokens"]
-        cache_creation_tokens = prompt_tokens_details["cache_creation_tokens"]
-        cache_creation_token_details = prompt_tokens_details["cache_creation_token_details"]
-    # Fall back to the private top-level counters the Usage constructor mirrors cache
-    # tokens onto, so providers/callers that bypass prompt_tokens_details are covered.
-    if not cache_read_tokens:
-        cache_read_tokens = _coerce_token_count(getattr(usage, "_cache_read_input_tokens", 0))
-    if not cache_creation_tokens:
-        cache_creation_tokens = _coerce_token_count(getattr(usage, "_cache_creation_input_tokens", 0))
 
-    cache_read_cost = float(cache_read_tokens) * cache_read_cost_rate
-    cache_creation_cost = calculate_cache_writing_cost(
-        cache_creation_tokens=cache_creation_tokens,
-        cache_creation_token_details=cache_creation_token_details,
-        cache_creation_cost_above_1hr=cache_creation_cost_above_1hr_rate,
-        cache_creation_cost=cache_creation_cost_rate,
+def get_billed_token_rates(
+    model: str,
+    custom_llm_provider: str | None,
+    usage: Usage,
+    service_tier: str | None = None,
+    data_residency: str | None = None,
+    vertex_location: str | None = None,
+    current_time: datetime | None = None,
+    custom_cost_per_token: CostPerToken | None = None,
+) -> BilledTokenRates | None:
+    """Rates the cost calculator bills ``usage`` at, resolved exactly as the totals and the token-type
+    breakdown resolve them. None when the model's pricing cannot be resolved."""
+    if custom_cost_per_token is not None:
+        return _custom_pricing_rates(custom_cost_per_token)
+    try:
+        model_info: Final = get_model_info(model=model, custom_llm_provider=custom_llm_provider)
+    except Exception:  # noqa: BLE001  # get_model_info raises a bare Exception for an unmapped model: no rates
+        return None
+    return _cost_map_billed_rates(
+        model_info=model_info,
+        usage=usage,
+        custom_llm_provider=custom_llm_provider,
+        service_tier=service_tier,
+        data_residency=data_residency,
+        vertex_location=vertex_location,
+        current_time=current_time,
     )
 
-    # Apply the same flat regional-processing uplift the totals get, so per-type
-    # costs stay reconciled with input_cost/output_cost for regionalized OpenAI hosts.
-    uplift: Final = _get_regional_uplift_multiplier(model_info, data_residency)
-    if uplift != 1.0:
-        reasoning_cost *= uplift
-        cache_read_cost *= uplift
-        cache_creation_cost *= uplift
 
-    vertex_uplift: Final = get_vertex_regional_endpoint_uplift(model_info, vertex_location)
-    if vertex_uplift != 1.0:
-        reasoning_cost *= vertex_uplift
-        cache_read_cost *= vertex_uplift
-        cache_creation_cost *= vertex_uplift
+def get_token_type_cost_breakdown(
+    model: str,
+    custom_llm_provider: str | None,
+    usage: Usage,
+    service_tier: str | None = None,
+    data_residency: str | None = None,
+    vertex_location: str | None = None,
+    current_time: datetime | None = None,
+    custom_cost_per_token: CostPerToken | None = None,
+) -> TokenTypeCostBreakdown:
+    """
+    Provider-agnostic cost of reasoning and cache tokens, derived from the usage
+    object and model pricing alone.
 
-    # Mirror the provider-specific geo uplift (e.g. Anthropic us: 1.1) the totals
-    # apply, so cache and reasoning line items stay reconciled with them.
-    geo_multiplier: Final = get_provider_specific_geo_multiplier(model_info=model_info, usage=usage)
-    if geo_multiplier != 1.0:
-        reasoning_cost *= geo_multiplier
-        cache_read_cost *= geo_multiplier
-        cache_creation_cost *= geo_multiplier
+    This works for every provider, including Perplexity/Cerebras/Dashscope whose
+    cost calculators bypass ``generic_cost_per_token``, because cache tokens always
+    land on ``prompt_tokens_details`` (via the Usage constructor and provider
+    transformations) and reasoning tokens on ``completion_tokens_details``. It reuses
+    the same rate resolution as the total-cost path (``get_billed_token_rates``) so the
+    breakdown can never drift from the totals. A deployment billed by
+    ``custom_cost_per_token`` is priced from those flat rates instead of the cost map and,
+    like its totals, bills cache writes flat rather than by their 5m/1h split.
+    Returns zeros (never raises) when the model or its pricing cannot be resolved.
+    """
+    rates: Final = get_billed_token_rates(
+        model=model,
+        custom_llm_provider=custom_llm_provider,
+        usage=usage,
+        service_tier=service_tier,
+        data_residency=data_residency,
+        vertex_location=vertex_location,
+        current_time=current_time,
+        custom_cost_per_token=custom_cost_per_token,
+    )
+    if rates is None:
+        return TokenTypeCostBreakdown(0.0, 0.0, 0.0)
 
+    cache_read_tokens, cache_creation_tokens, cache_creation_token_details = _cache_token_counts(usage)
+    cache_creation_cost: Final = (
+        float(cache_creation_tokens) * rates.cache_creation_input_token_cost
+        if custom_cost_per_token is not None
+        else calculate_cache_writing_cost(
+            cache_creation_tokens=cache_creation_tokens,
+            cache_creation_token_details=cache_creation_token_details,
+            cache_creation_cost_above_1hr=rates.cache_creation_input_token_cost_above_1hr,
+            cache_creation_cost=rates.cache_creation_input_token_cost,
+        )
+    )
     return TokenTypeCostBreakdown(
-        reasoning_cost=reasoning_cost,
-        cache_read_cost=cache_read_cost,
+        reasoning_cost=float(_reasoning_token_count(usage)) * rates.output_cost_per_reasoning_token,
+        cache_read_cost=float(cache_read_tokens) * rates.cache_read_input_token_cost,
         cache_creation_cost=cache_creation_cost,
+        rates=rates,
     )
 
 

litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py

@@ -3,7 +3,7 @@ import json
 import re
 import time
 import traceback
-from collections.abc import Iterable, Sequence
+from collections.abc import Mapping, Sequence
 from typing import Final, Literal, cast
 
 import litellm
@@ -151,6 +151,16 @@ def _clear_later_replay_slice_metadata(choice: StreamingChoices) -> None:
         del choice.enhancements
 
 
+def _invalid_choices_message(response_object: Mapping[str, object]) -> str:
+    raw_keys: Final = list(response_object.keys())
+    if "choices" not in response_object:
+        return f"LiteLLM: provider returned a response with no 'choices'. Raw keys: {raw_keys}"
+    return (
+        f"LiteLLM: provider returned 'choices' that is not a list ({type(response_object['choices']).__name__}). "
+        f"Raw keys: {raw_keys}"
+    )
+
+
 async def convert_to_streaming_response_async(
     response_object: dict | None = None,
 ):
@@ -179,14 +189,12 @@ async def convert_to_streaming_response_async(
 
     choice_list: Final[list[StreamingChoices]] = []
 
-    if not response_object.get("choices"):
+    if not isinstance(response_object.get("choices"), list):
         from litellm.exceptions import APIError
 
         raise APIError(
             status_code=500,
-            message=(
-                f"LiteLLM: provider returned a response with no 'choices'. Raw keys: {list(response_object.keys())}"
-            ),
+            message=_invalid_choices_message(response_object),
             llm_provider="",
             model="",
         )
@@ -287,14 +295,12 @@ def convert_to_streaming_response(
     model_response_object: Final = ModelResponseStream()
     choice_list: Final[list[StreamingChoices]] = []
 
-    if not response_object.get("choices"):
+    if not isinstance(response_object.get("choices"), list):
         from litellm.exceptions import APIError
 
         raise APIError(
             status_code=500,
-            message=(
-                f"LiteLLM: provider returned a response with no 'choices'. Raw keys: {list(response_object.keys())}"
-            ),
+            message=_invalid_choices_message(response_object),
             llm_provider="",
             model="",
         )
@@ -623,15 +629,12 @@ def convert_to_model_response_object(
                 return convert_to_streaming_response(response_object=response_object)
             choice_list: Final[list[Choices]] = []
 
-            if not response_object.get("choices") or not isinstance(response_object["choices"], Iterable):
+            if not isinstance(response_object.get("choices"), list):
                 from litellm.exceptions import APIError
 
                 raise APIError(
                     status_code=500,
-                    message=(
-                        "LiteLLM: provider returned a response with no 'choices'. "
-                        f"Raw keys: {list(response_object.keys())}"
-                    ),
+                    message=_invalid_choices_message(response_object),
                     llm_provider="",
                     model="",
                 )

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -6,7 +6,7 @@ import io
 import json
 import mimetypes
 import re
-from collections.abc import Iterable, Mapping, Sequence
+from collections.abc import Iterable, Iterator, Mapping, Sequence
 from itertools import groupby
 from os import PathLike
 from pathlib import Path
@@ -1823,14 +1823,11 @@ def _extract_reasoning_content(message: dict) -> tuple[str | None, str | None]:
     return None, message_content
 
 
-def _readable_thinking_text(
-    block: ChatCompletionThinkingBlock | ChatCompletionRedactedThinkingBlock,
-) -> str:
+def _readable_thinking_text(block: Mapping[str, object]) -> str:
     """The text a chat model can read back, empty for redacted blocks and malformed ones."""
     if block.get("type") != "thinking":
         return ""
-    thinking: Final = cast(ChatCompletionThinkingBlock, block).get("thinking")  # cast-ok: narrowed by the type tag
-    return str(thinking or "")
+    return str(block.get("thinking") or "")
 
 
 def reasoning_content_from_thinking_blocks(
@@ -1843,24 +1840,125 @@ def reasoning_content_from_thinking_blocks(
     return "\n".join(text for block in thinking_blocks if (text := _readable_thinking_text(block)))
 
 
-def responses_reasoning_item_from_thinking_blocks(
-    thinking_blocks: Iterable[ChatCompletionThinkingBlock | ChatCompletionRedactedThinkingBlock],
-) -> ChatCompletionReasoningItem | None:
-    """Build a Responses API `reasoning` input item from Anthropic thinking blocks.
+ENCRYPTED_REASONING_SIGNATURE_PREFIX: Final = "litellm_encrypted_reasoning:"
 
-    The item carries no `id`: the Responses API rejects an empty one and 404s on any id it
-    did not mint itself, while an item without an id is always accepted.
+
+def encrypted_reasoning_signature(encrypted_content: str) -> str:
+    """The opaque value a Responses API reasoning item's `encrypted_content` travels in.
+
+    Anthropic clients echo a thinking block's `signature` and a redacted block's `data`
+    back verbatim, so either field can carry the encrypted reasoning across turns; the
+    prefix tells the two apart from a signature Anthropic minted.
     """
+    return f"{ENCRYPTED_REASONING_SIGNATURE_PREFIX}{encrypted_content}"
+
+
+def _carries_encrypted_reasoning(signature: object) -> bool:
+    return isinstance(signature, str) and signature.startswith(ENCRYPTED_REASONING_SIGNATURE_PREFIX)
+
+
+def encrypted_content_from_signature(signature: object) -> str | None:
+    if not isinstance(signature, str) or not _carries_encrypted_reasoning(signature):
+        return None
+    return signature.removeprefix(ENCRYPTED_REASONING_SIGNATURE_PREFIX) or None
+
+
+def _encrypted_reasoning_field(block: Mapping[str, object]) -> object:
+    match block.get("type"):
+        case "thinking":
+            return block.get("signature")
+        case "redacted_thinking":
+            return block.get("data")
+        case _:
+            return None
+
+
+def encrypted_content_of_block(block: Mapping[str, object]) -> str | None:
+    return encrypted_content_from_signature(_encrypted_reasoning_field(block))
+
+
+def is_encrypted_reasoning_block(block: object) -> bool:
+    """A thinking or redacted_thinking block carrying Responses API encrypted reasoning.
+
+    Only the Responses API that minted the content can read it back, so an Anthropic
+    backend has to drop such a block rather than fail signature verification on it.
+    """
+    if not isinstance(block, Mapping):
+        return False
+    mapping: Final = cast(Mapping[str, object], block)  # cast-ok: narrowed by isinstance
+    return _carries_encrypted_reasoning(_encrypted_reasoning_field(mapping))
+
+
+def strip_encrypted_reasoning_from_messages(messages: object) -> None:
+    """Drop the bridge-tagged reasoning blocks a routed deployment cannot decrypt from
+    Anthropic-shaped history.
+
+    The whole block goes, the way #40280 drops undecryptable Responses ``input`` items: a
+    provider that did not mint the block rejects it signed (a foreign signature) and unsigned
+    (a missing signature) alike, so keeping its text as an unsigned thinking block only moves
+    the 400 from the router to the provider.
+
+    Mutates the content lists in place: the router's fallback snapshot shares these
+    message objects, so a rebound list would replay the stripped blocks on the fallback hop.
+    """
+    if not isinstance(messages, list):
+        return
+    for content in _anthropic_content_lists(cast(list[object], messages)):  # cast-ok: untyped client json
+        _strip_encrypted_reasoning_from_blocks(content)
+
+
+def _anthropic_content_lists(messages: Sequence[object]) -> Iterator[object]:
+    return (
+        cast(list[object], content)  # cast-ok: narrowed by isinstance
+        for message in messages
+        if isinstance(message, Mapping)
+        for content in (cast(Mapping[str, object], message).get("content"),)  # cast-ok: narrowed by isinstance
+        if isinstance(content, list)
+    )
+
+
+def _strip_encrypted_reasoning_from_blocks(content: object) -> None:
+    blocks: Final = cast(list[object], content)  # cast-ok: narrowed by the caller's isinstance
+    kept: Final = tuple(block for block in blocks if not is_encrypted_reasoning_block(block))
+    blocks[:] = kept  # rebind-ok: shared with fallback snapshot
+
+
+def _reasoning_replay_group_key(indexed_block: tuple[int, Mapping[str, object]]) -> str:
+    index, block = indexed_block
+    return f"encrypted:{index}" if is_encrypted_reasoning_block(block) else "summary"
+
+
+def _reasoning_item_from_block_group(group: tuple[Mapping[str, object], ...]) -> ChatCompletionReasoningItem | None:
     summary: Final[list[ChatCompletionReasoningSummaryTextBlock]] = [  # mutable-ok: API message payload
         ChatCompletionReasoningSummaryTextBlock(type="summary_text", text=text)
-        for block in thinking_blocks
+        for block in group
         if (text := _readable_thinking_text(block))
     ]
+    encrypted_content: Final = encrypted_content_of_block(group[0])
+    if encrypted_content is not None:
+        return ChatCompletionReasoningItem(type="reasoning", summary=summary, encrypted_content=encrypted_content)
     if not summary:
         return None
     return ChatCompletionReasoningItem(type="reasoning", summary=summary)
 
 
+def responses_reasoning_items_from_thinking_blocks(
+    thinking_blocks: Iterable[Mapping[str, object]],
+) -> tuple[ChatCompletionReasoningItem, ...]:
+    """Build Responses API `reasoning` input items from Anthropic thinking blocks.
+
+    A block carrying encrypted reasoning replays the item it came from byte for byte;
+    a run of plain thinking blocks collapses into one summary-only item. No item carries
+    an `id`: the Responses API 404s on any id it did not mint itself and rejects an empty
+    one, while an item without an id is always accepted.
+    """
+    return tuple(
+        item
+        for _, group in groupby(enumerate(thinking_blocks), key=_reasoning_replay_group_key)
+        if (item := _reasoning_item_from_block_group(tuple(block for _, block in group))) is not None
+    )
+
+
 def _parse_content_for_reasoning(
     message_text: str | None,
 ) -> tuple[str | None, str | None]:

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -46,6 +46,7 @@ from litellm.types.utils import GenericImageParsingChunk
 from .common_utils import (
     convert_content_list_to_str,
     infer_content_type_from_url_and_content,
+    is_encrypted_reasoning_block,
     is_non_content_values_set,
     parse_tool_call_arguments,
 )
@@ -2299,13 +2300,16 @@ def sanitize_messages_for_tool_calling(
 
 
 def _is_unsignable_thinking_block(block: object) -> bool:
-    """A `thinking` block that Anthropic cannot accept on input.
+    """A thinking block that Anthropic cannot accept on input.
 
     Anthropic verifies the thinking signature cryptographically, so a block whose
     signature is null, empty, or missing (e.g. from an open-source reasoning model)
-    is rejected with a 400 and must be dropped rather than blanked or repaired.
-    `redacted_thinking` blocks carry no signature and are always kept.
+    is rejected with a 400 and must be dropped rather than blanked or repaired, and
+    so is a block whose signature or data carries another provider's encrypted
+    reasoning. A `redacted_thinking` block Anthropic minted is always kept.
     """
+    if is_encrypted_reasoning_block(block):
+        return True
     if not isinstance(block, dict) or block.get("type") != "thinking":
         return False
     signature: Final = block.get("signature")

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -24,6 +24,7 @@ from litellm.types.utils import (
     Choices,
     CompletionTokensDetails,
     CompletionTokensDetailsWrapper,
+    Delta,
     Function,
     FunctionCall,
     ModelResponse,
@@ -326,6 +327,18 @@ class ChunkProcessor:
                 return chunk_id
         return ""
 
+    @staticmethod
+    def _get_role_from_chunks(chunks: Sequence["_BaseChunk"]) -> str:
+        return ChunkProcessor._role_of_choice(next((c["choices"][0] for c in chunks if c.get("choices")), None))
+
+    @staticmethod
+    def _role_of_choice(choice: object) -> str:
+        match choice:
+            case StreamingChoices(delta=Delta(role=str() as role)) | {"delta": {"role": str() as role}} if role:
+                return role
+            case _:
+                return "assistant"
+
     @staticmethod
     def _get_model_from_chunks(chunks: Sequence["_BaseChunk"], first_chunk_model: str) -> str:
         """
@@ -353,8 +366,7 @@ class ChunkProcessor:
         model: Final = ChunkProcessor._get_model_from_chunks(chunks, first_chunk_model)
         system_fingerprint: Final = chunk.get("system_fingerprint", None)
 
-        first_chunk_with_choices: Final = next((c for c in chunks if c.get("choices")), chunk)
-        role: Final = first_chunk_with_choices["choices"][0]["delta"]["role"]
+        role: Final = ChunkProcessor._get_role_from_chunks(chunks)
         finish_reason = "stop"
         for chunk in chunks:
             if "choices" in chunk and len(chunk["choices"]) > 0:

litellm/litellm_core_utils/streaming_handler.py

@@ -1473,17 +1473,14 @@ class CustomStreamWrapper:
                 self.received_finish_reason = response_obj["finish_reason"]
         elif self.custom_llm_provider == "cached_response":
             cached_chunk: Final = cast(ModelResponseStream, chunk)
-            chunk_finish_reason: Final = cached_chunk.choices[0].finish_reason
+            cached_choice: Final = cached_chunk.choices[0] if cached_chunk.choices else None
+            chunk_finish_reason: Final = cached_choice.finish_reason if cached_choice is not None else None
             response_obj = {
-                "text": cached_chunk.choices[0].delta.content,
+                "text": cached_choice.delta.content if cached_choice is not None else None,
                 "is_finished": chunk_finish_reason is not None,
                 "finish_reason": chunk_finish_reason,
                 "original_chunk": cached_chunk,
-                "tool_calls": (
-                    cached_chunk.choices[0].delta.tool_calls
-                    if hasattr(cached_chunk.choices[0].delta, "tool_calls")
-                    else None
-                ),
+                "tool_calls": (getattr(cached_choice.delta, "tool_calls", None) if cached_choice is not None else None),
             }
 
             completion_obj["content"] = response_obj["text"]

litellm/litellm_core_utils/token_counter.py

@@ -3,11 +3,15 @@
 import base64
 import io
 import struct
-from collections.abc import Callable, Iterable, Mapping, Sequence
+from collections.abc import Awaitable, Callable, Iterable, Mapping, Sequence
 from typing import Final, Literal, cast
 
+import anyio
+import anyio.lowlevel
 import httpx
 import tiktoken
+from tokenizers import Tokenizer
+from typing_extensions import ParamSpec, TypeVar
 
 import litellm
 from litellm import verbose_logger
@@ -21,7 +25,10 @@ from litellm.constants import (
     MAX_TILE_HEIGHT,
     MAX_TILE_WIDTH,
     TIKTOKEN_ENCODE_CHUNK_SIZE_CHARS,
+    TOKEN_COUNTER_MAX_CONCURRENT_COUNTS,
+    TOKEN_COUNTER_MAX_EXACT_CHARS,
 )
+from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.default_encoding import encoding as default_encoding
 from litellm.litellm_core_utils.url_utils import safe_get
 from litellm.llms.custom_httpx.http_handler import _get_httpx_client
@@ -317,6 +324,32 @@ TokenCounterFunction = Callable[[str], int]
 Type for a function that counts tokens in a string.
 """
 
+EXTRAPOLATION_SAMPLES: Final = 16
+T_ParamSpec: Final = ParamSpec("T_ParamSpec")
+T_Retval = TypeVar("T_Retval")
+_COUNT_OFFLOAD_LIMITER: Final = anyio.lowlevel.RunVar[anyio.CapacityLimiter]("litellm_count_offload_limiter")
+
+
+def _count_offload_limiter_for_this_loop() -> anyio.CapacityLimiter:
+    existing: Final = _COUNT_OFFLOAD_LIMITER.get(None)
+    if existing is not None:
+        return existing
+    created: Final = anyio.CapacityLimiter(TOKEN_COUNTER_MAX_CONCURRENT_COUNTS)
+    _COUNT_OFFLOAD_LIMITER.set(created)
+    return created
+
+
+def offload_token_count(
+    function: Callable[T_ParamSpec, T_Retval],
+) -> Callable[T_ParamSpec, Awaitable[T_Retval]]:
+    async def offloaded(
+        *args: T_ParamSpec.args,
+        **kwargs: T_ParamSpec.kwargs,  # kwargs-ok: ParamSpec keeps the wrapped function's own keyword contract
+    ) -> T_Retval:
+        return await asyncify(function, limiter=_count_offload_limiter_for_this_loop())(*args, **kwargs)
+
+    return offloaded
+
 
 def _get_tiktoken_count_function(
     encode_length: Callable[[str], int],
@@ -538,9 +571,40 @@ def _count_extra(
     return num_tokens
 
 
+def _get_extrapolating_count_function(
+    count_exactly: TokenCounterFunction,
+    max_exact_chars: int = TOKEN_COUNTER_MAX_EXACT_CHARS,
+) -> TokenCounterFunction:
+    def count_tokens(text: str) -> int:
+        if len(text) <= max_exact_chars:
+            return count_exactly(text)
+        samples: Final = _evenly_spaced_samples(text, max_exact_chars)
+        sampled_chars: Final = sum(len(sample) for sample in samples)
+        return round(sum(count_exactly(sample) for sample in samples) * len(text) / sampled_chars)
+
+    return count_tokens
+
+
+def _evenly_spaced_samples(text: str, total_chars: int) -> tuple[str, ...]:
+    sample_count: Final = min(EXTRAPOLATION_SAMPLES, total_chars)
+    sample_chars: Final = total_chars // sample_count
+    last_start: Final = len(text) - sample_chars
+    return tuple(
+        text[start : start + sample_chars]
+        for start in (last_start * index // max(sample_count - 1, 1) for index in range(sample_count))
+    )
+
+
 def _get_count_function(
     model: str | None,
     custom_tokenizer: dict | SelectTokenizerResponse | None = None,
+) -> TokenCounterFunction:
+    return _get_extrapolating_count_function(_get_exact_count_function(model, custom_tokenizer))
+
+
+def _get_exact_count_function(
+    model: str | None,
+    custom_tokenizer: dict | SelectTokenizerResponse | None = None,
 ) -> TokenCounterFunction:
     """
     Get the function to count tokens based on the model and custom tokenizer."""
@@ -549,10 +613,10 @@ def _get_count_function(
     if model is not None or custom_tokenizer is not None:
         tokenizer_json: Final = custom_tokenizer or _select_tokenizer(model)
         if tokenizer_json["type"] == "huggingface_tokenizer":
+            tokenizer: Final[Tokenizer] = tokenizer_json["tokenizer"]
 
             def count_tokens(text: str) -> int:
-                enc: Final = tokenizer_json["tokenizer"].encode(text)
-                return len(enc.ids)
+                return len(tokenizer.encode_batch_fast([text])[0])
 
             return count_tokens
         elif tokenizer_json["type"] == "openai_tokenizer":

litellm/llms/anthropic/chat/guardrail_translation/handler.py

@@ -13,10 +13,11 @@ Pattern Overview:
 """
 
 import json
-from collections.abc import Iterator, Mapping, Sequence
+from collections.abc import Mapping, MutableSequence, Sequence
 from copy import deepcopy
 from dataclasses import dataclass
 from itertools import chain, repeat
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Protocol, cast, overload, runtime_checkable
 
 from typing_extensions import ReadOnly, TypedDict, assert_never
@@ -41,6 +42,7 @@ from litellm.llms.base_llm.guardrail_translation.utils import (
     merge_guardrailed_scoped_messages,
     merge_returned_tools_into_request_tools,
     scoped_structured_message_indices,
+    stream_item_field,
     stream_item_fingerprint,
 )
 from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
@@ -153,6 +155,46 @@ class ExtractedInput:
 EMPTY_EXTRACTED_INPUT: Final = ExtractedInput(scanned=(), images=())
 
 
+@dataclass(frozen=True, slots=True)
+class _ToolCallShape:
+    name: str | None
+    arguments: str
+
+
+@dataclass(frozen=True, slots=True)
+class _SSEFieldRewrite:
+    """One field of one nested section of a buffered SSE event, rewritten."""
+
+    section: str
+    field: str
+    value: object
+
+
+class _SSEEventRewriter(Protocol):
+    def __call__(self, event: Mapping[str, object]) -> _SSEFieldRewrite | None: ...
+
+
+def _rewritten_event(event: Mapping[str, object], rewrite_event: _SSEEventRewriter) -> Mapping[str, object]:
+    rewrite: Final = rewrite_event(event)
+    section: Final = None if rewrite is None else event.get(rewrite.section)
+    if rewrite is None or not isinstance(section, Mapping):
+        return event
+    return {**event, rewrite.section: {**section, rewrite.field: rewrite.value}}  # mutable-ok: json.dumps needs a dict
+
+
+def _tool_call_shapes(tool_calls: Sequence[object]) -> tuple[_ToolCallShape, ...]:
+    """The guardrail-visible shape of each tool call, whether the guardrail handed
+    back the ``ChatCompletionMessageToolCall`` objects it was given or plain dicts."""
+    functions: Final = tuple(stream_item_field(tool_call, "function") for tool_call in tool_calls)
+    return tuple(
+        _ToolCallShape(
+            name=name if isinstance(name := stream_item_field(function, "name"), str) else None,
+            arguments=arguments if isinstance(arguments := stream_item_field(function, "arguments"), str) else "",
+        )
+        for function in functions
+    )
+
+
 class _AnthropicSSEDelta(TypedDict, total=False):
     type: ReadOnly[str]
     text: ReadOnly[str]
@@ -170,12 +212,18 @@ class AnthropicMessagesHandler(BaseTranslation):
     them through guardrail rewrites; downstream provider handling is out of scope.
     """
 
-    delivers_ended_stream_text_rewrites = True
+    delivers_ended_stream_rewrites = True
+    assembles_streamed_response = True
 
     def __init__(self):
         super().__init__()
         self.adapter = LiteLLMAnthropicMessagesAdapter()
 
+    def post_call_hook_response(self, response: object) -> object:
+        if not isinstance(response, ModelResponse):
+            return response
+        return self.adapter.translate_openai_response_to_anthropic(response)
+
     @staticmethod
     def _build_streaming_usage_response(
         responses_so_far: Sequence[object],
@@ -1050,6 +1098,7 @@ class AnthropicMessagesHandler(BaseTranslation):
                     first_choice.message.tool_calls,
                 )
                 string_so_far = first_choice.message.content
+                pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_list or ())
                 guardrail_inputs: Final = GenericGuardrailAPIInputs()
                 if string_so_far:
                     guardrail_inputs["texts"] = [string_so_far]
@@ -1084,6 +1133,19 @@ class AnthropicMessagesHandler(BaseTranslation):
                     and guardrailed_texts[0] != string_so_far
                 ):
                     self._write_ended_stream_text_rewrite(responses_so_far, guardrailed_texts[0])
+                if deliver_ended_stream_rewrites:
+                    returned_tool_calls: Final = _guardrailed_inputs.get("tool_calls")
+                    self._write_ended_stream_tool_call_rewrites(
+                        responses_so_far,
+                        pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+                        post_guardrail_tool_calls=_tool_call_shapes(
+                            returned_tool_calls
+                            if isinstance(returned_tool_calls, list)
+                            and len(returned_tool_calls) == len(pre_guardrail_tool_calls)
+                            else tool_calls_list or ()
+                        ),
+                        guardrail_name=guardrail_to_apply.guardrail_name or "unknown",
+                    )
             else:
                 verbose_proxy_logger.debug("Skipping output guardrail - model response has no choices")
             return responses_so_far
@@ -1206,44 +1268,124 @@ class AnthropicMessagesHandler(BaseTranslation):
 
     @staticmethod
     def _write_ended_stream_text_rewrite(
-        responses_so_far: list[Any],  # mutable-ok: rewrites the caller's buffered chunks in place
+        responses_so_far: MutableSequence[object],  # mutable-ok: rewrites the caller's buffered chunks in place
         rewritten_text: str,
     ) -> None:
         """Deliver an ended-stream guardrail text rewrite by rewriting the
         buffered chunks in place: the first ``text_delta`` carries the full
         rewritten text and every later one is blanked, leaving the surrounding
-        message and content-block framing untouched. Handles both chunk formats
-        this stream carries (parsed event dicts and raw SSE bytes)."""
+        message and content-block framing untouched."""
         replacements: Final = chain((rewritten_text,), repeat(""))
-        for idx, item in enumerate(responses_so_far):
-            if isinstance(item, dict):
-                delta = item.get("delta")
-                if item.get("type") == "content_block_delta" and isinstance(delta, dict):
-                    if delta.get("type") == "text_delta":
-                        delta["text"] = next(replacements)
-            elif isinstance(item, (bytes, bytearray)):
-                responses_so_far[idx] = (  # rebind-ok: delivers the rewrite into the caller's buffer
-                    AnthropicMessagesHandler._rewrite_sse_text_deltas(bytes(item), replacements)
-                )
+
+        def rewrite_text_delta(event: Mapping[str, object]) -> _SSEFieldRewrite | None:
+            delta: Final = event.get("delta")
+            if event.get("type") != "content_block_delta" or not isinstance(delta, Mapping):
+                return None
+            if delta.get("type") != "text_delta":
+                return None
+            return _SSEFieldRewrite("delta", "text", next(replacements))
+
+        AnthropicMessagesHandler._rewrite_ended_stream_events(responses_so_far, rewrite_text_delta)
+
+    @classmethod
+    def _write_ended_stream_tool_call_rewrites(
+        cls,
+        responses_so_far: MutableSequence[object],  # mutable-ok: rewrites the caller's buffered chunks in place
+        *,
+        pre_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+        post_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+        guardrail_name: str,
+    ) -> None:
+        """Deliver ended-stream guardrail tool-call rewrites by rewriting the
+        buffered chunks in place: the rebuilt response lists tool calls in the
+        order of the stream's ``tool_use`` blocks, so the nth rewritten call lands
+        on the nth block, its first ``input_json_delta`` carrying the full rewritten
+        arguments, every later one blanked, and ``content_block_start`` carrying the
+        rewritten name. Blocks that do not line up with the rebuilt tool calls make
+        the rewrite undeliverable, so the pipeline executor discards it and releases
+        the original chunks."""
+        if post_guardrail_tool_calls == pre_guardrail_tool_calls:
+            return
+        block_indices: Final = tuple(
+            index
+            for item in responses_so_far
+            for event in cls._iter_sse_events(item)
+            if event.get("type") == "content_block_start"
+            and isinstance(block := event.get("content_block"), Mapping)
+            and block.get("type") == "tool_use"
+            and isinstance(index := event.get("index"), int)
+        )
+        if len(block_indices) != len(post_guardrail_tool_calls):
+            from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+            raise UndeliverableStreamRewrite(guardrail_name)
+        rewrites_by_block: Final = MappingProxyType(
+            {
+                index: after
+                for index, before, after in zip(block_indices, pre_guardrail_tool_calls, post_guardrail_tool_calls)
+                if after != before
+            }
+        )
+        argument_replacements: Final = MappingProxyType(
+            {index: chain((rewrite.arguments,), repeat("")) for index, rewrite in rewrites_by_block.items()}
+        )
+
+        def rewrite_tool_use(event: Mapping[str, object]) -> _SSEFieldRewrite | None:
+            index: Final = event.get("index")
+            if not isinstance(index, int) or index not in rewrites_by_block:
+                return None
+            match event.get("type"):
+                case "content_block_start":
+                    name: Final = rewrites_by_block[index].name
+                    if name is None:
+                        return None
+                    return _SSEFieldRewrite("content_block", "name", name)
+                case "content_block_delta":
+                    delta: Final = event.get("delta")
+                    if not isinstance(delta, Mapping) or delta.get("type") != "input_json_delta":
+                        return None
+                    return _SSEFieldRewrite("delta", "partial_json", next(argument_replacements[index]))
+                case _:
+                    return None
+
+        cls._rewrite_ended_stream_events(responses_so_far, rewrite_tool_use)
 
     @staticmethod
-    def _rewrite_sse_text_deltas(sse_bytes: bytes, replacements: "Iterator[str]") -> bytes:
-        """Rewrite every ``text_delta`` data line in one SSE chunk with the next
-        replacement text, leaving all other events and framing byte-identical."""
+    def _rewrite_ended_stream_events(
+        responses_so_far: MutableSequence[object],  # mutable-ok: rewrites the caller's buffered chunks in place
+        rewrite_event: _SSEEventRewriter,
+    ) -> None:
+        """Replace every buffered event ``rewrite_event`` returns a rewrite for, in
+        both chunk formats this stream carries (parsed event dicts and raw SSE
+        bytes), leaving every other event and the framing untouched."""
+        rewritten_items: Final = tuple(
+            AnthropicMessagesHandler._rewrite_buffered_item(item, rewrite_event) for item in responses_so_far
+        )
+        responses_so_far[:] = rewritten_items  # rebind-ok: delivers the rewrites into the caller's buffer
+
+    @staticmethod
+    def _rewrite_buffered_item(item: object, rewrite_event: _SSEEventRewriter) -> object:
+        if isinstance(item, dict):
+            return _rewritten_event(_as_str_mapping(item), rewrite_event)
+        if isinstance(item, (bytes, bytearray)):
+            return AnthropicMessagesHandler._rewrite_sse_events(bytes(item), rewrite_event)
+        return item
+
+    @staticmethod
+    def _rewrite_sse_events(sse_bytes: bytes, rewrite_event: _SSEEventRewriter) -> bytes:
+        """Rewrite the data lines of one SSE chunk that ``rewrite_event`` rewrites,
+        leaving all other events and framing byte-identical."""
         try:
             decoded: Final = sse_bytes.decode("utf-8")
         except UnicodeDecodeError:
             return sse_bytes
         return "\n\n".join(
-            AnthropicMessagesHandler._rewrite_sse_block(block, replacements) for block in decoded.split("\n\n")
+            "\n".join(AnthropicMessagesHandler._rewrite_sse_line(line, rewrite_event) for line in block.split("\n"))
+            for block in decoded.split("\n\n")
         ).encode("utf-8")
 
     @staticmethod
-    def _rewrite_sse_block(block: str, replacements: "Iterator[str]") -> str:
-        return "\n".join(AnthropicMessagesHandler._rewrite_sse_line(line, replacements) for line in block.split("\n"))
-
-    @staticmethod
-    def _rewrite_sse_line(line: str, replacements: "Iterator[str]") -> str:
+    def _rewrite_sse_line(line: str, rewrite_event: _SSEEventRewriter) -> str:
         if not line.startswith("data:"):
             return line
         try:
@@ -1252,14 +1394,10 @@ class AnthropicMessagesHandler(BaseTranslation):
             )
         except json.JSONDecodeError:
             return line
-        if not isinstance(data, dict) or data.get("type") != "content_block_delta":
+        if not isinstance(data, dict):
             return line
-        delta: Final = data.get("delta")
-        if not isinstance(delta, dict) or delta.get("type") != "text_delta":
-            return line
-        return "data: " + json.dumps(
-            {**data, "delta": {**delta, "text": next(replacements)}}  # mutable-ok: json.dumps needs plain dicts
-        )
+        rewritten: Final = _rewritten_event(_as_str_mapping(data), rewrite_event)
+        return line if rewritten is data else "data: " + json.dumps(rewritten)
 
     def get_streaming_scan_key(self, responses_so_far: Sequence[object]) -> StreamingScanKey | None:
         stream_ended: Final = self._check_streaming_has_ended(responses_so_far)

litellm/llms/anthropic/common_utils.py

@@ -21,6 +21,7 @@ from litellm.constants import (
 )
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     get_file_ids_from_messages,
+    is_encrypted_reasoning_block,
 )
 from litellm.litellm_core_utils.prompt_templates.factory import (
     THOUGHT_SIGNATURE_SEPARATOR,
@@ -1201,6 +1202,32 @@ def strip_thinking_blocks_from_anthropic_messages(messages: list[Any]) -> list[A
     return out
 
 
+def _without_encrypted_reasoning_blocks(message: dict) -> dict | None:  # mutable-ok: Anthropic message payload shape
+    if not isinstance(message, Mapping):
+        return message
+    content: Final = message.get("content")
+    if not isinstance(content, list):
+        return message
+    kept: Final = [b for b in content if not is_encrypted_reasoning_block(b)]  # mutable-ok: API message payload
+    if len(kept) == len(content):
+        return message
+    if not kept:
+        return None
+    return {**message, "content": kept}  # mutable-ok: API message payload
+
+
+def strip_encrypted_reasoning_blocks_from_anthropic_messages(
+    messages: Sequence[dict],  # mutable-ok: Anthropic message payload shape
+) -> list[dict]:  # mutable-ok: AnthropicMessagesRequest.messages is typed list[dict]
+    """
+    Drop thinking / redacted_thinking blocks that carry another provider's encrypted
+    reasoning (a turn the Responses API bridge served) before the request reaches
+    Anthropic, which cannot verify them. Anthropic's own signed blocks are kept.
+    """
+    stripped: Final = (_without_encrypted_reasoning_blocks(m) for m in messages)
+    return [m for m in stripped if m is not None]  # mutable-ok: API message payload
+
+
 def strip_thinking_blocks_from_anthropic_messages_request_dict(
     data: dict[str, Any],
 ) -> None:

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

@@ -113,6 +113,7 @@ from litellm.litellm_core_utils.reasoning_effort_utils import (
 from litellm.llms.anthropic.common_utils import (
     is_empty_unsigned_thinking_block,
     normalize_anthropic_tool_use_id,
+    strip_encrypted_reasoning_blocks_from_anthropic_messages,
 )
 from litellm.llms.anthropic.experimental_pass_through.context_management import (
     PolyfillResult,
@@ -417,7 +418,8 @@ class LiteLLMAnthropicMessagesAdapter:
         model: str | None = None,
     ) -> list:
         new_messages: Final[list[AllMessageValues]] = []
-        for m in messages:
+        replayable_messages: Final = strip_encrypted_reasoning_blocks_from_anthropic_messages(messages)
+        for m in replayable_messages:
             user_message: ChatCompletionUserMessage | None = None
             tool_message_list: list[ChatCompletionToolMessage] = []
             new_user_content_list: list[ChatCompletionTextObject | ChatCompletionImageObject] = []
@@ -1487,8 +1489,9 @@ class LiteLLMAnthropicMessagesAdapter:
             anthropic_content.insert(0, polyfill_result.compaction_block)
 
         ## extract finish reason
+        openai_finish_reason: Final = response.choices[0].finish_reason if response.choices else "stop"
         translated_finish_reason: Final = self._translate_openai_finish_reason_to_anthropic(
-            openai_finish_reason=response.choices[0].finish_reason
+            openai_finish_reason=openai_finish_reason
         )
         anthropic_finish_reason: Final = (
             "refusal"

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

@@ -25,6 +25,7 @@ from ...common_utils import (
     AnthropicModelInfo,
     optionally_handle_anthropic_oauth,
     strip_advisor_blocks_from_messages,
+    strip_encrypted_reasoning_blocks_from_anthropic_messages,
 )
 
 DEFAULT_ANTHROPIC_API_VERSION: Final = "2023-06-01"
@@ -613,7 +614,7 @@ class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
             messages = strip_advisor_blocks_from_messages(messages)
 
         anthropic_messages_request: Final[AnthropicMessagesRequest] = AnthropicMessagesRequest(
-            messages=messages,
+            messages=strip_encrypted_reasoning_blocks_from_anthropic_messages(messages),
             max_tokens=max_tokens,
             model=model,
             **anthropic_messages_optional_request_params,

litellm/llms/anthropic/experimental_pass_through/responses_adapters/handler.py

@@ -19,6 +19,7 @@ from litellm.types.llms.anthropic_messages.anthropic_response import (
     AnthropicMessagesResponse,
 )
 from litellm.types.llms.openai import ResponsesAPIResponse
+from litellm.utils import ProviderConfigManager
 
 from ..utils import litellm_logging_obj_from_kwargs, local_model_name
 from .streaming_iterator import AnthropicResponsesStreamWrapper
@@ -34,6 +35,15 @@ def _forwarded_kwargs(extra_kwargs: Mapping[str, object] | None) -> Mapping[str,
     return extra_kwargs or {}
 
 
+def _provider_returns_encrypted_reasoning(model: str, custom_llm_provider: object) -> bool:
+    provider: Final = (
+        custom_llm_provider if isinstance(custom_llm_provider, str) else litellm.get_llm_provider(model=model)[1]
+    )
+    provider_model: Final = local_model_name(model, provider)
+    responses_config: Final = ProviderConfigManager.get_provider_responses_api_config(provider, provider_model)
+    return responses_config is not None and "include" in responses_config.get_supported_openai_params(provider_model)
+
+
 def _build_responses_kwargs(
     *,
     max_tokens: int,
@@ -85,8 +95,13 @@ def _build_responses_kwargs(
         request_data["output_format"] = output_format
 
     anthropic_request: Final = AnthropicMessagesRequest(**request_data)
-    responses_kwargs: Final = _ADAPTER.translate_request(anthropic_request)
     forwarded_kwargs: Final = _forwarded_kwargs(extra_kwargs)
+    responses_kwargs: Final = _ADAPTER.translate_request(
+        anthropic_request,
+        include_encrypted_reasoning=_provider_returns_encrypted_reasoning(
+            model, forwarded_kwargs.get("custom_llm_provider")
+        ),
+    )
 
     # Normalize reasoning effort based on model capabilities
     # (e.g. "max" → "xhigh"/"high", "minimal" → "low" if unsupported)
@@ -111,7 +126,7 @@ def _build_responses_kwargs(
         responses_kwargs["stream"] = True
 
     # Forward litellm-specific kwargs (api_key, api_base, logging obj, etc.)
-    excluded: Final = {"anthropic_messages"}
+    excluded: Final = frozenset(("anthropic_messages",))
     for key, value in forwarded_kwargs.items():
         if key == "litellm_logging_obj" and value is not None:
             from litellm.litellm_core_utils.litellm_logging import (
@@ -132,6 +147,14 @@ def _build_responses_kwargs(
     if explicit_prompt_cache_key is not None:
         responses_kwargs["prompt_cache_key"] = explicit_prompt_cache_key
 
+    deployment_include: Final = forwarded_kwargs.get("include")
+    bridge_include: Final = responses_kwargs.get("include")
+    if isinstance(deployment_include, list) and isinstance(bridge_include, list):
+        responses_kwargs["include"] = [
+            *bridge_include,
+            *(item for item in deployment_include if item not in bridge_include),
+        ]
+
     return responses_kwargs
 
 

litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py

@@ -9,13 +9,19 @@ from typing import TYPE_CHECKING, Any, Final
 
 from litellm import verbose_logger
 from litellm._uuid import uuid
+from litellm.litellm_core_utils.prompt_templates.common_utils import (
+    encrypted_reasoning_signature,
+)
 from litellm.llms.anthropic.experimental_pass_through.messages.utils import (
     refusal_stop_details,
     responses_output_refusal_text,
 )
 from litellm.types.llms.anthropic_messages.anthropic_response import AnthropicUsage
 
-from .transformation import LiteLLMAnthropicToResponsesAPIAdapter
+from .transformation import (
+    REASONING_SUMMARY_PART_SEPARATOR,
+    LiteLLMAnthropicToResponsesAPIAdapter,
+)
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObject
@@ -29,9 +35,10 @@ class AnthropicResponsesStreamWrapper:
       response.created                   -> message_start
       response.output_item.added         -> content_block_start (if message/function_call)
       response.output_text.delta         -> content_block_delta (text_delta)
+      response.reasoning_summary_part.added -> content_block_delta (thinking_delta separator)
       response.reasoning_summary_text.delta -> content_block_delta (thinking_delta)
       response.function_call_arguments.delta -> content_block_delta (input_json_delta)
-      response.output_item.done          -> content_block_stop
+      response.output_item.done          -> content_block_delta (signature_delta) + content_block_stop
       response.completed                 -> message_delta + message_stop
     """
 
@@ -94,6 +101,38 @@ class AnthropicResponsesStreamWrapper:
         )
         return block_idx
 
+    @staticmethod
+    def _field(source: object, name: str) -> object:
+        return source.get(name) if isinstance(source, dict) else getattr(source, name, None)
+
+    def _close_reasoning_item(self, item: object, item_id: str | None) -> None:
+        block_idx: Final = self._item_id_to_block_index.get(item_id, -1) if item_id else self._current_block_index
+        encrypted_content: Final = self._field(item, "encrypted_content")
+        signature: Final = (
+            encrypted_reasoning_signature(encrypted_content)
+            if isinstance(encrypted_content, str) and encrypted_content
+            else None
+        )
+        if block_idx < 0 and signature is None:
+            return
+        if block_idx < 0:
+            redacted_idx: Final = self._open_block(
+                item_id,
+                {"type": "redacted_thinking", "data": signature},  # mutable-ok: API message payload
+            )
+            stop: Final = {"type": "content_block_stop", "index": redacted_idx}  # mutable-ok: API message payload
+            self._chunk_queue.append(stop)
+            return
+        if signature is not None:
+            self._chunk_queue.append(
+                {  # mutable-ok: API message payload
+                    "type": "content_block_delta",
+                    "index": block_idx,
+                    "delta": {"type": "signature_delta", "signature": signature},  # mutable-ok: API message payload
+                }
+            )
+        self._chunk_queue.append({"type": "content_block_stop", "index": block_idx})  # mutable-ok: API message payload
+
     def _process_event(self, event: object) -> None:
         """Convert one Responses API event into zero or more Anthropic chunks queued for emission."""
         event_type = getattr(event, "type", None)
@@ -175,6 +214,26 @@ class AnthropicResponsesStreamWrapper:
             )
             return
 
+        if event_type == "response.reasoning_summary_part.added":
+            part_item_id: Final = self._field(event, "item_id")
+            summary_index: Final = self._field(event, "summary_index")
+            part_block_idx: Final = (
+                self._item_id_to_block_index.get(part_item_id, -1) if isinstance(part_item_id, str) else -1
+            )
+            if part_block_idx < 0 or not isinstance(summary_index, int) or summary_index == 0:
+                return
+            self._chunk_queue.append(
+                {  # mutable-ok: API message payload
+                    "type": "content_block_delta",
+                    "index": part_block_idx,
+                    "delta": {  # mutable-ok: API message payload
+                        "type": "thinking_delta",
+                        "thinking": REASONING_SUMMARY_PART_SEPARATOR,
+                    },
+                }
+            )
+            return
+
         # ---- reasoning summary text delta ----
         if event_type == "response.reasoning_summary_text.delta":
             item_id = getattr(event, "item_id", None) or (event.get("item_id") if isinstance(event, dict) else None)
@@ -220,6 +279,9 @@ class AnthropicResponsesStreamWrapper:
             item_id = (
                 getattr(item, "id", None) or (item.get("id") if isinstance(item, dict) else None) if item else None
             )
+            if self._field(item, "type") == "reasoning":
+                self._close_reasoning_item(item, item_id)
+                return
             block_idx = self._item_id_to_block_index.get(item_id, -1) if item_id else self._current_block_index
             if block_idx < 0:
                 return

litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py

@@ -13,7 +13,8 @@ from typing import Any, Final, cast
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     TOOL_RESULT_IMAGE_BOUNDARY,
     TOOL_RESULT_IMAGE_PLACEHOLDER,
-    responses_reasoning_item_from_thinking_blocks,
+    encrypted_reasoning_signature,
+    responses_reasoning_items_from_thinking_blocks,
     with_prompt_cache_breakpoint,
 )
 from litellm.litellm_core_utils.reasoning_effort_utils import (
@@ -33,6 +34,7 @@ from litellm.types.llms.anthropic import (
     AnthropicFinishReason,
     AnthropicMessagesRequest,
     AnthropicMessagesToolChoice,
+    AnthropicResponseContentBlockRedactedThinking,
     AnthropicResponseContentBlockText,
     AnthropicResponseContentBlockThinking,
     AnthropicResponseContentBlockToolUse,
@@ -43,11 +45,13 @@ from litellm.types.llms.anthropic_messages.anthropic_response import (
     AnthropicUsage,
 )
 from litellm.types.llms.openai import (
-    ChatCompletionThinkingBlock,
     ResponseAPIUsage,
     ResponsesAPIResponse,
 )
 
+REASONING_SUMMARY_PART_SEPARATOR: Final = "\n\n"
+RESPONSES_INCLUDE_ENCRYPTED_REASONING: Final = "reasoning.encrypted_content"
+
 
 class LiteLLMAnthropicToResponsesAPIAdapter:
     """
@@ -163,49 +167,55 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
         return str(getattr(part, "text", None) or "")
 
     @classmethod
-    def _thinking_blocks_from_reasoning_item(
+    def _thinking_block_from_reasoning_item(
         cls,
         summary: Iterable[object],
-    ) -> tuple[dict[str, Any], ...]:  # mutable-ok: API message payload
-        """Anthropic thinking blocks for one Responses reasoning item.
+        encrypted_content: object,
+    ) -> dict[str, Any] | None:  # mutable-ok: API message payload
+        """The one Anthropic block for a Responses reasoning item.
 
-        The signature stays empty: only Anthropic can sign a thinking block, and a stand-in
-        value would be replayed as a real one and rejected by every backend that verifies it.
+        The item's encrypted reasoning rides the block's opaque field (`signature`, or
+        `data` when there is no summary text) so the client echoes it back and the next
+        turn replays the very item OpenAI produced; without it the signature stays empty,
+        since only Anthropic can sign a thinking block.
         """
-        return tuple(
-            AnthropicResponseContentBlockThinking(
-                type="thinking",
-                thinking=text,
-                signature=None,
-            ).model_dump()
-            for part in summary
-            if (text := cls._summary_part_text(part))
+        text: Final = REASONING_SUMMARY_PART_SEPARATOR.join(
+            part_text for part in summary if (part_text := cls._summary_part_text(part))
         )
+        if not isinstance(encrypted_content, str) or not encrypted_content:
+            if not text:
+                return None
+            return AnthropicResponseContentBlockThinking(type="thinking", thinking=text, signature=None).model_dump()
+        signature: Final = encrypted_reasoning_signature(encrypted_content)
+        if not text:
+            return AnthropicResponseContentBlockRedactedThinking(type="redacted_thinking", data=signature).model_dump()
+        return AnthropicResponseContentBlockThinking(type="thinking", thinking=text, signature=signature).model_dump()
 
     @staticmethod
     def _assistant_block_group_key(indexed_block: tuple[int, Mapping[str, object]]) -> str:
         """Group a run of consecutive thinking blocks together; keep every other block alone."""
         index, block = indexed_block
-        return "thinking" if block.get("type") == "thinking" else f"block:{index}"
+        return "thinking" if block.get("type") in ("thinking", "redacted_thinking") else f"block:{index}"
 
     @classmethod
-    def _assistant_group_to_input_item(
+    def _assistant_group_to_input_items(
         cls, group: tuple[Mapping[str, object], ...]
-    ) -> dict[str, Any] | None:  # mutable-ok: API message payload
+    ) -> tuple[dict[str, Any], ...]:  # mutable-ok: API message payload
         first: Final = group[0]
         btype: Final = first.get("type")
-        if btype == "thinking":
-            blocks: Final = cast(tuple[ChatCompletionThinkingBlock, ...], group)  # cast-ok: untrusted client payload
-            reasoning_item: Final = responses_reasoning_item_from_thinking_blocks(blocks)
-            return None if reasoning_item is None else dict(reasoning_item)  # mutable-ok: API message payload
+        if btype in ("thinking", "redacted_thinking"):
+            replayed: Final = responses_reasoning_items_from_thinking_blocks(group)
+            return tuple(dict(item) for item in replayed)  # mutable-ok: API message payload
         if btype == "tool_use":
-            return {  # mutable-ok: API message payload
-                "type": "function_call",
-                "call_id": first.get("id", ""),
-                "name": first.get("name", ""),
-                "arguments": json.dumps(first.get("input", {})),  # mutable-ok: API message payload
-            }
-        return None
+            return (
+                {  # mutable-ok: API message payload
+                    "type": "function_call",
+                    "call_id": first.get("id", ""),
+                    "name": first.get("name", ""),
+                    "arguments": json.dumps(first.get("input", {})),  # mutable-ok: API message payload
+                },
+            )
+        return ()
 
     def translate_messages_to_responses_input(
         self,
@@ -362,7 +372,7 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
                     input_items.extend(
                         item
                         for _, group in groupby(enumerate(blocks), key=self._assistant_block_group_key)
-                        if (item := self._assistant_group_to_input_item(tuple(block for _, block in group))) is not None
+                        for item in self._assistant_group_to_input_items(tuple(block for _, block in group))
                     )
                     asst_parts: list[dict[str, Any]] = [  # mutable-ok: API message payload
                         {"type": "output_text", "text": block.get("text", "")}  # mutable-ok: API message payload
@@ -495,10 +505,16 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
     def translate_request(
         self,
         anthropic_request: AnthropicMessagesRequest,
+        include_encrypted_reasoning: bool = True,
     ) -> dict[str, Any]:
         """
         Translate a full Anthropic /v1/messages request dict to
         litellm.responses() / litellm.aresponses() kwargs.
+
+        ``include_encrypted_reasoning`` asks the provider for ``reasoning.encrypted_content``
+        on every call, so a reasoning model's items can be replayed intact next turn even
+        when the client sent no ``thinking`` block; pass False for a provider whose
+        Responses API rejects ``include``.
         """
         model: Final[str] = anthropic_request["model"]
         messages_list: Final = cast(
@@ -528,6 +544,8 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
             "model": model,
             "input": input_items,
         }
+        if include_encrypted_reasoning:
+            responses_kwargs["include"] = [RESPONSES_INCLUDE_ENCRYPTED_REASONING]  # mutable-ok: API request payload
 
         if system and not developer_parts:
             if isinstance(system, str):
@@ -634,7 +652,9 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
 
         for item in response.output:
             if isinstance(item, ResponseReasoningItem):
-                content.extend(self._thinking_blocks_from_reasoning_item(item.summary))
+                reasoning_block = self._thinking_block_from_reasoning_item(item.summary, item.encrypted_content)
+                if reasoning_block is not None:
+                    content.append(reasoning_block)
 
             elif isinstance(item, ResponseOutputMessage):
                 for part in item.content:
@@ -684,11 +704,12 @@ class LiteLLMAnthropicToResponsesAPIAdapter:
                                     ).model_dump()
                                 )
                 elif item_type == "reasoning":
-                    content.extend(
-                        self._thinking_blocks_from_reasoning_item(
-                            cast(Iterable[object], item.get("summary") or ()),  # cast-ok: untyped provider json
-                        )
+                    reasoning_block = self._thinking_block_from_reasoning_item(
+                        cast(Iterable[object], item.get("summary") or ()),  # cast-ok: untyped provider json
+                        item.get("encrypted_content"),
                     )
+                    if reasoning_block is not None:
+                        content.append(reasoning_block)
                 elif item_type == "function_call":
                     try:
                         input_data = json.loads(item.get("arguments", "{}"))

litellm/llms/azure/audio_transcriptions.py

@@ -37,6 +37,7 @@ class AzureAudioTranscription(AzureChatCompletion):
         azure_ad_token: str | None = None,
         atranscription: bool = False,
         litellm_params: dict | None = None,
+        custom_llm_provider: str = "azure",
     ) -> TranscriptionResponse | Coroutine[Any, Any, TranscriptionResponse]:
         data: Final = {"model": model, "file": audio_file, **optional_params}
 
@@ -53,6 +54,7 @@ class AzureAudioTranscription(AzureChatCompletion):
                 logging_obj=logging_obj,
                 model=model,
                 litellm_params=litellm_params,
+                custom_llm_provider=custom_llm_provider,
             )
 
         azure_client: Final = self.get_azure_openai_client(
@@ -99,7 +101,7 @@ class AzureAudioTranscription(AzureChatCompletion):
             additional_args={"complete_input_dict": data},
             original_response=stringified_response,
         )
-        hidden_params: Final = {"model": model, "custom_llm_provider": "azure"}
+        hidden_params: Final = {"model": model, "custom_llm_provider": custom_llm_provider}
         final_response: Final[TranscriptionResponse] = convert_to_model_response_object(
             response_object=stringified_response,
             model_response_object=model_response,
@@ -122,6 +124,7 @@ class AzureAudioTranscription(AzureChatCompletion):
         client=None,
         max_retries=None,
         litellm_params: dict | None = None,
+        custom_llm_provider: str = "azure",
     ) -> TranscriptionResponse:
         response = None
         try:
@@ -178,7 +181,7 @@ class AzureAudioTranscription(AzureChatCompletion):
                 },
                 original_response=stringified_response,
             )
-            hidden_params: Final = {"model": model, "custom_llm_provider": "azure"}
+            hidden_params: Final = {"model": model, "custom_llm_provider": custom_llm_provider}
             response = convert_to_model_response_object(
                 _response_headers=headers,
                 response_object=stringified_response,

litellm/llms/azure/common_utils.py

@@ -14,6 +14,7 @@ from typing_extensions import ReadOnly, TypedDict
 import litellm
 from litellm._logging import verbose_logger
 from litellm.caching.caching import DualCache
+from litellm.constants import DEFAULT_MAX_RETRIES
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.openai.common_utils import BaseOpenAILLM
 from litellm.secret_managers.get_azure_ad_token_provider import (
@@ -582,7 +583,8 @@ class BaseAzureLLM(BaseOpenAILLM):
         if scope is None:
             scope = "https://cognitiveservices.azure.com/.default"
 
-        max_retries: Final = litellm_params.get("max_retries")
+        configured_max_retries: Final = litellm_params.get("max_retries")
+        max_retries: Final = DEFAULT_MAX_RETRIES if configured_max_retries is None else configured_max_retries
         timeout: Final = litellm_params.get("timeout")
         if not api_key and azure_ad_token_provider is None and tenant_id and client_id and client_secret:
             verbose_logger.debug("Using Azure AD Token Provider from Entra ID for Azure Auth")
@@ -642,8 +644,7 @@ class BaseAzureLLM(BaseOpenAILLM):
         else:
             azure_client_params["http_client"] = self._get_sync_http_client()
 
-        if max_retries is not None:
-            azure_client_params["max_retries"] = max_retries
+        azure_client_params["max_retries"] = max_retries
         if timeout is not None:
             azure_client_params["timeout"] = timeout
 

litellm/llms/azure_ai/cost_calculator.py

@@ -11,7 +11,7 @@ from litellm.types.utils import Usage
 from litellm.utils import get_model_info
 
 
-def _is_azure_model_router(model: str) -> bool:
+def is_azure_model_router(model: str) -> bool:
     """
     Check if the model is Azure AI Foundry Model Router.
 
@@ -31,6 +31,18 @@ def _is_azure_model_router(model: str) -> bool:
     return "model-router" in model_lower or "model_router" in model_lower or model_lower == "azure-model-router"
 
 
+ROUTER_FEE_ENTRY_NAMES: Final = frozenset({"model-router", "model_router"})
+
+
+def is_router_fee_entry(model: str) -> bool:
+    return model.lower().removeprefix("azure_ai/") in ROUTER_FEE_ENTRY_NAMES
+
+
+def _router_fee_entry_name(model: str) -> str:
+    entry_name: Final = model.lower().removeprefix("azure_ai/")
+    return entry_name if entry_name in ROUTER_FEE_ENTRY_NAMES else "model_router"
+
+
 def calculate_azure_model_router_flat_cost(model: str, prompt_tokens: int) -> float:
     """
     Calculate the flat cost for Azure AI Foundry Model Router.
@@ -42,20 +54,39 @@ def calculate_azure_model_router_flat_cost(model: str, prompt_tokens: int) -> fl
     Returns:
         float: The flat cost in USD, or 0.0 if not applicable
     """
-    if not _is_azure_model_router(model):
+    if not is_azure_model_router(model):
         return 0.0
-
-    # Get the model router pricing from model_prices_and_context_window.json
-    # Use "model_router" as the key (without actual model name suffix)
-    model_info: Final = get_model_info(model="model_router", custom_llm_provider="azure_ai")
+    model_info: Final = get_model_info(model=_router_fee_entry_name(model), custom_llm_provider="azure_ai")
     router_flat_cost_per_token: Final = model_info.get("input_cost_per_token", 0)
-
     if router_flat_cost_per_token and router_flat_cost_per_token > 0:
         return prompt_tokens * router_flat_cost_per_token
-
     return 0.0
 
 
+def _response_model_cost(model: str, usage: Usage, service_tier: str | None) -> tuple[float, float]:
+    try:
+        return generic_cost_per_token(
+            model=model, usage=usage, custom_llm_provider="azure_ai", service_tier=service_tier
+        )
+    except Exception as e:
+        if not is_azure_model_router(model):
+            raise
+        verbose_logger.debug(
+            "Azure AI Model Router: model '%s' not in cost map, only the routing fee applies. Error: %s", model, e
+        )
+        return 0.0, 0.0
+
+
+def _router_fee_name(model: str, request_model: str | None) -> str | None:
+    if is_router_fee_entry(model):
+        return None
+    if is_azure_model_router(model):
+        return model
+    if request_model is not None and is_azure_model_router(request_model):
+        return request_model
+    return None
+
+
 def cost_per_token(
     model: str,
     usage: Usage,
@@ -64,68 +95,31 @@ def cost_per_token(
     service_tier: str | None = None,
 ) -> tuple[float, float]:
     """
-    Calculate the cost per token for Azure AI models.
+    Price the response model's own tokens for Azure AI, plus the Model Router fee exactly once when either the
+    priced name or request_model is a Model Router name.
 
-    For Azure AI Foundry Model Router:
-    - Adds a flat cost of $0.14 per million input tokens (from model_prices_and_context_window.json)
-    - Plus the cost of the actual model used (handled by generic_cost_per_token)
+    A response priced as the router entry itself already carries the fee, so nothing is added on top of it. A
+    router deployment name that is missing from the cost map prices at the fee alone.
+
+    completion_cost passes only the priced name: when that name is a routed model it adds the fee itself through
+    AzureModelRouterConfig.calculate_additional_costs as the "Azure Model Router Flat Cost" line of the cost
+    breakdown, and when the name is router-shaped the fee is already in the prompt cost returned here.
 
     Args:
         model: str, the model name without provider prefix (from response)
         usage: LiteLLM Usage block
         response_time_ms: Optional response time in milliseconds
-        request_model: Optional[str], the original request model name (to detect router usage)
+        request_model: Optional[str], the original request model name; a Model Router name adds the routing fee
+        service_tier: Optional service tier the request was priced on
 
     Returns:
         Tuple[float, float] - prompt_cost_in_usd, completion_cost_in_usd
 
     Raises:
-        ValueError: If the model is not found in the cost map and cost cannot be calculated
-            (except for Model Router models where we return just the routing flat cost)
+        ValueError: If a model that is not a Model Router name is missing from the cost map
     """
-    prompt_cost = 0.0
-    completion_cost = 0.0
-
-    # Determine if this was a model router request
-    # Check both the response model and the request model
-    is_router_request: Final = _is_azure_model_router(model) or (
-        request_model is not None and _is_azure_model_router(request_model)
-    )
-
-    # Calculate base cost using generic cost calculator
-    # This may raise an exception if the model is not in the cost map
-    try:
-        prompt_cost, completion_cost = generic_cost_per_token(
-            model=model,
-            usage=usage,
-            custom_llm_provider="azure_ai",
-            service_tier=service_tier,
-        )
-    except Exception as e:
-        # For Model Router, the model name (e.g., "azure-model-router") may not be in the cost map
-        # because it's a routing service, not an actual model. In this case, we continue
-        # to calculate just the routing flat cost.
-        if not _is_azure_model_router(model):
-            # Re-raise for non-router models - they should have pricing defined
-            raise
-        verbose_logger.debug(
-            "Azure AI Model Router: model '%s' not in cost map, calculating routing flat cost only. Error: %s", model, e
-        )
-
-    # Add flat cost for Azure Model Router
-    # The flat cost is defined in model_prices_and_context_window.json for azure_ai/model_router
-    if is_router_request:
-        # Use the request model for flat cost calculation if available, otherwise use response model
-        router_model_for_calc: Final = request_model if request_model else model
-        router_flat_cost: Final = calculate_azure_model_router_flat_cost(router_model_for_calc, usage.prompt_tokens)
-
-        if router_flat_cost > 0:
-            verbose_logger.debug(
-                f"Azure AI Model Router flat cost: ${router_flat_cost:.6f} "
-                f"({usage.prompt_tokens} tokens × ${router_flat_cost / usage.prompt_tokens:.9f}/token)"
-            )
-
-            # Add flat cost to prompt cost
-            prompt_cost += router_flat_cost
-
-    return prompt_cost, completion_cost
+    prompt_cost, completion_cost = _response_model_cost(model=model, usage=usage, service_tier=service_tier)
+    fee_name: Final = _router_fee_name(model=model, request_model=request_model)
+    if fee_name is None:
+        return prompt_cost, completion_cost
+    return prompt_cost + calculate_azure_model_router_flat_cost(fee_name, usage.prompt_tokens), completion_cost

litellm/llms/azure_ai/image_edit/__init__.py

@@ -23,7 +23,7 @@ def get_azure_ai_image_edit_config(model: str) -> BaseImageEditConfig:
     """
     Get the appropriate image edit config for an Azure AI model.
 
-    - MAI models use /mai/v1/images/edits with multipart form data and size
+    - MAI models use /mai/v1/images/edits with multipart form data
     - FLUX 2 models use JSON with base64 image
     - FLUX 1 models use multipart/form-data
     """

litellm/llms/azure_ai/image_edit/mai_transformation.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, Any, Final, cast
+from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 from httpx._types import RequestFiles
@@ -13,7 +13,6 @@ from litellm.llms.azure_ai.image_generation.mai_transformation import (
 from litellm.llms.openai.common_utils import OpenAIError
 from litellm.llms.openai.image_edit.transformation import OpenAIImageEditConfig
 from litellm.secret_managers.main import get_secret_str
-from litellm.types.images.main import ImageEditOptionalRequestParams
 from litellm.types.llms.openai import FileTypes
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import ImageResponse
@@ -26,65 +25,8 @@ if TYPE_CHECKING:
 class AzureFoundryMAIImageEditConfig(OpenAIImageEditConfig):
     """Azure AI Foundry MAI image editing (e.g. MAI-Image-2.5)."""
 
-    DEFAULT_SIZE = "1024x1024"
-
     def get_supported_openai_params(self, model: str) -> list:
-        return ["prompt", "image", "model", "n", "size"]
-
-    def map_openai_params(
-        self,
-        image_edit_optional_params: ImageEditOptionalRequestParams,
-        model: str,
-        drop_params: bool,
-    ) -> dict:
-        optional_params: Final[dict[str, Any]] = {}
-        supported_params: Final = self.get_supported_openai_params(model)
-
-        for key, value in dict(image_edit_optional_params).items():
-            if value is None or key in optional_params:
-                continue
-
-            if key in supported_params:
-                if key == "size" and value:
-                    size_param = cast(str, value)
-                    self._validate_size_param(size_param)
-                    optional_params[key] = size_param
-                else:
-                    optional_params[key] = value
-            elif not drop_params:
-                raise ValueError(
-                    f"Parameter {key} is not supported for model {model}. "
-                    f"Supported parameters are {supported_params}. "
-                    f"Set drop_params=True to drop unsupported parameters."
-                )
-
-        if "size" not in optional_params:
-            optional_params["size"] = self.DEFAULT_SIZE
-
-        return optional_params
-
-    def _validate_size_param(self, size: str) -> None:
-        known_sizes: Final = {
-            "1024x1024",
-            "1792x1024",
-            "1024x1792",
-            "512x512",
-            "256x256",
-        }
-
-        if size in known_sizes:
-            return
-
-        if "x" in size:
-            try:
-                tuple(map(int, size.lower().split("x", 1)))
-                return
-            except ValueError:
-                raise ValueError(f"Invalid size format: '{size}'. Expected format 'WIDTHxHEIGHT' (e.g., '1024x1024').")
-
-        raise ValueError(
-            f"Unsupported size value: '{size}'. Use a known size (e.g., '1024x1024') or a custom 'WIDTHxHEIGHT' string."
-        )
+        return ["prompt", "image", "model", "n"]
 
     def validate_environment(
         self,

litellm/llms/azure_ai/image_generation/mai_transformation.py

@@ -2,6 +2,7 @@ from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
+from litellm.exceptions import UnsupportedParamsError
 from litellm.llms.base_llm.image_generation.transformation import (
     BaseImageGenerationConfig,
 )
@@ -21,6 +22,10 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
     DEFAULT_WIDTH = 1024
     DEFAULT_HEIGHT = 1024
 
+    MAX_IMAGES_PER_REQUEST: Final = 1
+    MIN_DIMENSION_PX: Final = 768
+    MAX_TOTAL_PX: Final = 1_056_768
+
     @staticmethod
     def get_mai_image_generation_url(
         api_base: str | None,
@@ -145,16 +150,27 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
 
             if k in supported_params:
                 if k == "size" and v:
-                    self._map_size_param(v, optional_params)
+                    self._map_size_param(v, optional_params, model)
+                elif k == "n" and v is not None and self._image_count(v, model) != self.MAX_IMAGES_PER_REQUEST:
+                    if not drop_params:
+                        raise self._unsupported(
+                            model,
+                            f"n={v} is not supported for model {model}. The Azure AI MAI image "
+                            f"endpoint returns exactly {self.MAX_IMAGES_PER_REQUEST} image per "
+                            "request and ignores any count, so a larger value would silently "
+                            "return fewer images than requested. Send one request per image, or "
+                            "set drop_params=True to drop n.",
+                        )
                 else:
                     optional_params[k] = v
             elif k in ("width", "height"):
                 optional_params[k] = v
             elif not drop_params:
-                raise ValueError(
+                raise self._unsupported(
+                    model,
                     f"Parameter {k} is not supported for model {model}. "
                     f"Supported parameters are {supported_params} and width/height. "
-                    f"Set drop_params=True to drop unsupported parameters."
+                    f"Set drop_params=True to drop unsupported parameters.",
                 )
 
         if "width" not in optional_params:
@@ -165,7 +181,19 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
         optional_params.pop("size", None)
         return optional_params
 
-    def _map_size_param(self, size: str, optional_params: dict) -> None:
+    @staticmethod
+    def _unsupported(model: str, message: str) -> UnsupportedParamsError:
+        return UnsupportedParamsError(message=message, llm_provider="azure_ai", model=model)
+
+    def _image_count(self, n: object, model: str) -> int:
+        if isinstance(n, int):
+            return n
+        try:
+            return int(str(n))
+        except ValueError:
+            raise self._unsupported(model, f"n={n!r} is not a whole number of images for model {model}.")
+
+    def _map_size_param(self, size: str, optional_params: dict, model: str) -> None:
         size_mapping: Final = {
             "1024x1024": (1024, 1024),
             "1792x1024": (1792, 1024),
@@ -176,19 +204,36 @@ class AzureFoundryMAIImageGenerationConfig(BaseImageGenerationConfig):
 
         if size in size_mapping:
             width, height = size_mapping[size]
-            optional_params["width"] = width
-            optional_params["height"] = height
         elif "x" in size:
             try:
                 width, height = map(int, size.lower().split("x"))
-                optional_params["width"] = width
-                optional_params["height"] = height
             except ValueError:
-                raise ValueError(f"Invalid size format: '{size}'. Expected format 'WIDTHxHEIGHT' (e.g., '1024x1024').")
+                raise self._unsupported(
+                    model, f"Invalid size format: '{size}'. Expected format 'WIDTHxHEIGHT' (e.g., '1024x1024')."
+                )
         else:
-            raise ValueError(
+            raise self._unsupported(
+                model,
                 f"Unsupported size value: '{size}'. "
-                f"Use a known size (e.g., '1024x1024') or a custom 'WIDTHxHEIGHT' string."
+                f"Use a known size (e.g., '1024x1024') or a custom 'WIDTHxHEIGHT' string.",
+            )
+
+        self._validate_dimensions(model=model, size=size, width=width, height=height)
+        optional_params["width"] = width
+        optional_params["height"] = height
+
+    def _validate_dimensions(self, model: str, size: str, width: int, height: int) -> None:
+        if width < self.MIN_DIMENSION_PX or height < self.MIN_DIMENSION_PX:
+            raise self._unsupported(
+                model,
+                f"Unsupported size value: '{size}'. Azure AI MAI image models require width and "
+                f"height of at least {self.MIN_DIMENSION_PX} pixels.",
+            )
+        if width * height > self.MAX_TOTAL_PX:
+            raise self._unsupported(
+                model,
+                f"Unsupported size value: '{size}'. Azure AI MAI image models accept at most "
+                f"{self.MAX_TOTAL_PX} total pixels ({width}x{height} is {width * height}).",
             )
 
     def transform_image_generation_response(

litellm/llms/base_llm/guardrail_translation/base_translation.py

@@ -52,13 +52,28 @@ class StreamingScanKey:
 
 
 class BaseTranslation(ABC):
-    delivers_ended_stream_text_rewrites: ClassVar[bool] = False
+    delivers_ended_stream_rewrites: ClassVar[bool] = False
     """Whether ``process_output_streaming_response`` accepts
     ``deliver_ended_stream_rewrites=True`` and, on an ended (fully buffered)
-    stream, writes guardrail text rewrites back across ``responses_so_far`` so
-    a buffered pipeline can release rewritten chunks. Tool-call rewrites, and
-    text rewrites on every other translation, are undeliverable: the pipeline
-    executor discards them and releases the original chunks."""
+    stream, writes guardrail text and tool-call rewrites back across
+    ``responses_so_far`` so a buffered pipeline can release rewritten chunks,
+    raising ``UndeliverableStreamRewrite`` for a shape it cannot place. Rewrites
+    on every other translation are undeliverable: the pipeline executor
+    discards them and releases the original chunks."""
+
+    assembles_streamed_response: ClassVar[bool] = False
+    """Whether ``process_output_streaming_response`` stores the assembled response of an
+    ended stream under ``request_data["response"]`` before scanning it, the way the chat,
+    Responses, and Messages translations do. A streaming pipeline runs a guardrail that only
+    has the legacy post-call hook against that response, so on a translation without it such
+    a guardrail keeps running on its own."""
+
+    def post_call_hook_response(self, response: object) -> object:
+        """The ``response`` this endpoint's non-streaming post-call hooks receive, derived from
+        the object the translation stores under ``request_data["response"]`` while scanning an
+        ended stream. Chat and Responses scan that shape already; a translation that scans a
+        different one (Messages scans an OpenAI-shaped ModelResponse) overrides this."""
+        return response
 
     @staticmethod
     def transform_user_api_key_dict_to_metadata(
@@ -175,9 +190,9 @@ class BaseTranslation(ABC):
         transformations (see ``StreamTransformSink``); base handlers ignore it.
         ``deliver_ended_stream_rewrites`` is passed True only when the caller
         holds the whole buffered stream and the subclass declares
-        ``delivers_ended_stream_text_rewrites``: the handler then writes
-        guardrail text rewrites back across ``responses_so_far`` instead of
-        discarding them.
+        ``delivers_ended_stream_rewrites``: the handler then writes
+        guardrail text and tool-call rewrites back across ``responses_so_far``
+        instead of discarding them.
         """
         return responses_so_far
 

litellm/llms/bedrock/base_aws_llm.py

@@ -1,13 +1,17 @@
+import asyncio
 import base64
+import contextvars
 import hashlib
 import json
 import os
 import re
 import urllib.parse
 from collections.abc import Callable, Mapping
+from concurrent.futures import ThreadPoolExecutor
 from datetime import datetime
+from functools import partial
 from threading import Lock
-from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, cast, get_args, overload
+from typing import TYPE_CHECKING, Any, ClassVar, Final, Literal, ParamSpec, TypeVar, cast, get_args, overload
 
 import httpx
 from pydantic import BaseModel, ValidationError
@@ -16,6 +20,7 @@ from litellm._logging import verbose_logger
 from litellm.caching.caching import DualCache
 from litellm.caching.in_memory_cache import InMemoryCache
 from litellm.constants import (
+    AWS_SIGNING_MAX_THREADS,
     BEDROCK_EMBEDDING_PROVIDERS_LITERAL,
     BEDROCK_IAM_CACHE_FETCH_LOCK_STRIPES,
     BEDROCK_IAM_CACHE_MAX_ENTRIES,
@@ -80,7 +85,11 @@ class AwsAuthError(Exception):
         super().__init__(self.message)  # Call the base class constructor with the parameters it needs
 
 
-class BaseAWSLLM:
+class SignsRequestsWithAWS:
+    pass
+
+
+class BaseAWSLLM(SignsRequestsWithAWS):
     # Process-wide IAM credential cache (shared across instances — Bedrock passthrough is per-request).
     # Storage is in-process memory only: no Redis backend unless attached elsewhere. Entry TTL: static
     # access-key + secret + region use ``_get_default_ttl_for_boto3_credentials`` (~59 minutes); ambient
@@ -1668,3 +1677,52 @@ class BaseAWSLLM:
             request_headers_dict["Authorization"] = incoming_authorization
 
         return request_headers_dict, request.body
+
+
+def sign_aws_json_post(
+    get_credentials: Callable[[], Credentials],
+    service_name: str,
+    aws_region_name: str | None,
+    url: str,
+    body: str,
+    headers: Mapping[str, str],
+) -> AWSPreparedRequest:
+    try:
+        from botocore.auth import SigV4Auth
+        from botocore.awsrequest import AWSRequest
+    except ImportError:
+        raise ImportError(f"Missing boto3 to call {service_name}. Run 'pip install boto3'.")
+
+    aws_request: Final = AWSRequest(method="POST", url=url, data=body, headers=headers)
+    SigV4Auth(get_credentials(), service_name, aws_region_name).add_auth(aws_request)
+    return aws_request.prepare()
+
+
+_SignParams = ParamSpec("_SignParams")
+_SignedRequest = TypeVar("_SignedRequest")
+
+AWS_SIGNING_EXECUTOR: Final = ThreadPoolExecutor(max_workers=AWS_SIGNING_MAX_THREADS, thread_name_prefix="aws-signing")
+
+
+async def run_aws_signing(
+    sign: Callable[_SignParams, _SignedRequest],
+    /,
+    *args: _SignParams.args,
+    **kwargs: _SignParams.kwargs,  # kwargs-ok: ParamSpec forwarding keeps the wrapped signing signature
+) -> _SignedRequest:
+    context: Final = contextvars.copy_context()
+    return await asyncio.get_running_loop().run_in_executor(
+        AWS_SIGNING_EXECUTOR, partial(context.run, sign, *args, **kwargs)
+    )
+
+
+async def sign_request_off_loop_if_aws(
+    provider_config: object,
+    sign_request: Callable[_SignParams, _SignedRequest],
+    /,
+    *args: _SignParams.args,
+    **kwargs: _SignParams.kwargs,  # kwargs-ok: ParamSpec forwarding keeps the wrapped sign_request signature
+) -> _SignedRequest:
+    if isinstance(provider_config, SignsRequestsWithAWS):
+        return await run_aws_signing(sign_request, *args, **kwargs)
+    return sign_request(*args, **kwargs)

litellm/llms/bedrock/chat/converse_handler.py

@@ -21,7 +21,7 @@ from litellm.rust_bridge.chat_completions import rust_chat_completions_accepts
 from litellm.types.utils import ModelResponse
 from litellm.utils import CustomStreamWrapper
 
-from ..base_aws_llm import BaseAWSLLM, Credentials, bedrock_bearer_token
+from ..base_aws_llm import BaseAWSLLM, Credentials, bedrock_bearer_token, run_aws_signing
 from ..common_utils import BedrockError, _get_all_bedrock_regions, error_response_text
 from .invoke_handler import AWSEventStreamDecoder, MockResponseIterator, make_call
 
@@ -136,7 +136,8 @@ class BedrockConverseLLM(BaseAWSLLM):
         )
         data: Final = json.dumps(request_data)
 
-        prepped: Final = self.get_request_headers(
+        prepped: Final = await run_aws_signing(
+            self.get_request_headers,
             credentials=credentials,
             aws_region_name=litellm_params.get("aws_region_name") or "us-west-2",
             extra_headers=headers,
@@ -206,7 +207,8 @@ class BedrockConverseLLM(BaseAWSLLM):
         )
         data: Final = json.dumps(request_data)
 
-        prepped: Final = self.get_request_headers(
+        prepped: Final = await run_aws_signing(
+            self.get_request_headers,
             credentials=credentials,
             aws_region_name=litellm_params.get("aws_region_name") or "us-west-2",
             extra_headers=headers,

litellm/llms/bedrock/chat/converse_transformation.py

@@ -4,6 +4,7 @@ Translating between OpenAI's `/chat/completion` format and Amazon's `/converse`
 
 import copy
 import json
+import re
 import time
 import types
 from collections.abc import Mapping
@@ -293,6 +294,10 @@ class AmazonConverseConfig(BaseConfig):
                     llm_provider="bedrock",
                 )
 
+    @staticmethod
+    def _is_openai_gpt_reasoning_model(model: str) -> bool:
+        return re.search(r"openai\.gpt-\d", model) is not None
+
     def _is_nova_2_model(self, model: str) -> bool:
         """
         Check if the model is a Nova 2 model that supports reasoningConfig.
@@ -422,15 +427,15 @@ class AmazonConverseConfig(BaseConfig):
         """
         Handle the reasoning_effort parameter based on the model type.
 
-        - GPT-OSS models: passed through unchanged via additionalModelRequestFields.
-        - OpenAI GPT-5.x models: mapped to ``reasoning.effort`` via additionalModelRequestFields.
+        - GPT-OSS and DeepSeek V3 models: passed through unchanged via additionalModelRequestFields.
+        - OpenAI GPT-5.x and GPT-6 models: mapped to ``reasoning.effort`` via additionalModelRequestFields.
         - Nova 2 models: transformed to reasoningConfig.
         - Anthropic models: mapped to ``thinking`` (and ``output_config.effort`` on
           adaptive Claude 4.6 / 4.7).
         """
-        if "gpt-oss" in model:
+        if "gpt-oss" in model or "deepseek" in model:
             optional_params["reasoning_effort"] = reasoning_effort
-        elif "openai.gpt-5" in model:
+        elif self._is_openai_gpt_reasoning_model(model):
             reasoning: Final[BedrockConverseGptReasoningEffortBlock] = {"effort": reasoning_effort}
             optional_params["reasoning"] = reasoning
         elif self._is_nova_2_model(model):
@@ -509,6 +514,36 @@ class AmazonConverseConfig(BaseConfig):
                 )
                 thinking["budget_tokens"] = BEDROCK_MIN_THINKING_BUDGET_TOKENS
 
+    def _is_deepseek_model(self, model: str, base_model: str) -> bool:
+        return "deepseek" in model or "deepseek" in base_model
+
+    def _is_deepseek_r1_model(self, model: str, base_model: str) -> bool:
+        return "deepseek.r1" in model or "deepseek.r1" in base_model
+
+    def _model_accepts_anthropic_thinking_param(self, model: str, base_model: str) -> bool:
+        """Whether the model accepts the Anthropic-shaped ``thinking`` request field.
+
+        Only Claude reasoning models accept it. DeepSeek advertises ``supports_reasoning`` but reasons
+        natively: R1 returns a 400 when the field is sent and V3 silently ignores it.
+        """
+        if self._is_deepseek_model(model=model, base_model=base_model):
+            return False
+        return (
+            "claude-3-7" in model
+            or "claude-sonnet-4" in model
+            or "claude-opus-4" in model
+            or supports_reasoning(model=model, custom_llm_provider=self.custom_llm_provider)
+            or supports_reasoning(model=base_model, custom_llm_provider=self.custom_llm_provider)
+        )
+
+    def _model_rejects_reasoning_effort_param(self, model: str, base_model: str) -> bool:
+        """Whether the model returns a 400 for every ``reasoning_effort`` shape on Converse.
+
+        DeepSeek R1 always reasons and rejects any reasoning request field. DeepSeek V3 accepts a raw
+        ``reasoning_effort`` like gpt-oss does, and every other model maps it to a shape it accepts.
+        """
+        return self._is_deepseek_r1_model(model=model, base_model=base_model)
+
     def get_supported_openai_params(self, model: str) -> list[str]:
         from litellm.utils import supports_function_calling
 
@@ -564,23 +599,20 @@ class AmazonConverseConfig(BaseConfig):
             # only anthropic and mistral support tool choice config. otherwise (E.g. cohere) will fail the call - https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ToolChoice.html
             supported_params.append("tool_choice")
 
-        if "gpt-oss" in model or "openai.gpt-5" in model or "openai.gpt-5" in base_model:
+        if (
+            "gpt-oss" in model
+            or self._is_openai_gpt_reasoning_model(model)
+            or self._is_openai_gpt_reasoning_model(base_model)
+        ):
             supported_params.append("reasoning_effort")
+        elif self._is_deepseek_model(model=model, base_model=base_model):
+            if not self._is_deepseek_r1_model(model=model, base_model=base_model):
+                supported_params.append("reasoning_effort")
         elif self._is_nova_2_model(model):
             # Nova 2 models support reasoning_effort (transformed to reasoningConfig)
             # These models use a different reasoning structure than Anthropic's thinking parameter
             supported_params.append("reasoning_effort")
-        elif (
-            "claude-3-7" in model
-            or "claude-sonnet-4" in model
-            or "claude-opus-4" in model
-            or "deepseek.r1" in model
-            or supports_reasoning(
-                model=model,
-                custom_llm_provider=self.custom_llm_provider,
-            )
-            or supports_reasoning(model=base_model, custom_llm_provider=self.custom_llm_provider)
-        ):
+        elif self._model_accepts_anthropic_thinking_param(model=model, base_model=base_model):
             supported_params.append("thinking")
             supported_params.append("reasoning_effort")
             supported_params.append("output_config")
@@ -872,6 +904,11 @@ class AmazonConverseConfig(BaseConfig):
         drop_params: bool,
     ) -> dict:
         is_thinking_enabled: Final = self.is_thinking_enabled(non_default_params)
+        base_model: Final = BedrockModelInfo.get_base_model(model)
+        drop_thinking_param: Final = self._is_deepseek_model(model=model, base_model=base_model)
+        drop_reasoning_effort_param: Final = self._model_rejects_reasoning_effort_param(
+            model=model, base_model=base_model
+        )
 
         for param, value in non_default_params.items():
             if param == "response_format" and isinstance(value, dict):
@@ -920,7 +957,12 @@ class AmazonConverseConfig(BaseConfig):
                 optional_params["_parallel_tool_use_config"] = {
                     "tool_choice": {"type": "auto", "disable_parallel_tool_use": not value}
                 }
-            if param == "thinking" and "openai.gpt-5" not in model:
+            if param == "thinking" and drop_thinking_param:
+                verbose_logger.debug(
+                    "Dropping unsupported `thinking` param for Bedrock model=%s; it reasons natively.",
+                    model,
+                )
+            elif param == "thinking" and not self._is_openai_gpt_reasoning_model(model):
                 if (
                     isinstance(value, dict)
                     and value.get("type") == "adaptive"
@@ -946,6 +988,11 @@ class AmazonConverseConfig(BaseConfig):
                     AnthropicModelInfo.translate_legacy_thinking_for_adaptive_model(
                         model=model, optional_params=optional_params, custom_llm_provider="bedrock"
                     )
+            elif param == "reasoning_effort" and isinstance(value, str) and drop_reasoning_effort_param:
+                verbose_logger.debug(
+                    "Dropping unsupported `reasoning_effort` param for Bedrock model=%s; it always reasons and rejects it.",
+                    model,
+                )
             elif param == "reasoning_effort" and isinstance(value, str):
                 self._handle_reasoning_effort_parameter(
                     model=model, reasoning_effort=value, optional_params=optional_params
@@ -1805,6 +1852,7 @@ class AmazonConverseConfig(BaseConfig):
             data=request_data,
             messages=messages,
             encoding=encoding,
+            json_mode=json_mode,
         )
 
     def _transform_reasoning_content(self, reasoning_content_blocks: list[BedrockConverseReasoningContentBlock]) -> str:
@@ -2237,6 +2285,7 @@ class AmazonConverseConfig(BaseConfig):
         data: dict | str,
         messages: list,
         encoding,
+        json_mode: bool | None = None,
     ) -> ModelResponse:
         ## LOGGING
         if logging_obj is not None:
@@ -2247,7 +2296,9 @@ class AmazonConverseConfig(BaseConfig):
                 additional_args={"complete_input_dict": data},
             )
 
-        json_mode: Final[bool | None] = optional_params.get("json_mode", None)
+        resolved_json_mode: Final[bool | None] = (
+            json_mode if json_mode is not None else optional_params.get("json_mode", None)
+        )
         ## RESPONSE OBJECT
         try:
             completion_response: Final = ConverseResponseBlock(**response.json())
@@ -2339,7 +2390,7 @@ class AmazonConverseConfig(BaseConfig):
             chat_completion_message["thinking_blocks"] = self._transform_thinking_blocks(reasoningContentBlocks)
         chat_completion_message["content"] = content_str
         filtered_tools: Final = self._filter_json_mode_tools(
-            json_mode=json_mode,
+            json_mode=resolved_json_mode,
             tools=tools,
             chat_completion_message=chat_completion_message,
         )
@@ -2363,7 +2414,7 @@ class AmazonConverseConfig(BaseConfig):
         # When json_mode filtered out all synthetic tool calls the response
         # is plain content, not a pending tool invocation. Fix finish_reason
         # so callers (e.g. OpenAI SDK) don't misinterpret it.
-        if json_mode and not filtered_tools and tools:
+        if resolved_json_mode and not filtered_tools and tools:
             initial_finish_reason = "stop"
 
         (

litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py

@@ -340,6 +340,7 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
                     optional_params=optional_params,
                     litellm_params=litellm_params,
                     encoding=encoding,
+                    json_mode=json_mode,
                 )
             elif provider == "twelvelabs":
                 return litellm.AmazonTwelveLabsPegasusConfig().transform_response(

litellm/llms/bedrock/count_tokens/handler.py

@@ -10,9 +10,10 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_logger
+from litellm.llms.bedrock.base_aws_llm import run_aws_signing
 from litellm.llms.bedrock.common_utils import BedrockError
 from litellm.llms.bedrock.count_tokens.transformation import BedrockCountTokensConfig
-from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, get_async_httpx_client
 
 
 class BedrockCountTokensHandler(BedrockCountTokensConfig):
@@ -27,6 +28,7 @@ class BedrockCountTokensHandler(BedrockCountTokensConfig):
         request_data: dict[str, Any],
         litellm_params: dict[str, Any],
         resolved_model: str,
+        client: AsyncHTTPHandler | None = None,
     ) -> dict[str, Any]:
         """
         Handle a CountTokens request using existing LiteLLM patterns.
@@ -75,7 +77,8 @@ class BedrockCountTokensHandler(BedrockCountTokensConfig):
             # Extract api_key for bearer token auth if provided
             api_key: Final = litellm_params.get("api_key", None)
             headers: Final = {"Content-Type": "application/json"}
-            signed_headers, signed_body = self._sign_request(
+            signed_headers, signed_body = await run_aws_signing(
+                self._sign_request,
                 service_name="bedrock",
                 headers=headers,
                 optional_params=litellm_params,
@@ -85,7 +88,7 @@ class BedrockCountTokensHandler(BedrockCountTokensConfig):
                 api_key=api_key,
             )
 
-            async_client: Final = get_async_httpx_client(llm_provider=litellm.LlmProviders.BEDROCK)
+            async_client: Final = client or get_async_httpx_client(llm_provider=litellm.LlmProviders.BEDROCK)
 
             response: Final = await async_client.post(
                 endpoint_url,

litellm/llms/bedrock/embed/embedding.py

@@ -5,7 +5,7 @@ Handles embedding calls to Bedrock's `/invoke` endpoint
 import copy
 import json
 import urllib.parse
-from collections.abc import Callable
+from collections.abc import Callable, Mapping
 from typing import TYPE_CHECKING, Final, get_args, overload
 
 import httpx
@@ -26,7 +26,7 @@ from litellm.types.llms.bedrock import (
 )
 from litellm.types.utils import EmbeddingResponse, LlmProviders
 
-from ..base_aws_llm import BaseAWSLLM, Credentials, bedrock_bearer_token
+from ..base_aws_llm import AWSPreparedRequest, BaseAWSLLM, Credentials, bedrock_bearer_token, run_aws_signing
 from ..common_utils import BedrockError
 from .amazon_nova_transformation import AmazonNovaEmbeddingConfig
 from .amazon_titan_g1_transformation import AmazonTitanG1Config
@@ -41,6 +41,20 @@ if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 
 
+def _sign_get_request(
+    credentials: Credentials, url: str, headers: Mapping[str, str], aws_region_name: str
+) -> AWSPreparedRequest:
+    try:
+        from botocore.auth import SigV4Auth
+        from botocore.awsrequest import AWSRequest
+    except ImportError:
+        raise ImportError("Missing boto3 to call bedrock. Run 'pip install boto3'.")
+
+    request: Final = AWSRequest(method="GET", url=url, data=None, headers=headers)
+    SigV4Auth(credentials, "bedrock", aws_region_name).add_auth(request)
+    return request.prepare()
+
+
 class BedrockEmbedding(BaseAWSLLM):
     @overload
     def _load_credentials(
@@ -342,7 +356,8 @@ class BedrockEmbedding(BaseAWSLLM):
             if extra_headers is not None:
                 headers = {"Content-Type": "application/json", **extra_headers}
 
-            prepped = self.get_request_headers(
+            prepped = await run_aws_signing(
+                self.get_request_headers,
                 credentials=credentials,
                 aws_region_name=aws_region_name,
                 extra_headers=extra_headers,
@@ -600,9 +615,6 @@ class BedrockEmbedding(BaseAWSLLM):
             dict: Status response from AWS Bedrock
         """
 
-        # Get AWS credentials using the same method as other Bedrock methods
-        credentials, _ = self._load_credentials(kwargs)
-
         # Get the runtime endpoint
         endpoint_url, _ = self.get_runtime_endpoint(
             api_base=None,
@@ -619,27 +631,13 @@ class BedrockEmbedding(BaseAWSLLM):
         # Prepare headers for GET request
         headers: Final = {"Content-Type": "application/json"}
 
-        # Use AWSRequest directly for GET requests (get_request_headers hardcodes POST)
-        try:
-            from botocore.auth import SigV4Auth
-            from botocore.awsrequest import AWSRequest
-        except ImportError:
-            raise ImportError("Missing boto3 to call bedrock. Run 'pip install boto3'.")
+        def sign_status_request() -> AWSPreparedRequest:
+            credentials, _ = self._load_credentials(kwargs)
+            return _sign_get_request(
+                credentials=credentials, url=status_url, headers=headers, aws_region_name=aws_region_name
+            )
 
-        # Create AWSRequest with GET method and encoded URL
-        request: Final = AWSRequest(
-            method="GET",
-            url=status_url,
-            data=None,  # GET request, no body
-            headers=headers,
-        )
-
-        # Sign the request - SigV4Auth will create canonical string from request URL
-        sigv4: Final = SigV4Auth(credentials, "bedrock", aws_region_name)
-        sigv4.add_auth(request)
-
-        # Prepare the request
-        prepped: Final = request.prepare()
+        prepped: Final = await run_aws_signing(sign_status_request)
 
         # LOGGING
         if logging_obj is not None:

litellm/llms/bedrock/files/transformation.py

@@ -7,13 +7,13 @@ from contextlib import suppress
 from functools import cache
 from itertools import chain
 from types import MappingProxyType
-from typing import Any, Final, TypeAlias, TypedDict
+from typing import Any, Final, Literal, TypeAlias, TypedDict
 from urllib.parse import unquote
 
 import httpx
 from httpx import Headers, Response
 from openai.types.file_deleted import FileDeleted
-from pydantic import BaseModel, ConfigDict, TypeAdapter
+from pydantic import BaseModel, ConfigDict, Field, TypeAdapter
 from typing_extensions import ReadOnly
 
 from litellm._logging import verbose_logger
@@ -60,11 +60,12 @@ from litellm.utils import get_llm_provider
 from ..base_aws_llm import BaseAWSLLM
 from ..common_utils import BedrockError, merge_bedrock_aws_request_params, resolve_s3_encryption_key_id
 
-# litellm_params key used to hand the SigV4-signed GET headers from
-# `transform_file_content_request` to `validate_environment` (the only hook
-# the shared file-content HTTP handler exposes for setting request headers).
-# Same pattern as the `upload_url` handoff in `transform_create_file_request`.
-S3_SIGNED_GET_HEADERS_PARAM: Final = "_s3_signed_get_headers"
+S3_SIGNED_REQUEST_HEADERS_PARAM: Final = "_s3_signed_request_headers"
+
+
+class _S3DeleteContext(BaseModel):
+    file_id: str = Field(min_length=1)
+
 
 # litellm_params key carrying the size of the body uploaded to S3, handed from
 # `transform_create_file_request` to `transform_create_file_response`.
@@ -291,7 +292,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
     ) -> dict:
         result: Final[dict[str, object]] = {}
         result.update(headers)
-        signed_headers: Final = litellm_params.pop(S3_SIGNED_GET_HEADERS_PARAM, None)
+        signed_headers: Final = litellm_params.pop(S3_SIGNED_REQUEST_HEADERS_PARAM, None)
         if isinstance(signed_headers, Mapping):
             result.update(signed_headers)  # any-ok: untyped handoff headers
         # otherwise no extra headers - AWS credentials are handled by BaseAWSLLM
@@ -1187,18 +1188,27 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
     def transform_delete_file_request(
         self,
         file_id: str,
-        optional_params: dict,
-        litellm_params: dict,
-    ) -> tuple[str, dict]:
-        raise NotImplementedError("BedrockFilesConfig does not support file deletion")
+        optional_params: Mapping[str, object],
+        litellm_params: MutableMapping[str, object],
+    ) -> tuple[str, dict[str, str]]:
+        return self._transform_s3_file_request(
+            file_id=file_id, method="DELETE", optional_params=optional_params, litellm_params=litellm_params
+        )
 
     def transform_delete_file_response(
         self,
         raw_response: httpx.Response,
         logging_obj: LiteLLMLoggingObj,
-        litellm_params: dict,
+        litellm_params: Mapping[str, object],
     ) -> FileDeleted:
-        raise NotImplementedError("BedrockFilesConfig does not support file deletion")
+        if raw_response.status_code != 204:
+            raise BedrockError(
+                status_code=raw_response.status_code if raw_response.status_code >= 400 else 502,
+                message=raw_response.text or f"S3 file deletion returned HTTP {raw_response.status_code}",
+                headers=raw_response.headers,
+            )
+        context: Final = _S3DeleteContext.model_validate(logging_obj.model_call_details.get("additional_args"))
+        return FileDeleted(id=context.file_id, deleted=True, object="file")
 
     def transform_list_files_request(
         self,
@@ -1233,6 +1243,18 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
         if not file_id:
             raise ValueError("file_id is required for Bedrock file content retrieval")
 
+        return self._transform_s3_file_request(
+            file_id=file_id, method="GET", optional_params=optional_params, litellm_params=litellm_params
+        )
+
+    def _transform_s3_file_request(
+        self,
+        *,
+        file_id: str,
+        method: Literal["GET", "DELETE"],
+        optional_params: Mapping[str, object],
+        litellm_params: MutableMapping[str, object],
+    ) -> tuple[str, dict[str, str]]:
         s3_uri: Final = extract_s3_uri_from_file_id(file_id)
         bucket_name, object_key = _validate_file_id_against_configured_buckets(
             s3_uri=s3_uri,
@@ -1240,40 +1262,32 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
             allow_legacy_cloud_file_ids=should_allow_legacy_cloud_file_ids(litellm_params),
         )
 
-        # The shared file-content handler passes optional_params={}, so AWS
-        # credentials/region arrive via litellm_params here (unlike the upload
-        # path). s3_region_name wins over aws_region_name, same priority as
-        # get_complete_file_url above.
-        merged_params: Final[dict[str, object]] = {}
-        merged_params.update(litellm_params)
-        merged_params.update(optional_params)
-        request_params: Final = _BedrockS3RequestParams.model_validate(merged_params)
+        request_params: Final = _BedrockS3RequestParams.model_validate({**litellm_params, **optional_params})
 
         region_preference: Final = request_params.s3_region_name or request_params.aws_region_name
         region_params: Final[dict[str, str | None]] = {"aws_region_name": region_preference}
         aws_region_name: Final = self._get_aws_region_name(optional_params=region_params, model="")
 
-        s3_endpoint_url = (
+        s3_endpoint_url: Final = (
             request_params.s3_endpoint_url or f"https://s3.{aws_region_name}.{get_aws_dns_suffix(aws_region_name)}"
         ).rstrip("/")
         url: Final = f"{s3_endpoint_url}/{bucket_name}/{encode_s3_object_key_for_url(object_key)}"
 
-        litellm_params[S3_SIGNED_GET_HEADERS_PARAM] = self._sign_s3_get_request(
+        litellm_params[S3_SIGNED_REQUEST_HEADERS_PARAM] = self._sign_s3_request_without_body(
             api_base=url,
             aws_region_name=aws_region_name,
             request_params=request_params,
+            method=method,
         )
         return url, {}
 
-    def _sign_s3_get_request(
+    def _sign_s3_request_without_body(
         self,
         api_base: str,
         aws_region_name: str,
         request_params: _BedrockS3RequestParams,
+        method: Literal["GET", "DELETE"] = "GET",
     ) -> dict[str, str]:
-        """
-        SigV4-sign an S3 GetObject request, mirroring `_sign_s3_request` (PUT).
-        """
         try:
             import hashlib
 
@@ -1297,7 +1311,7 @@ class BedrockFilesConfig(BaseAWSLLM, BaseFilesConfig):
 
         empty_body_hash: Final = hashlib.sha256(b"").hexdigest()
         aws_request: Final = AWSRequest(  # any-ok: botocore AWSRequest is untyped
-            method="GET",
+            method=method,
             url=api_base,
             headers={"x-amz-content-sha256": empty_body_hash},
         )

litellm/llms/bedrock/realtime/handler.py

@@ -21,7 +21,7 @@ from litellm.litellm_core_utils.realtime_streaming import DefaultLoggedRealTimeE
 from litellm.types.llms.openai import OpenAIRealtimeEvents
 from litellm.types.realtime import RealtimeResponseTransformInput
 
-from ..base_aws_llm import BaseAWSLLM
+from ..base_aws_llm import BaseAWSLLM, run_aws_signing
 from ..common_utils import BedrockError
 from .transformation import BedrockRealtimeConfig
 
@@ -149,7 +149,8 @@ class BedrockRealtime(BaseAWSLLM):
 
         verbose_proxy_logger.debug("Bedrock Realtime: Connecting to %s with model %s", endpoint_uri, model)
 
-        credentials: Final = self.get_credentials(
+        credentials: Final = await run_aws_signing(
+            self.get_credentials,
             aws_access_key_id=aws_access_key_id,
             aws_secret_access_key=aws_secret_access_key,
             aws_session_token=aws_session_token,
@@ -169,7 +170,7 @@ class BedrockRealtime(BaseAWSLLM):
                     "or configure credentials in the environment"
                 ),
             )
-        frozen_credentials: Final = credentials.get_frozen_credentials()
+        frozen_credentials: Final = await run_aws_signing(credentials.get_frozen_credentials)
 
         # Initialize Bedrock client with aws_sdk_bedrock_runtime
         config: Final = Config(

litellm/llms/bedrock_mantle/common_utils.py

@@ -23,7 +23,7 @@ from botocore.exceptions import (
     ProfileNotFound,
 )
 
-from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
+from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, SignsRequestsWithAWS
 from litellm.secret_managers.main import get_secret_str
 
 BEDROCK_MANTLE_DEFAULT_REGION: Final = "us-east-1"
@@ -55,7 +55,7 @@ def resolve_mantle_region(params: Mapping[str, object]) -> str:
     )
 
 
-class BedrockMantleAuthMixin:
+class BedrockMantleAuthMixin(SignsRequestsWithAWS):
     _aws_signer: BaseAWSLLM
 
     @staticmethod

litellm/llms/custom_httpx/llm_http_handler.py

@@ -77,6 +77,7 @@ from litellm.llms.base_llm.vector_store_files.transformation import (
     BaseVectorStoreFilesConfig,
 )
 from litellm.llms.base_llm.videos.transformation import BaseVideoConfig
+from litellm.llms.bedrock.base_aws_llm import SignsRequestsWithAWS, run_aws_signing, sign_request_off_loop_if_aws
 from litellm.llms.custom_httpx.container_handler import raise_for_error_status
 from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
@@ -579,7 +580,7 @@ class BaseLLMHTTPHandler:
             data: dict[str, object],  # mutable-ok: async_completion takes dict
             signed_headers: dict[str, object],  # mutable-ok: async_completion takes dict
             signed_json_body: bytes | None,
-        ):
+        ) -> Coroutine[object, object, ModelResponse | CustomStreamWrapper]:
             async_client: Final = client if isinstance(client, AsyncHTTPHandler) else None
             if stream is True:
                 return self.acompletion_stream_function(
@@ -626,7 +627,7 @@ class BaseLLMHTTPHandler:
 
         if acompletion is True and provider_config.uses_async_transform_request:
 
-            async def transform_then_dispatch():
+            async def transform_then_dispatch() -> ModelResponse | CustomStreamWrapper:
                 transformed: Final = cast(  # cast-ok: async_transform_request is declared as a bare dict
                     "dict[str, object]",
                     await provider_config.async_transform_request(
@@ -637,7 +638,12 @@ class BaseLLMHTTPHandler:
                         headers=request_headers,
                     ),
                 )
-                return await dispatch_async(*await asyncio.to_thread(sign_and_log, transformed))
+                signed_request: Final = await (
+                    run_aws_signing(sign_and_log, transformed)
+                    if isinstance(provider_config, SignsRequestsWithAWS)
+                    else asyncio.to_thread(sign_and_log, transformed)
+                )
+                return await dispatch_async(*signed_request)
 
             return transform_then_dispatch()
 
@@ -1973,7 +1979,9 @@ class BaseLLMHTTPHandler:
             api_key=api_key,
         )
 
-        signed_headers, signed_json_body = provider_config.sign_request(
+        signed_headers, signed_json_body = await sign_request_off_loop_if_aws(
+            provider_config,
+            provider_config.sign_request,
             headers=headers,
             optional_params=optional_params,
             request_data=data,
@@ -2074,7 +2082,9 @@ class BaseLLMHTTPHandler:
                         max_attempts,
                     )
                     provider_config.transform_anthropic_messages_request_on_http_error(e=e, request_data=request_body)
-                    headers, signed_json_body = provider_config.sign_request(
+                    headers, signed_json_body = await sign_request_off_loop_if_aws(
+                        provider_config,
+                        provider_config.sign_request,
                         headers=headers,
                         optional_params=optional_params_dict,
                         request_data=request_body,
@@ -2234,7 +2244,9 @@ class BaseLLMHTTPHandler:
             stream=stream,
         )
 
-        headers, signed_json_body = anthropic_messages_provider_config.sign_request(
+        headers, signed_json_body = await sign_request_off_loop_if_aws(
+            anthropic_messages_provider_config,
+            anthropic_messages_provider_config.sign_request,
             headers=headers,
             optional_params=dict(litellm_params),  # dynamic aws_* params are passed under litellm_params
             request_data=request_body,
@@ -2910,7 +2922,9 @@ class BaseLLMHTTPHandler:
                 fake_stream=fake_stream,
             )
 
-        headers, signed_body = responses_api_provider_config.sign_request(
+        headers, signed_body = await sign_request_off_loop_if_aws(
+            responses_api_provider_config,
+            responses_api_provider_config.sign_request,
             headers=headers,
             optional_params=dict(litellm_params),
             request_data=data,
@@ -4618,7 +4632,9 @@ class BaseLLMHTTPHandler:
         )
         data = BaseResponsesAPIConfig.normalize_responses_api_request_dict(data)
 
-        headers, signed_body = responses_api_provider_config.sign_request(
+        headers, signed_body = await sign_request_off_loop_if_aws(
+            responses_api_provider_config,
+            responses_api_provider_config.sign_request,
             headers=headers,
             optional_params=dict(litellm_params),
             request_data=data,
@@ -9845,7 +9861,9 @@ class BaseLLMHTTPHandler:
             )
         all_optional_params: Final[dict[str, object]] = dict(litellm_params)
         all_optional_params.update(vector_store_search_optional_params or {})
-        headers, signed_json_body = vector_store_provider_config.sign_request(
+        headers, signed_json_body = await sign_request_off_loop_if_aws(
+            vector_store_provider_config,
+            vector_store_provider_config.sign_request,
             headers=headers,
             optional_params=all_optional_params,
             request_data=request_body,

litellm/llms/databricks/chat/transformation.py

@@ -15,6 +15,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo
     _should_convert_tool_call_to_json_mode,
 )
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
+    _extract_reasoning_content,  # pyright: ignore[reportPrivateUsage]  # same import as the OpenAI transformation
     strip_litellm_internal_message_fields,
     strip_name_from_message,
 )
@@ -23,7 +24,9 @@ from litellm.types.llms.anthropic import AllAnthropicToolsValues
 from litellm.types.llms.databricks import (
     AllDatabricksContentValues,
     DatabricksChoice,
+    DatabricksDelta,
     DatabricksFunction,
+    DatabricksMessage,
     DatabricksResponse,
     DatabricksTool,
 )
@@ -247,8 +250,10 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
         litellm_params: dict,
         stream: bool | None = None,
     ) -> str:
-        api_base = self._get_api_base(api_base)
-        complete_url: Final = f"{api_base}/chat/completions"
+        use_ai_gateway: Final = model.removeprefix("databricks/").count(".") >= 2
+        api_base = self._get_api_base(api_base, use_ai_gateway=use_ai_gateway)
+        url_base: Final = api_base.rstrip("/") if use_ai_gateway else api_base
+        complete_url: Final = f"{url_base}/chat/completions"
         return complete_url
 
     def get_supported_openai_params(self, model: str | None = None) -> list:
@@ -534,6 +539,19 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
                             thinking_blocks.append(thinking_block)
         return reasoning_content, thinking_blocks
 
+    @staticmethod
+    def extract_top_level_reasoning_content(delta: DatabricksDelta) -> str | None:
+        return delta.get("reasoning_content")
+
+    @staticmethod
+    def resolve_reasoning_and_content(
+        message: DatabricksMessage, block_reasoning_content: str | None
+    ) -> tuple[str | None, str | None]:
+        content_str: Final = DatabricksConfig.extract_content_str(message["content"])
+        if block_reasoning_content is not None:
+            return block_reasoning_content, content_str
+        return _extract_reasoning_content({**message, "content": content_str})
+
     @staticmethod
     def extract_citations(
         content: AllDatabricksContentValues | None,
@@ -577,14 +595,13 @@ class DatabricksConfig(DatabricksBase, OpenAILikeChatConfig, AnthropicConfig):
                     finish_reason = "stop"
 
             if translated_message is None:
-                ## get the content str
-                content_str = DatabricksConfig.extract_content_str(choice["message"]["content"])
-
-                ## get the reasoning content
                 (
-                    reasoning_content,
+                    block_reasoning_content,
                     thinking_blocks,
                 ) = DatabricksConfig.extract_reasoning_content(choice["message"].get("content"))
+                reasoning_content, content_str = DatabricksConfig.resolve_reasoning_and_content(
+                    choice["message"], block_reasoning_content
+                )
 
                 citations = DatabricksConfig.extract_citations(choice["message"].get("content"))
 
@@ -738,12 +755,16 @@ class DatabricksChatResponseIterator(BaseModelResponseIterator):
 
                 # extract the reasoning content
                 (
-                    reasoning_content,
+                    block_reasoning_content,
                     thinking_blocks,
                 ) = DatabricksConfig.extract_reasoning_content(choice["delta"].get("content"))
 
                 choice["delta"]["content"] = content_str
-                choice["delta"]["reasoning_content"] = reasoning_content
+                choice["delta"]["reasoning_content"] = (
+                    block_reasoning_content
+                    if block_reasoning_content is not None
+                    else DatabricksConfig.extract_top_level_reasoning_content(choice["delta"])
+                )
                 choice["delta"]["thinking_blocks"] = thinking_blocks
                 translated_choices.append(choice)
             return ModelResponseStream(

litellm/llms/databricks/common_utils.py

@@ -177,19 +177,13 @@ class DatabricksBase:
         # Default: just litellm
         return f"litellm/{version}"
 
-    def _get_api_base(self, api_base: str | None) -> str:
-        """
-        Get the Databricks API base URL.
-
-        If not provided, attempts to get it from the Databricks SDK.
-        """
+    def _get_api_base(self, api_base: str | None, use_ai_gateway: bool = False) -> str:
         if api_base is None:
             try:
                 from databricks.sdk import WorkspaceClient
 
                 databricks_client: Final = WorkspaceClient()
                 api_base = f"{databricks_client.config.host}/serving-endpoints"
-                return api_base
             except ImportError:
                 raise DatabricksException(
                     status_code=400,
@@ -198,6 +192,18 @@ class DatabricksBase:
                         "or install the databricks-sdk Python library."
                     ),
                 )
+
+        if not use_ai_gateway:
+            return api_base
+
+        normalized_api_base: Final = api_base.rstrip("/")
+        if normalized_api_base.endswith("/ai-gateway/mlflow/v1"):
+            return normalized_api_base
+        if normalized_api_base.endswith("/serving-endpoints"):
+            return f"{normalized_api_base.removesuffix('/serving-endpoints')}/ai-gateway/mlflow/v1"
+        api_base_parts: Final = urlsplit(normalized_api_base)
+        if api_base_parts.path in ("", "/"):
+            return f"{normalized_api_base}/ai-gateway/mlflow/v1"
         return api_base
 
     def _get_oauth_m2m_token(

litellm/llms/hosted_vllm/image_edit/__init__.py

@@ -0,0 +1,9 @@
+from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
+
+from .transformation import HostedVLLMImageEditConfig
+
+__all__ = ("HostedVLLMImageEditConfig",)
+
+
+def get_hosted_vllm_image_edit_config(model: str) -> BaseImageEditConfig:
+    return HostedVLLMImageEditConfig()

litellm/llms/hosted_vllm/image_edit/transformation.py

@@ -0,0 +1,43 @@
+from typing import Final
+
+from litellm.llms.openai.image_edit.transformation import OpenAIImageEditConfig
+from litellm.secret_managers.main import get_secret_str
+
+PARAMS_VLLM_OMNI_DOES_NOT_ACCEPT: Final = frozenset({"mask", "quality", "input_fidelity"})
+
+
+class HostedVLLMImageEditConfig(OpenAIImageEditConfig):
+    def get_supported_openai_params(self, model: str) -> list:  # mutable-ok: BaseImageEditConfig contract
+        return [  # mutable-ok: BaseImageEditConfig returns list
+            param
+            for param in super().get_supported_openai_params(model)
+            if param not in PARAMS_VLLM_OMNI_DOES_NOT_ACCEPT
+        ]
+
+    def validate_environment(
+        self,
+        headers: dict,  # mutable-ok: BaseImageEditConfig contract
+        model: str,
+        api_key: str | None = None,
+        litellm_params: dict | None = None,  # mutable-ok: BaseImageEditConfig contract
+        api_base: str | None = None,
+    ) -> dict:  # mutable-ok: BaseImageEditConfig contract
+        resolved_key: Final = api_key or get_secret_str("HOSTED_VLLM_API_KEY") or "fake-api-key"
+        return {**headers, "Authorization": f"Bearer {resolved_key}"}  # mutable-ok: httpx headers are a dict
+
+    def get_complete_url(
+        self,
+        model: str,
+        api_base: str | None,
+        litellm_params: dict,  # mutable-ok: BaseImageEditConfig contract
+    ) -> str:
+        resolved_api_base: Final = api_base or get_secret_str("HOSTED_VLLM_API_BASE")
+        if resolved_api_base is None:
+            raise ValueError(
+                "api_base not set for Hosted VLLM images edits API. "
+                "Set via api_base parameter or HOSTED_VLLM_API_BASE environment variable"
+            )
+        trimmed: Final = resolved_api_base.rstrip("/")
+        if trimmed.endswith("/v1"):
+            return f"{trimmed}/images/edits"
+        return f"{trimmed}/v1/images/edits"

litellm/llms/openai/chat/guardrail_translation/handler.py

@@ -49,6 +49,8 @@ from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import
     coerce_stream_holdback_value,
 )
 from litellm.types.utils import (
+    ChatCompletionDeltaToolCall,
+    ChatCompletionMessageToolCall,
     Choices,
     GenericGuardrailAPIInputs,
     ModelResponse,
@@ -78,7 +80,8 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
     Methods can be overridden to customize behavior for different message formats.
     """
 
-    delivers_ended_stream_text_rewrites = True
+    delivers_ended_stream_rewrites = True
+    assembles_streamed_response = True
 
     def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None:
         """
@@ -610,13 +613,14 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
         deliver_ended_stream_rewrites: bool,
     ) -> None:
         """Ended-stream path: rebuild the full response, run the non-streaming
-        output guardrail against it, and (when opted in) write any text rewrite
-        back across the buffered chunks."""
+        output guardrail against it, and (when opted in) write any text or
+        tool-call rewrite back across the buffered chunks."""
         model_response: Final = cast(
             ModelResponse,
             stream_chunk_builder(chunks=responses_so_far, logging_obj=litellm_logging_obj),
         )
         pre_guardrail_texts: Final = self._string_choice_contents(model_response)
+        pre_guardrail_tool_calls: Final = self._function_tool_call_shapes(model_response)
         await self.process_output_response(
             response=model_response,
             guardrail_to_apply=guardrail_to_apply,
@@ -624,13 +628,21 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
             user_api_key_dict=user_api_key_dict,
             request_data=request_data,
         )
-        if deliver_ended_stream_rewrites:
-            await self._write_ended_stream_text_rewrites(
-                responses_so_far=responses_so_far,
-                guardrailed_response=model_response,
-                pre_guardrail_texts=pre_guardrail_texts,
-                guardrail_name=guardrail_to_apply.guardrail_name or "unknown",
-            )
+        if not deliver_ended_stream_rewrites:
+            return
+        guardrail_name: Final = guardrail_to_apply.guardrail_name or "unknown"
+        await self._write_ended_stream_text_rewrites(
+            responses_so_far=responses_so_far,
+            guardrailed_response=model_response,
+            pre_guardrail_texts=pre_guardrail_texts,
+            guardrail_name=guardrail_name,
+        )
+        self._write_ended_stream_tool_call_rewrites(
+            responses_so_far=responses_so_far,
+            guardrailed_response=model_response,
+            pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+            guardrail_name=guardrail_name,
+        )
 
     def build_stream_error_items(
         self,
@@ -1043,6 +1055,71 @@ class OpenAIChatCompletionsHandler(BaseTranslation):
             task_mappings=[(target_choice_index, None) for _ in changed],  # mutable-ok: callee takes lists
         )
 
+    @staticmethod
+    def _function_tool_call_shapes(response: "ModelResponse") -> tuple[tuple[str | None, str], ...]:
+        return tuple(
+            (tool_call.function.name, tool_call.function.arguments)
+            for choice in response.choices
+            for tool_call in choice.message.tool_calls or ()
+            if isinstance(tool_call, ChatCompletionMessageToolCall)
+        )
+
+    @staticmethod
+    def _function_tool_call_fragments(
+        responses_so_far: Sequence["ModelResponseStream"],
+    ) -> tuple[tuple[ChatCompletionDeltaToolCall, ...], ...]:
+        """Group the stream's function tool-call fragments by their tool-call index, in
+        the index order ``stream_chunk_builder`` lists the rebuilt tool calls, keeping
+        only the indices the builder keeps (an id and a name somewhere in the stream)."""
+        fragments: Final = tuple(
+            tool_call
+            for response in responses_so_far
+            for choice in response.choices
+            for tool_call in choice.delta.tool_calls or ()
+            if isinstance(tool_call, ChatCompletionDeltaToolCall)
+        )
+        identified: Final = frozenset(fragment.index for fragment in fragments if fragment.id)
+        named: Final = frozenset(fragment.index for fragment in fragments if fragment.function.name)
+        return tuple(
+            tuple(fragment for fragment in fragments if fragment.index == index) for index in sorted(identified & named)
+        )
+
+    def _write_ended_stream_tool_call_rewrites(
+        self,
+        responses_so_far: list["ModelResponseStream"],  # mutable-ok: rewrites the caller's buffered chunks in place
+        guardrailed_response: "ModelResponse",
+        pre_guardrail_tool_calls: tuple[tuple[str | None, str], ...],
+        guardrail_name: str,
+    ) -> None:
+        """Write ended-stream guardrail tool-call rewrites back across the buffered
+        chunks: the rewritten name and full arguments land in the tool call's first
+        fragment and the arguments of its later fragments are blanked, mirroring the
+        text write-back. A rewrite on a stream carrying more than one distinct choice
+        index, or whose fragments do not line up with the rebuilt tool calls, is
+        reported as undeliverable, so the pipeline executor discards it and releases
+        the original chunks."""
+        post_guardrail_tool_calls: Final = self._function_tool_call_shapes(guardrailed_response)
+        if post_guardrail_tool_calls == pre_guardrail_tool_calls:
+            return
+        stream_choice_indices: Final = frozenset(
+            choice.index for response in responses_so_far for choice in response.choices
+        )
+        fragments_by_tool_call: Final = self._function_tool_call_fragments(responses_so_far)
+        if len(stream_choice_indices) != 1 or len(fragments_by_tool_call) != len(post_guardrail_tool_calls):
+            from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+            raise UndeliverableStreamRewrite(guardrail_name)
+        for before, (name, arguments), fragments in zip(
+            pre_guardrail_tool_calls, post_guardrail_tool_calls, fragments_by_tool_call
+        ):
+            if (name, arguments) == before:
+                continue
+            head, *tail = fragments
+            head.function.name = name
+            head.function.arguments = arguments
+            for fragment in tail:
+                fragment.function.arguments = ""
+
     async def _apply_guardrail_responses_to_output_streaming(
         self,
         responses: list["ModelResponseStream"],

litellm/llms/openai/responses/guardrail_translation/handler.py

@@ -37,14 +37,14 @@ from itertools import accumulate, chain, repeat
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, NamedTuple, Union, cast
 
-from openai.types.responses.response_function_tool_call import ResponseFunctionToolCall
-from pydantic import BaseModel, TypeAdapter
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 from typing_extensions import ReadOnly, TypedDict
 
 from litellm._logging import verbose_proxy_logger
 from litellm.completion_extras.litellm_responses_transformation.transformation import (
     LiteLLMResponsesTransformationHandler,
     OpenAiResponsesToChatCompletionStreamIterator,
+    tool_call_dict_from_output_item,
 )
 from litellm.llms.base_llm.guardrail_translation.base_translation import (
     BaseTranslation,
@@ -84,7 +84,6 @@ from litellm.types.llms.openai import (
 )
 from litellm.types.responses.main import (
     GenericResponseOutputItem,
-    OutputFunctionToolCall,
     OutputText,
 )
 from litellm.types.utils import GenericGuardrailAPIInputs
@@ -101,6 +100,72 @@ if TYPE_CHECKING:
     from litellm.types.llms.openai import ResponseInputParam
 
 
+class _ToolCallShape(NamedTuple):
+    name: str | None
+    arguments: str
+
+
+class _ToolCallFunctionFields(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    name: str | None = None
+    arguments: str = ""
+
+
+class _ToolCallFields(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    function: _ToolCallFunctionFields
+
+
+def _tool_call_shapes(tool_calls: Sequence[ChatCompletionToolCallChunk]) -> tuple[_ToolCallShape, ...]:
+    return tuple(
+        _ToolCallShape(name=tool_call["function"].get("name"), arguments=tool_call["function"].get("arguments", ""))
+        for tool_call in tool_calls
+    )
+
+
+def _returned_tool_call_shape(tool_call: object) -> _ToolCallShape | None:
+    payload: Final = tool_call.model_dump() if isinstance(tool_call, BaseModel) else tool_call
+    try:
+        fields: Final = _ToolCallFields.model_validate(payload)
+    except ValidationError:
+        return None
+    return _ToolCallShape(name=fields.function.name, arguments=fields.function.arguments)
+
+
+def _post_guardrail_tool_call_shapes(
+    returned_tool_calls: Sequence[object] | None,
+    pre_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+    guardrail_name: str | None,
+) -> tuple[_ToolCallShape, ...]:
+    if not pre_guardrail_tool_calls:
+        return pre_guardrail_tool_calls
+    if returned_tool_calls is None or len(returned_tool_calls) != len(pre_guardrail_tool_calls):
+        verbose_proxy_logger.warning(
+            "OpenAI Responses API: guardrail %s returned %s tool calls for the %d scanned, "
+            "leaving the tool call output items unchanged",
+            guardrail_name,
+            "no" if returned_tool_calls is None else len(returned_tool_calls),
+            len(pre_guardrail_tool_calls),
+        )
+        return pre_guardrail_tool_calls
+    returned_shapes: Final = tuple(_returned_tool_call_shape(tool_call) for tool_call in returned_tool_calls)
+    validated_shapes: Final = tuple(shape for shape in returned_shapes if shape is not None)
+    if len(validated_shapes) != len(returned_shapes):
+        verbose_proxy_logger.warning(
+            "OpenAI Responses API: guardrail %s returned tool calls without a function name and arguments, "
+            "leaving the tool call output items unchanged",
+            guardrail_name,
+        )
+        return pre_guardrail_tool_calls
+    return validated_shapes
+
+
+def _tool_call_rewrite(before: _ToolCallShape, after: _ToolCallShape) -> _ToolCallShape:
+    return _ToolCallShape(name=after.name if after.name != before.name else None, arguments=after.arguments)
+
+
 class ResponseOutputEnvelope(TypedDict, total=False):
     """Dict form of a Responses API response, as far as guardrail write-back reads it."""
 
@@ -128,6 +193,20 @@ _TERMINAL_ENVELOPE_EVENT_TYPES: Final = frozenset(
 )
 
 
+_TOOL_CALL_ITEM_TYPES: Final = frozenset({"function_call", "custom_tool_call"})
+_TOOL_CALL_PAYLOAD_FIELDS: Final[Mapping[str, str]] = MappingProxyType(
+    {"function_call": "arguments", "custom_tool_call": "input"}
+)
+_TOOL_CALL_PAYLOAD_DELTA_EVENT_TYPES: Final = frozenset(
+    {"response.function_call_arguments.delta", "response.custom_tool_call_input.delta"}
+)
+_TOOL_CALL_PAYLOAD_DONE_EVENT_FIELDS: Final[Mapping[str, str]] = MappingProxyType(
+    {"response.function_call_arguments.done": "arguments", "response.custom_tool_call_input.done": "input"}
+)
+_TOOL_CALL_PAYLOAD_EVENT_TYPES: Final = _TOOL_CALL_PAYLOAD_DELTA_EVENT_TYPES | frozenset(
+    _TOOL_CALL_PAYLOAD_DONE_EVENT_FIELDS
+)
+_OUTPUT_ITEM_EVENT_TYPES: Final = frozenset({"response.output_item.added", "response.output_item.done"})
 _PATCHABLE_ITEM_FIELDS: Final[Mapping[str, str]] = MappingProxyType(
     {"function_call_output": "output", "message": "content"}
 )
@@ -164,8 +243,20 @@ def _rewritten_input_item(item: Mapping[str, object], rewritten: object) -> Mapp
     return {**item, field: converted_value}  # mutable-ok: request input items must stay JSON-plain dicts
 
 
-def _is_function_call_item(item: object) -> bool:
-    return isinstance(item, Mapping) and item.get("type") in ("function_call", "custom_tool_call")
+def _is_tool_call_item(item: object) -> bool:
+    return isinstance(item, Mapping) and item.get("type") in _TOOL_CALL_ITEM_TYPES
+
+
+def _tool_call_output_item_mapping(item: object) -> Mapping[str, object] | None:
+    if stream_item_field(item, "type") not in _TOOL_CALL_ITEM_TYPES:
+        return None
+    if isinstance(item, Mapping):
+        return cast("Mapping[str, object]", item)  # cast-ok: output items are str-keyed JSON objects
+    return item.model_dump() if isinstance(item, BaseModel) else None
+
+
+def _is_tool_call_output_item(item: object) -> bool:
+    return _tool_call_output_item_mapping(item) is not None
 
 
 def _last_message_role(messages: Sequence[object]) -> str | None:
@@ -189,7 +280,7 @@ def _provenance_unit_bounds(
     start_indexes: Final = tuple(
         index
         for index in range(len(raw_input))
-        if index == 0 or not (_is_function_call_item(raw_input[index]) and trailing_roles[index - 1] == "assistant")
+        if index == 0 or not (_is_tool_call_item(raw_input[index]) and trailing_roles[index - 1] == "assistant")
     )
     return tuple(zip(start_indexes, (*start_indexes[1:], len(raw_input))))
 
@@ -340,7 +431,8 @@ class OpenAIResponsesHandler(BaseTranslation):
     Methods can be overridden to customize behavior for different message formats.
     """
 
-    delivers_ended_stream_text_rewrites = True
+    delivers_ended_stream_rewrites = True
+    assembles_streamed_response = True
 
     def get_structured_messages(self, data: dict) -> list[AllMessageValues] | None:
         """
@@ -587,7 +679,7 @@ class OpenAIResponsesHandler(BaseTranslation):
             - response.output is a list of output items
             - Each output item can be:
               * GenericResponseOutputItem with a content list of OutputText objects
-              * ResponseFunctionToolCall with tool call data
+              * ResponseFunctionToolCall or CustomToolCallOutputItem with tool call data
             - Each OutputText object has a text field
         """
 
@@ -652,6 +744,7 @@ class OpenAIResponsesHandler(BaseTranslation):
             if response_model:
                 inputs["model"] = response_model
 
+            pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_to_check)
             guardrailed_inputs: Final = await guardrail_to_apply.apply_guardrail(
                 inputs=inputs,
                 request_data=request_data,
@@ -660,6 +753,11 @@ class OpenAIResponsesHandler(BaseTranslation):
             )
 
             guardrailed_texts: Final = guardrailed_inputs.get("texts", [])
+            post_guardrail_tool_calls: Final = _post_guardrail_tool_call_shapes(
+                returned_tool_calls=guardrailed_inputs.get("tool_calls"),
+                pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+                guardrail_name=guardrail_to_apply.guardrail_name,
+            )
 
             # Step 3: Map guardrail responses back to original response structure
             await self._apply_guardrail_responses_to_output(
@@ -667,6 +765,11 @@ class OpenAIResponsesHandler(BaseTranslation):
                 responses=guardrailed_texts,
                 task_mappings=task_mappings,
             )
+            self._write_tool_call_rewrites_to_output(
+                tool_call_items=tuple(item for item in response_output if _is_tool_call_output_item(item)),
+                pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+                post_guardrail_tool_calls=post_guardrail_tool_calls,
+            )
 
         verbose_proxy_logger.debug("OpenAI Responses API: Processed output response: %s", response)
 
@@ -754,6 +857,7 @@ class OpenAIResponsesHandler(BaseTranslation):
                 if response_model:
                     inputs["model"] = response_model
 
+                pre_guardrail_tool_calls: Final = _tool_call_shapes(tool_calls_to_check)
                 guardrailed_inputs: Final = await guardrail_to_apply.apply_guardrail(
                     inputs=inputs,
                     request_data=request_data,
@@ -762,6 +866,11 @@ class OpenAIResponsesHandler(BaseTranslation):
                 )
 
                 guardrailed_texts: Final = guardrailed_inputs.get("texts", [])
+                post_guardrail_tool_calls: Final = _post_guardrail_tool_call_shapes(
+                    returned_tool_calls=guardrailed_inputs.get("tool_calls"),
+                    pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+                    guardrail_name=guardrail_to_apply.guardrail_name,
+                )
 
                 # Write guardrailed texts back into the output items in-place.
                 # final_chunk is a reference into responses_so_far so this
@@ -784,6 +893,13 @@ class OpenAIResponsesHandler(BaseTranslation):
                             stream_events=responses_so_far[:-1],
                             rewrites_by_position=rewrites_by_position,
                         )
+                    self._deliver_ended_stream_tool_call_rewrites(
+                        responses_so_far=responses_so_far,
+                        outputs=outputs,
+                        pre_guardrail_tool_calls=pre_guardrail_tool_calls,
+                        post_guardrail_tool_calls=post_guardrail_tool_calls,
+                        guardrail_name=guardrail_to_apply.guardrail_name or "unknown",
+                    )
                 return responses_so_far
 
         # ------------------------------------------------------------------ #
@@ -894,6 +1010,148 @@ class OpenAIResponsesHandler(BaseTranslation):
                 continue
             OpenAIResponsesHandler._write_event_field(content[content_idx], "text", rewritten)
 
+    def _deliver_ended_stream_tool_call_rewrites(
+        self,
+        responses_so_far: Sequence[object],
+        outputs: Sequence[object],
+        pre_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+        post_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+        guardrail_name: str,
+    ) -> None:
+        """Write ended-stream guardrail tool-call rewrites into the completed
+        envelope's ``function_call`` and ``custom_tool_call`` items and sync the
+        earlier stream events, keyed by ``call_id``. The guardrail sees the
+        envelope's tool calls in output order, which is how a rewritten call
+        finds its ``call_id``; the stream events find their call through the
+        ``call_id`` on ``output_item`` events and the ``item_id`` on argument
+        and custom-input events, since an
+        event's ``output_index`` need not match the envelope's (the chat bridge
+        numbers tool calls from 1 while the envelope lists them after the
+        message). A rewrite whose calls do not line up with the envelope, or
+        whose events cannot be found, is reported as undeliverable, so the
+        pipeline executor discards it and releases the original events."""
+        if post_guardrail_tool_calls == pre_guardrail_tool_calls:
+            return
+        tool_call_items: Final = tuple(output_item for output_item in outputs if _is_tool_call_output_item(output_item))
+        call_ids: Final = tuple(
+            call_id
+            for output_item in tool_call_items
+            if isinstance(call_id := stream_item_field(output_item, "call_id"), str) and call_id
+        )
+        stream_events: Final = responses_so_far[:-1]
+        call_id_by_item_id: Final = self._tool_call_ids_by_item_id(stream_events)
+        event_call_ids: Final = tuple(
+            self._tool_call_event_call_id(event, call_id_by_item_id) for event in stream_events
+        )
+        rewrites_by_call_id: Final = MappingProxyType(
+            {
+                call_id: _tool_call_rewrite(before, after)
+                for call_id, before, after in zip(call_ids, pre_guardrail_tool_calls, post_guardrail_tool_calls)
+                if after != before
+            }
+        )
+        unresolved_argument_event: Final = any(
+            call_id is None and stream_item_field(event, "type") in _TOOL_CALL_PAYLOAD_EVENT_TYPES
+            for event, call_id in zip(stream_events, event_call_ids)
+        )
+        if (
+            len(call_ids) != len(tool_call_items)
+            or len(frozenset(call_ids)) != len(call_ids)
+            or len(call_ids) != len(post_guardrail_tool_calls)
+            or unresolved_argument_event
+            or not rewrites_by_call_id.keys() <= frozenset(event_call_ids)
+        ):
+            from litellm.proxy.policy_engine.pipeline_executor import UndeliverableStreamRewrite
+
+            raise UndeliverableStreamRewrite(guardrail_name)
+        for output_item, rewrite in (
+            (output_item, rewrites_by_call_id[call_id])
+            for output_item, call_id in zip(tool_call_items, call_ids)
+            if call_id in rewrites_by_call_id
+        ):
+            self._write_tool_call_item(output_item, rewrite.name, rewrite.arguments)
+        delta_replacements: Final = MappingProxyType(
+            {call_id: chain((rewrite.arguments,), repeat("")) for call_id, rewrite in rewrites_by_call_id.items()}
+        )
+        for event, call_id in zip(stream_events, event_call_ids):
+            if call_id not in rewrites_by_call_id:
+                continue
+            match stream_item_field(event, "type"):
+                case str() as event_type if event_type in _TOOL_CALL_PAYLOAD_DELTA_EVENT_TYPES:
+                    self._write_event_field(event, "delta", next(delta_replacements[call_id]))
+                case str() as event_type if event_type in _TOOL_CALL_PAYLOAD_DONE_EVENT_FIELDS:
+                    self._write_event_field(
+                        event, _TOOL_CALL_PAYLOAD_DONE_EVENT_FIELDS[event_type], rewrites_by_call_id[call_id].arguments
+                    )
+                case "response.output_item.added":
+                    self._write_tool_call_item(
+                        stream_item_field(event, "item"), rewrites_by_call_id[call_id].name, None
+                    )
+                case "response.output_item.done":
+                    self._write_tool_call_item(
+                        stream_item_field(event, "item"),
+                        rewrites_by_call_id[call_id].name,
+                        rewrites_by_call_id[call_id].arguments,
+                    )
+                case _:
+                    pass
+
+    def _write_tool_call_rewrites_to_output(
+        self,
+        tool_call_items: Sequence[object],
+        pre_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+        post_guardrail_tool_calls: tuple[_ToolCallShape, ...],
+    ) -> None:
+        if len(tool_call_items) != len(post_guardrail_tool_calls):
+            return
+        for output_item, rewrite in (
+            (output_item, _tool_call_rewrite(before, after))
+            for output_item, before, after in zip(tool_call_items, pre_guardrail_tool_calls, post_guardrail_tool_calls)
+            if after != before
+        ):
+            self._write_tool_call_item(output_item, rewrite.name, rewrite.arguments)
+
+    @staticmethod
+    def _tool_call_ids_by_item_id(stream_events: Sequence[object]) -> Mapping[str, str]:
+        items: Final = tuple(
+            stream_item_field(event, "item")
+            for event in stream_events
+            if stream_item_field(event, "type") in _OUTPUT_ITEM_EVENT_TYPES
+        )
+        return MappingProxyType(
+            {
+                item_id: call_id
+                for item in items
+                if stream_item_field(item, "type") in _TOOL_CALL_ITEM_TYPES
+                and isinstance(item_id := stream_item_field(item, "id"), str)
+                and isinstance(call_id := stream_item_field(item, "call_id"), str)
+            }
+        )
+
+    @staticmethod
+    def _tool_call_event_call_id(event: object, call_id_by_item_id: Mapping[str, str]) -> str | None:
+        event_type: Final = stream_item_field(event, "type")
+        if event_type in _TOOL_CALL_PAYLOAD_EVENT_TYPES:
+            item_id: Final = stream_item_field(event, "item_id")
+            return call_id_by_item_id.get(item_id) if isinstance(item_id, str) else None
+        if event_type not in _OUTPUT_ITEM_EVENT_TYPES:
+            return None
+        item: Final = stream_item_field(event, "item")
+        call_id: Final = stream_item_field(item, "call_id")
+        return (
+            call_id if stream_item_field(item, "type") in _TOOL_CALL_ITEM_TYPES and isinstance(call_id, str) else None
+        )
+
+    @staticmethod
+    def _write_tool_call_item(item: object, name: str | None, payload: str | None) -> None:
+        if item is None:
+            return
+        if name is not None:
+            OpenAIResponsesHandler._write_event_field(item, "name", name)
+        item_type: Final = stream_item_field(item, "type")
+        if payload is not None and isinstance(item_type, str) and item_type in _TOOL_CALL_PAYLOAD_FIELDS:
+            OpenAIResponsesHandler._write_event_field(item, _TOOL_CALL_PAYLOAD_FIELDS[item_type], payload)
+
     def _check_streaming_has_ended(self, responses_so_far: Sequence[object]) -> bool:
         """
         Check if the streaming has ended.
@@ -920,7 +1178,7 @@ class OpenAIResponsesHandler(BaseTranslation):
     def _completed_response_scan_key(response: object) -> StreamingScanKey:
         output_items: Final = stream_item_items(response, "output")
         message_items: Final = tuple(
-            item for item in output_items if stream_item_field(item, "type") != "function_call"
+            item for item in output_items if stream_item_field(item, "type") not in _TOOL_CALL_ITEM_TYPES
         )
         return StreamingScanKey(
             texts=tuple(
@@ -932,7 +1190,7 @@ class OpenAIResponsesHandler(BaseTranslation):
             tool_calls=tuple(
                 stream_item_fingerprint(item)
                 for item in output_items
-                if stream_item_field(item, "type") == "function_call"
+                if stream_item_field(item, "type") in _TOOL_CALL_ITEM_TYPES
             ),
             stream_ended=True,
         )
@@ -1043,34 +1301,10 @@ class OpenAIResponsesHandler(BaseTranslation):
         Override this method to customize text/image/tool extraction logic.
         """
 
-        # Check if this is a tool call (OutputFunctionToolCall)
-        if isinstance(output_item, OutputFunctionToolCall) or (
-            isinstance(output_item, BaseModel)
-            and hasattr(output_item, "type")
-            and getattr(output_item, "type") == "function_call"
-        ):
+        tool_call_item: Final = _tool_call_output_item_mapping(output_item)
+        if tool_call_item is not None:
             if tool_calls_to_check is not None:
-                tool_call_dict = (
-                    LiteLLMCompletionResponsesConfig.convert_response_function_tool_call_to_chat_completion_tool_call(
-                        tool_call_item=output_item,
-                        index=output_idx,
-                    )
-                )
-                tool_calls_to_check.append(cast(ChatCompletionToolCallChunk, tool_call_dict))
-            return
-        elif isinstance(output_item, dict) and output_item.get("type") == "function_call":
-            # Handle dict representation of tool call
-            if tool_calls_to_check is not None:
-                # Convert dict to ResponseFunctionToolCall for processing
-                try:
-                    tool_call_obj: Final = ResponseFunctionToolCall(**output_item)
-                    tool_call_dict = LiteLLMCompletionResponsesConfig.convert_response_function_tool_call_to_chat_completion_tool_call(
-                        tool_call_item=tool_call_obj,
-                        index=output_idx,
-                    )
-                    tool_calls_to_check.append(cast(ChatCompletionToolCallChunk, tool_call_dict))
-                except Exception:
-                    pass
+                tool_calls_to_check.append(tool_call_dict_from_output_item(tool_call_item, output_idx))
             return
 
         # Handle both GenericResponseOutputItem and dict

litellm/main.py

@@ -5398,6 +5398,14 @@ def completion(
         if dynamic_api_key is not None:
             api_key = dynamic_api_key
         # check if user passed in any of the OpenAI optional params
+        bridges_to_responses_api: Final = (
+            responses_api_model_info.get("mode") == "responses" and not skip_responses_api_bridge
+        )
+        allowed_openai_params: Final[list[str] | None] = (
+            [*(kwargs.get("allowed_openai_params") or []), "reasoning_effort"]
+            if bridges_to_responses_api
+            else kwargs.get("allowed_openai_params")
+        )
         optional_param_args: Final = {
             "functions": functions,
             "function_call": function_call,
@@ -5442,7 +5450,7 @@ def completion(
             "service_tier": service_tier,
             "store": store,
             "prompt_cache_key": prompt_cache_key,
-            "allowed_openai_params": kwargs.get("allowed_openai_params"),
+            "allowed_openai_params": allowed_openai_params,
             "base_model": base_model,
         }
         optional_params = get_optional_params(**optional_param_args, **non_default_params)
@@ -7805,6 +7813,7 @@ def transcription(
             azure_ad_token=azure_ad_token,
             max_retries=max_retries,
             litellm_params=litellm_params_dict,
+            custom_llm_provider=custom_llm_provider,
         )
     elif custom_llm_provider == "openai" or (custom_llm_provider in litellm.openai_compatible_providers):
         api_base = (

litellm/proxy/_experimental/mcp_server/rest_endpoints.py

@@ -3,12 +3,15 @@ import importlib
 from collections.abc import Awaitable, Callable, Mapping, Sequence
 from dataclasses import dataclass
 from datetime import datetime
+from traceback import walk_tb
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal
+from uuid import uuid4
 
 import anyio
 import httpx
 from fastapi import APIRouter, Depends, HTTPException, Query, Request, status
+from pydantic import ValidationError
 from starlette.datastructures import Headers
 
 from litellm._logging import verbose_logger
@@ -30,6 +33,8 @@ from litellm.proxy._experimental.mcp_server.faults.list_outcomes import (
     list_fault_http_status,
     outcome_wire_value,
 )
+from litellm.proxy._experimental.mcp_server.faults.traversal import iter_exception_tree
+from litellm.proxy._experimental.mcp_server.oauth_utils import _redact_mcp_resource_url
 from litellm.proxy._experimental.mcp_server.ui_session_utils import (
     acting_user_auth,
     build_effective_auth_contexts,
@@ -78,11 +83,39 @@ _MCP_GUARDRAIL_REJECTIONS: Final = (
 
 
 def _connection_error_message(exc: BaseException, url: str | None, timeout_seconds: float) -> str:
+    reference: Final = uuid4().hex
+    verbose_logger.error(
+        "MCP connection test failed (reference=%s): %s",
+        reference,
+        tuple(
+            (
+                type(cause).__name__,
+                tuple(
+                    (frame.f_code.co_filename, lineno, frame.f_code.co_name)
+                    for frame, lineno in walk_tb(cause.__traceback__)
+                ),
+            )
+            for cause in iter_exception_tree(exc)
+        ),
+    )
+    return next(
+        (
+            message
+            for cause in iter_exception_tree(exc)
+            if (message := _known_connection_error_message(cause, url, timeout_seconds)) is not None
+        ),
+        "An unexpected error occurred while testing the MCP connection. "
+        f"Retry; if it persists, share reference {reference} with your gateway administrator.",
+    )
+
+
+def _known_connection_error_message(exc: BaseException, url: str | None, timeout_seconds: float) -> str | None:
     if isinstance(exc, MCPServerURLCredentialsError):
         return str(exc.detail)
     if isinstance(exc, TimeoutError):
         return (
-            f"Failed to connect to MCP server: no response from {url or 'the server'} "
+            "Failed to connect to MCP server: no valid MCP response received from "
+            f"{_redact_mcp_resource_url(url) or 'the server'} "
             f"within {timeout_seconds:.0f}s. Check that the LiteLLM proxy can reach this URL "
             "from its network (DNS, egress rules, firewalls) and that the server answers MCP requests."
         )
@@ -99,13 +132,45 @@ def _connection_error_message(exc: BaseException, url: str | None, timeout_secon
         return "Failed to connect to MCP server: the connection timed out."
     if isinstance(exc, httpx.HTTPStatusError):
         return f"Failed to connect to MCP server: it returned HTTP {exc.response.status_code}."
-    return "Failed to connect to MCP server. Check proxy logs for details."
+    if isinstance(exc, (httpx.NetworkError, httpx.RemoteProtocolError, ConnectionError)):
+        return (
+            "Failed to connect to MCP server: the connection was interrupted. "
+            "Check the server and network connection, then retry."
+        )
+    if isinstance(exc, ValueError) and str(exc).startswith("Unexpected content type:"):
+        return (
+            "Failed to connect to MCP server: the endpoint returned an unsupported content type. "
+            "Check that the URL is an MCP endpoint, not a web page, and matches the selected transport."
+        )
+    if isinstance(exc, ValidationError) and exc.title in ("JSONRPCMessage", "InitializeResult", "ListToolsResult"):
+        return (
+            "Failed to connect to MCP server: the endpoint returned invalid JSON or an invalid MCP response. "
+            "Check the MCP endpoint URL and the server's protocol implementation."
+        )
+    if MCP_AVAILABLE and isinstance(exc, McpError):
+        if exc.error.code == -32000 and exc.error.message == "Connection closed":
+            return (
+                "Failed to connect to MCP server: the connection was closed before the request completed. "
+                "Check that the server stays running and returns a complete MCP response, then retry."
+            )
+        if exc.error.code == 32600 and exc.error.message == "Session terminated":
+            return (
+                "Failed to connect to MCP server: the MCP session was terminated. "
+                "Check that the URL points to an MCP endpoint and matches the selected transport, "
+                "then retry to start a new session."
+            )
+        return (
+            f"Failed to connect to MCP server: the MCP request failed (JSON-RPC code {exc.error.code}). "
+            "Check that the endpoint supports MCP initialization and tool listing, and check the upstream server logs."
+        )
+    return None
 
 
 if MCP_AVAILABLE:
+    from mcp.shared.exceptions import McpError
     from mcp.types import Tool as MCPTool
 
-    from litellm.experimental_mcp_client.client import MCPClient
+    from litellm.experimental_mcp_client.client import MCPClient, as_mcp_read_timeout
     from litellm.llms.litellm_proxy.skills.skill_search import (
         DEFAULT_SKILL_SEARCH_TOP_K,
     )
@@ -1169,7 +1234,7 @@ if MCP_AVAILABLE:
         return client_id, client_secret, scopes
 
     _STAGED_AUTH_VALUE_AUTH_TYPES: Final = frozenset(
-        (MCPAuth.api_key, MCPAuth.bearer_token, MCPAuth.basic, MCPAuth.authorization)
+        (MCPAuth.api_key, MCPAuth.bearer_token, MCPAuth.basic, MCPAuth.authorization, MCPAuth.token)
     )
 
     @dataclass(frozen=True, slots=True)
@@ -1178,6 +1243,17 @@ if MCP_AVAILABLE:
         mcp_auth_header: str | None
         oauth2_headers: dict[str, str] | None
 
+    def _preview_origin(url: str | None) -> tuple[str, str, int | None] | None:
+        if not url:
+            return None
+        try:
+            parsed: Final = httpx.URL(url)
+        except httpx.InvalidURL:
+            return None
+        if parsed.scheme not in ("http", "https") or not parsed.host:
+            return None
+        return parsed.scheme, parsed.host, parsed.port
+
     def _stage_server_test(new_mcp_server_request: NewMCPServerRequest, headers: Headers) -> _StagedServerTest:
         """
         Resolve the credentials a not-yet-saved server config carries for a preview call.
@@ -1190,7 +1266,19 @@ if MCP_AVAILABLE:
             MCPRequestHandler,
         )
 
-        request: Final = _inherit_credentials_from_existing_server(new_mcp_server_request)
+        saved_server: Final = (
+            global_mcp_server_manager.get_mcp_server_by_id(new_mcp_server_request.server_id)
+            if new_mcp_server_request.server_id
+            else None
+        )
+        saved_origin: Final = _preview_origin(saved_server.url) if saved_server else None
+        preview_origin: Final = _preview_origin(new_mcp_server_request.url)
+        may_inherit: Final = new_mcp_server_request.auth_type not in _STAGED_AUTH_VALUE_AUTH_TYPES or (
+            saved_origin is not None and saved_origin == preview_origin
+        )
+        request: Final = (
+            _inherit_credentials_from_existing_server(new_mcp_server_request) if may_inherit else new_mcp_server_request
+        )
         mcp_auth_header: Final = (
             request.credentials.get("auth_value")
             if request.auth_type in _STAGED_AUTH_VALUE_AUTH_TYPES and isinstance(request.credentials, dict)
@@ -1253,8 +1341,15 @@ if MCP_AVAILABLE:
             if _oauth2_flow == "client_credentials" and not request.token_url:
                 _oauth2_flow = None
 
+            # Static previews inherit credentials before this step, but must not resolve back to
+            # the saved record during client creation and discard the edited connection settings.
+            preview_server_id: Final = (
+                ""
+                if request.auth_type in _STAGED_AUTH_VALUE_AUTH_TYPES or request.auth_type in (None, MCPAuth.none)
+                else request.server_id or ""
+            )
             server_model: Final = MCPServer(
-                server_id=request.server_id or "",
+                server_id=preview_server_id,
                 name=request.alias or request.server_name or "",
                 url=request.url,
                 transport=request.transport,
@@ -1342,11 +1437,18 @@ if MCP_AVAILABLE:
         except (KeyboardInterrupt, SystemExit, asyncio.CancelledError):
             raise
         except BaseException as e:
-            verbose_logger.error("Error in MCP operation: %s", e, exc_info=True)
+            effective_timeout: Final = (
+                min(request.timeout if request.timeout is not None else MCP_CLIENT_TIMEOUT, timeout_seconds)
+                if any(
+                    isinstance(cause, McpError) and as_mcp_read_timeout(cause) is not None
+                    for cause in iter_exception_tree(e)
+                )
+                else timeout_seconds
+            )
             return {
                 "status": "error",
                 "error": True,
-                "message": _connection_error_message(e, request.url, timeout_seconds),
+                "message": _connection_error_message(e, request.url, effective_timeout),
             }
 
     async def _preview_openapi_tools(spec_path: str) -> dict:

litellm/proxy/_experimental/mcp_server/server.py

@@ -767,12 +767,12 @@ if MCP_AVAILABLE:
                     _stateful_auth_context_cleanup_task.cancel()
                     with contextlib.suppress(asyncio.CancelledError):
                         await _stateful_auth_context_cleanup_task
-                if _session_manager_cm:
-                    await _session_manager_cm.__aexit__(None, None, None)
-                if _session_manager_stateful_cm:
-                    await _session_manager_stateful_cm.__aexit__(None, None, None)
                 if _sse_session_manager_cm:
                     await _sse_session_manager_cm.__aexit__(None, None, None)
+                if _session_manager_stateful_cm:
+                    await _session_manager_stateful_cm.__aexit__(None, None, None)
+                if _session_manager_cm:
+                    await _session_manager_cm.__aexit__(None, None, None)
             except Exception as e:
                 verbose_logger.exception("Error during session manager shutdown: %s", e)
 
@@ -1005,6 +1005,7 @@ if MCP_AVAILABLE:
 
         if _mcp_proxy_mode.get() and name in MCP_PROXY_TOOL_NAMES:
             assert user_api_key_auth is not None
+            proxy_call_start: Final = datetime.now()  # noqa: DTZ005  # logging pipeline uses naive datetimes
             proxy_logging_obj: Final = (
                 await _build_virtual_call_logging_obj(
                     name=name,
@@ -1016,18 +1017,55 @@ if MCP_AVAILABLE:
                 if name == MCP_PROXY_CALL_TOOL_NAME
                 else None
             )
-            return await handle_mcp_proxy_tool(
-                name=name,
-                arguments=arguments or {},  # mutable-ok: proxy handler payload
-                user_api_key_dict=user_api_key_auth,
-                client_ip=client_ip,
-                mcp_servers=mcp_servers,
-                mcp_auth_header=mcp_auth_header,
-                mcp_server_auth_headers=mcp_server_auth_headers,
-                oauth2_headers=oauth2_headers,
-                raw_headers=raw_headers,
-                litellm_logging_obj=proxy_logging_obj,
-            )
+            try:
+                proxy_result: Final = await handle_mcp_proxy_tool(
+                    name=name,
+                    arguments=arguments or {},  # mutable-ok: proxy handler payload
+                    user_api_key_dict=user_api_key_auth,
+                    client_ip=client_ip,
+                    mcp_servers=mcp_servers,
+                    mcp_auth_header=mcp_auth_header,
+                    mcp_server_auth_headers=mcp_server_auth_headers,
+                    oauth2_headers=oauth2_headers,
+                    raw_headers=raw_headers,
+                    litellm_logging_obj=proxy_logging_obj,
+                )
+            except Exception as exc:
+                if proxy_logging_obj is not None:
+                    from litellm.proxy.proxy_server import proxy_logging_obj as request_logging_obj
+
+                    failure_end: Final = datetime.now()  # noqa: DTZ005  # matches the logging pipeline start time
+                    failure_traceback: Final = traceback.format_exc(limit=MAXIMUM_TRACEBACK_LINES_TO_LOG)
+                    try:
+                        proxy_logging_obj.failure_handler(exc, failure_traceback, proxy_call_start, failure_end)
+                        await proxy_logging_obj.async_failure_handler(
+                            exc, failure_traceback, proxy_call_start, failure_end
+                        )
+                        if not isinstance(exc, MCPUpstreamAuthError):
+                            await request_logging_obj.post_call_failure_hook(
+                                request_data={  # mutable-ok: failure hook mutates its request payload
+                                    "name": name,
+                                    "arguments": arguments,
+                                    "litellm_logging_obj": proxy_logging_obj,
+                                },
+                                original_exception=exc,
+                                user_api_key_dict=user_api_key_auth,
+                                route="/mcp/call_tool",
+                                traceback_str=failure_traceback,
+                            )
+                    except Exception:  # noqa: BLE001  # a failing failure hook must not mask the tool call's own error
+                        verbose_logger.exception("Error logging failed MCP proxy tool call")
+                raise
+            if proxy_logging_obj is not None:
+                return await _fire_mcp_tool_call_logging(
+                    logging_obj=proxy_logging_obj,
+                    result=proxy_result,
+                    start_time=proxy_call_start,
+                    end_time=datetime.now(),  # noqa: DTZ005  # matches the logging pipeline start time
+                    user_api_key_auth=user_api_key_auth,
+                    request_data=types.MappingProxyType({"name": name, "arguments": arguments}),
+                )
+            return proxy_result
 
         if name not in VIRTUAL_TOOL_NAMES:
             return None
@@ -3493,7 +3531,9 @@ if MCP_AVAILABLE:
         server_name: str | None,
         session_id: str | None = None,
     ) -> StandardLoggingMCPToolCall:
-        mcp_server: Final = global_mcp_server_manager._get_mcp_server_from_tool_name(name)
+        mcp_server: Final = global_mcp_server_manager._get_mcp_server_from_tool_name(
+            add_server_prefix_to_name(name, server_name) if server_name else name
+        )
         namespaced_tool_name: Final = f"{server_name}/{name}" if server_name else name
         if mcp_server:
             mcp_info: Final = mcp_server.mcp_info or {}

litellm/proxy/_experimental/mcp_server/toolset_db.py

@@ -132,10 +132,18 @@ async def update_mcp_toolset(
     data: UpdateMCPToolsetRequest,
     touched_by: str,
 ) -> MCPToolset | None:
-    data_dict: Final = data.model_dump(exclude_none=True, exclude={"toolset_id"})
-    if "tools" in data_dict:
-        data_dict["tools"] = json.dumps(data_dict["tools"])
-    data_dict["updated_by"] = touched_by
+    """A partial update: absent keeps, null clears. A toolset always has a name and a
+    tool list, so a null ``toolset_name`` or ``tools`` is a no-op rather than a clear;
+    emptying the tool selection is an explicit ``[]``, which cannot be mistaken for a
+    caller that left the field out."""
+    data_dict: Final = dict(  # mutable-ok: Prisma requires a plain dict for JSON query serialization
+        (
+            (field, json.dumps(value) if field == "tools" else value)
+            for field, value in data.model_dump(exclude_unset=True).items()
+            if field != "toolset_id" and (field not in ("toolset_name", "tools") or value is not None)
+        ),
+        updated_by=touched_by,
+    )
     try:
         row: Final = await _toolset_table(prisma_client).update(
             where={"toolset_id": data.toolset_id},

litellm/proxy/_types.py

@@ -3,6 +3,7 @@ import json
 import os
 from collections.abc import Callable, Mapping
 from datetime import datetime
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, NamedTuple
 
 import httpx
@@ -1294,6 +1295,13 @@ class UpdateKeyRequest(KeyRequestBase):
     rotation_interval: str | None = None
     organization_id: str | None = None
 
+    @model_validator(mode="before")
+    @classmethod
+    def drop_blank_team_id(cls, values: object) -> object:
+        if isinstance(values, Mapping) and values.get("team_id") == "":
+            return MappingProxyType({k: v for k, v in values.items() if k != "team_id"})
+        return values
+
     @field_validator("organization_id", mode="before")
     @classmethod
     def treat_cleared_organization_id_as_unset(cls, v: object) -> object:
@@ -2828,6 +2836,18 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
             "UI username/password login. Default is False."
         ),
     )
+    disable_env_credential_login: bool | None = Field(
+        None,
+        description=(
+            "If True, disables signing in to the Admin UI with the environment credentials: "
+            "UI_USERNAME/UI_PASSWORD, or the master key when UI_PASSWORD is unset (that fallback "
+            "means env-credential login is always live by default). Database users with passwords "
+            "are unaffected. LOCKOUT RISK: create at least one proxy admin user with a password "
+            "before enabling, or nobody can sign in to the UI. A locked-out admin can still "
+            "administer the proxy over the API with the master key, and can unset this setting "
+            "and restart the proxy to restore env-credential login. Default is False."
+        ),
+    )
     disable_budget_reservation: bool | None = Field(
         None,
         description=(
@@ -5148,9 +5168,26 @@ class CostEstimateRequest(LiteLLMPydanticObjectBase):
     model: str = Field(description="Model name (from /model_group/info)")
     input_tokens: int = Field(description="Expected input tokens per request", ge=0)
     output_tokens: int = Field(description="Expected output tokens per request", ge=0)
+    cache_read_input_tokens: int = Field(
+        default=0, description="Input tokens read from the prompt cache; counted within input_tokens", ge=0
+    )
+    cache_creation_input_tokens: int = Field(
+        default=0, description="Input tokens written to the prompt cache; counted within input_tokens", ge=0
+    )
+    reasoning_tokens: int = Field(
+        default=0, description="Reasoning tokens the model emits; counted within output_tokens", ge=0
+    )
     num_requests_per_day: int | None = Field(default=None, description="Number of requests per day", ge=0)
     num_requests_per_month: int | None = Field(default=None, description="Number of requests per month", ge=0)
 
+    @model_validator(mode="after")
+    def validate_token_subsets(self) -> "CostEstimateRequest":
+        if self.cache_read_input_tokens + self.cache_creation_input_tokens > self.input_tokens:
+            raise ValueError("cache_read_input_tokens plus cache_creation_input_tokens cannot exceed input_tokens")
+        if self.reasoning_tokens > self.output_tokens:
+            raise ValueError("reasoning_tokens cannot exceed output_tokens")
+        return self
+
 
 class CostEstimateResponse(LiteLLMPydanticObjectBase):
     """Response body for /cost/estimate endpoint."""
@@ -5158,6 +5195,9 @@ class CostEstimateResponse(LiteLLMPydanticObjectBase):
     model: str
     input_tokens: int
     output_tokens: int
+    cache_read_input_tokens: int = 0
+    cache_creation_input_tokens: int = 0
+    reasoning_tokens: int = 0
     num_requests_per_day: int | None = None
     num_requests_per_month: int | None = None
     # Per-request costs
@@ -5165,17 +5205,33 @@ class CostEstimateResponse(LiteLLMPydanticObjectBase):
     input_cost_per_request: float = Field(description="Input token cost per request (before margin)")
     output_cost_per_request: float = Field(description="Output token cost per request (before margin)")
     margin_cost_per_request: float = Field(default=0.0, description="Margin/fee added per request")
+    cache_read_cost_per_request: float = Field(default=0.0, description="Cache-read share of input_cost_per_request")
+    cache_creation_cost_per_request: float = Field(
+        default=0.0, description="Cache-write share of input_cost_per_request"
+    )
+    reasoning_cost_per_request: float = Field(default=0.0, description="Reasoning share of output_cost_per_request")
     # Daily costs (if num_requests_per_day provided)
     daily_cost: float | None = Field(default=None, description="Total daily cost (includes margin)")
     daily_input_cost: float | None = Field(default=None, description="Daily input token cost")
     daily_output_cost: float | None = Field(default=None, description="Daily output token cost")
     daily_margin_cost: float | None = Field(default=None, description="Daily margin/fee")
+    daily_cache_read_cost: float | None = Field(default=None, description="Cache-read share of daily_input_cost")
+    daily_cache_creation_cost: float | None = Field(default=None, description="Cache-write share of daily_input_cost")
+    daily_reasoning_cost: float | None = Field(default=None, description="Reasoning share of daily_output_cost")
     # Monthly costs (if num_requests_per_month provided)
     monthly_cost: float | None = Field(default=None, description="Total monthly cost (includes margin)")
     monthly_input_cost: float | None = Field(default=None, description="Monthly input token cost")
     monthly_output_cost: float | None = Field(default=None, description="Monthly output token cost")
     monthly_margin_cost: float | None = Field(default=None, description="Monthly margin/fee")
-    # Pricing info
-    input_cost_per_token: float | None = None
-    output_cost_per_token: float | None = None
+    monthly_cache_read_cost: float | None = Field(default=None, description="Cache-read share of monthly_input_cost")
+    monthly_cache_creation_cost: float | None = Field(
+        default=None, description="Cache-write share of monthly_input_cost"
+    )
+    monthly_reasoning_cost: float | None = Field(default=None, description="Reasoning share of monthly_output_cost")
+    # Pricing info: the rates this request's usage bills at, after token tiers and regional multipliers
+    input_cost_per_token: float | None = Field(default=None, description="Rate billed per input token")
+    output_cost_per_token: float | None = Field(default=None, description="Rate billed per output token")
+    cache_read_input_token_cost: float | None = Field(default=None, description="Rate billed per cache-read token")
+    cache_creation_input_token_cost: float | None = Field(default=None, description="Rate billed per cache-write token")
+    output_cost_per_reasoning_token: float | None = Field(default=None, description="Rate billed per reasoning token")
     provider: str | None = None

litellm/proxy/auth/auth_checks.py

@@ -475,6 +475,7 @@ def _is_model_cost_zero(model: str | list[str] | None, llm_router: Router | None
 
 
 _NO_MODEL_INFO: Final[Mapping[str, object]] = MappingProxyType({})
+_TEAM_GRANT_RELATIONS: Final[Mapping[str, object]] = MappingProxyType({"litellm_model_table": True})
 
 
 def _has_ptu_flat_cost(model: str, llm_router: "Router") -> bool:
@@ -2858,7 +2859,9 @@ class TeamNotFoundError(HTTPException):
 async def _get_team_db_check(
     team_id: str, prisma_client: PrismaClient, team_id_upsert: bool | None = None
 ) -> "_PrismaTeamRow | None":
-    response = await _team_table(TeamRepository(prisma_client)).find_unique(where={"team_id": team_id})
+    response = await _team_table(TeamRepository(prisma_client)).find_unique(
+        where={"team_id": team_id}, include=_TEAM_GRANT_RELATIONS
+    )
 
     if response is None and team_id_upsert:
         from litellm.proxy.management_endpoints.team_endpoints import new_team
@@ -3158,7 +3161,9 @@ async def get_team_object_by_alias(
 
     # Query database by team_alias
     try:
-        teams: Final = await _team_table(TeamRepository(prisma_client)).find_many(where={"team_alias": team_alias})
+        teams: Final = await _team_table(TeamRepository(prisma_client)).find_many(
+            where={"team_alias": team_alias}, include=_TEAM_GRANT_RELATIONS
+        )
 
         if not teams:
             raise HTTPException(

litellm/proxy/auth/handle_jwt.py

@@ -53,6 +53,7 @@ from litellm.proxy._types import (
 )
 from litellm.proxy.auth.auth_checks import can_team_access_model
 from litellm.proxy.auth.route_checks import RouteChecks
+from litellm.proxy.auth.team_grants import team_model_aliases
 from litellm.proxy.common_utils.user_api_key_cache import (
     UserApiKeyCache,
     get_management_object_ttl,
@@ -1595,7 +1596,7 @@ class JWTAuthManager:
                             model=requested_model,
                             team_object=team_object,
                             llm_router=llm_router,
-                            team_model_aliases=None,
+                            team_model_aliases=team_model_aliases(team_object),
                         )
                     ):
                         is_allowed = allowed_routes_check(
@@ -2132,7 +2133,7 @@ class JWTAuthManager:
                         model=requested_model,
                         team_object=team_object,
                         llm_router=llm_router,
-                        team_model_aliases=None,
+                        team_model_aliases=team_model_aliases(team_object),
                     )
                 except ProxyException:
                     continue

litellm/proxy/auth/login_utils.py

@@ -85,6 +85,29 @@ def get_ui_credentials(master_key: str | None) -> tuple[str, str]:
     return ui_username, ui_password
 
 
+def _matches_env_credentials(username: str, password: str, master_key: str | None) -> bool:
+    ui_username, ui_password = get_ui_credentials(master_key)
+    return secrets.compare_digest(username.encode("utf-8"), ui_username.encode("utf-8")) and secrets.compare_digest(
+        password.encode("utf-8"), ui_password.encode("utf-8")
+    )
+
+
+def is_env_credential_login_enabled(general_settings: Mapping[str, object]) -> bool:
+    """Whether a login with UI_USERNAME/UI_PASSWORD (or the master-key fallback) can succeed.
+
+    Two settings can turn it off: `disable_env_credential_login` unconditionally, and
+    `disable_password_login_when_sso_enabled` as a side effect, since its gate rejects
+    every username/password login before the env comparison runs. Feeds both the
+    `authenticate_user` gate and the Admin UI warning banner, so the banner never nags
+    about a login path that is already unreachable.
+    """
+    if general_settings.get("disable_env_credential_login") is True:
+        return False
+    if general_settings.get("disable_password_login_when_sso_enabled") is True and is_sso_provider_fully_configured():
+        return False
+    return True
+
+
 class LoginResult:
     """Result object containing authentication data from login."""
 
@@ -129,7 +152,8 @@ async def authenticate_user(
         master_key: Master key for the proxy (required)
         prisma_client: Prisma database client (optional)
         general_settings: Proxy general_settings, checked for
-            `disable_password_login_when_sso_enabled`
+            `disable_password_login_when_sso_enabled` and
+            `disable_env_credential_login`
 
     Returns:
         LoginResult: Object containing authentication data
@@ -170,8 +194,6 @@ async def authenticate_user(
             code=500,
         )
 
-    ui_username, ui_password = get_ui_credentials(master_key)
-
     # Check if we can find the `username` in the db. On the UI, users can enter username=their email
     _user_row: LiteLLM_UserTable | None = None
     user_role: (
@@ -197,8 +219,8 @@ async def authenticate_user(
     - Login with UI_USERNAME and UI_PASSWORD
     - Login with Invite Link `user_email` and `password` combination
     """
-    if secrets.compare_digest(username.encode("utf-8"), ui_username.encode("utf-8")) and secrets.compare_digest(
-        password.encode("utf-8"), ui_password.encode("utf-8")
+    if general_settings.get("disable_env_credential_login") is not True and _matches_env_credentials(
+        username, password, master_key
     ):
         # Non SSO -> If user is using UI_USERNAME and UI_PASSWORD they are Proxy admin
         user_role = LitellmUserRoles.PROXY_ADMIN
@@ -340,8 +362,13 @@ async def authenticate_user(
                 code=401,
             )
     else:
+        env_credentials_hint: Final = (
+            "\nCheck 'UI_USERNAME', 'UI_PASSWORD' in .env file"
+            if is_env_credential_login_enabled(general_settings)
+            else ""
+        )
         raise ProxyException(
-            message="Invalid credentials used to access UI.\nCheck 'UI_USERNAME', 'UI_PASSWORD' in .env file",
+            message=f"Invalid credentials used to access UI.{env_credentials_hint}",
             type=ProxyErrorTypes.auth_error,
             param="invalid_credentials",
             code=401,

litellm/proxy/auth/team_grants.py

@@ -0,0 +1,122 @@
+"""Project a team row (plus the caller's membership in it) onto the ``team_*`` fields of ``UserAPIKeyAuth``.
+
+The virtual-key path gets these fields for free from the combined-view SQL join. Every other auth path
+starts from a ``LiteLLM_TeamTable`` object instead and has to copy them over by hand, which is how JWT
+callers kept losing grants (aliases, permissions, limits) one field at a time. Build the badge through
+``team_grants`` and the two paths cannot drift.
+"""
+
+from collections.abc import Mapping, Sequence
+from types import MappingProxyType
+from typing import Annotated, Final
+
+from pydantic import BaseModel, BeforeValidator, ConfigDict, TypeAdapter, ValidationError
+from pydantic.main import IncEx
+from typing_extensions import ReadOnly, TypedDict
+
+from litellm.proxy._types import (
+    LiteLLM_ObjectPermissionTable,
+    LiteLLM_TeamMembership,
+    LiteLLM_TeamTable,
+    Member,
+)
+
+_MODEL_ALIASES_ADAPTER: Final = TypeAdapter(dict[str, str])
+_JSON_COLUMNS: Final[Mapping[str, IncEx | bool]] = MappingProxyType(
+    {"metadata": True, "litellm_model_table": MappingProxyType({"model_aliases": True})}
+)
+
+
+def _decode_model_aliases(value: object) -> object:
+    """``LiteLLM_ModelTable.model_aliases`` is typed ``str | dict``; writers hand Prisma ``json.dumps(...)``, so take both."""
+    if not isinstance(value, str):
+        return value
+    try:
+        return _MODEL_ALIASES_ADAPTER.validate_json(value)
+    except ValidationError:
+        return None
+
+
+class TeamModelAliasTable(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
+    model_aliases: Annotated[Mapping[str, str] | None, BeforeValidator(_decode_model_aliases)] = None
+
+
+class _TeamJsonColumns(BaseModel):
+    """The two loosely typed columns on ``LiteLLM_TeamTable``, re-read with the shape the badge needs."""
+
+    metadata: Mapping[str, object] | None = None
+    litellm_model_table: TeamModelAliasTable | None = None
+
+
+class TeamGrants(TypedDict, total=False):
+    """Keyword arguments for ``UserAPIKeyAuth``. Empty when the caller has no team, so the model's own defaults apply."""
+
+    team_alias: ReadOnly[str | None]
+    team_tpm_limit: ReadOnly[int | None]
+    team_rpm_limit: ReadOnly[int | None]
+    team_max_budget: ReadOnly[float | None]
+    team_soft_budget: ReadOnly[float | None]
+    team_spend: ReadOnly[float | None]
+    team_models: ReadOnly[Sequence[str]]
+    team_blocked: ReadOnly[bool]
+    team_metadata: ReadOnly[Mapping[str, object] | None]
+    team_model_aliases: ReadOnly[Mapping[str, str] | None]
+    team_object_permission_id: ReadOnly[str | None]
+    team_object_permission: ReadOnly[LiteLLM_ObjectPermissionTable | None]
+    team_member: ReadOnly[Member | None]
+    team_member_spend: ReadOnly[float | None]
+    team_member_tpm_limit: ReadOnly[int | None]
+    team_member_rpm_limit: ReadOnly[int | None]
+
+
+def _json_columns(team_object: LiteLLM_TeamTable) -> _TeamJsonColumns:
+    try:
+        return _TeamJsonColumns.model_validate(team_object.model_dump(include=_JSON_COLUMNS))
+    except ValidationError:
+        return _TeamJsonColumns()
+
+
+def team_model_aliases(team_object: LiteLLM_TeamTable | None) -> Mapping[str, str] | None:
+    if team_object is None:
+        return None
+    alias_table: Final = _json_columns(team_object).litellm_model_table
+    return alias_table.model_aliases if alias_table is not None else None
+
+
+def team_grants(
+    team_object: LiteLLM_TeamTable | None,
+    team_membership: LiteLLM_TeamMembership | None,
+    user_id: str | None,
+) -> TeamGrants:
+    if team_object is None:
+        return TeamGrants()
+    json_columns: Final = _json_columns(team_object)
+    return TeamGrants(
+        team_alias=team_object.team_alias,
+        team_tpm_limit=team_object.tpm_limit,
+        team_rpm_limit=team_object.rpm_limit,
+        team_max_budget=team_object.max_budget,
+        team_soft_budget=team_object.soft_budget,
+        team_spend=team_object.spend,
+        team_models=tuple(team_object.models),
+        team_blocked=team_object.blocked,
+        team_metadata=json_columns.metadata,
+        team_model_aliases=(
+            json_columns.litellm_model_table.model_aliases if json_columns.litellm_model_table is not None else None
+        ),
+        team_object_permission_id=team_object.object_permission_id,
+        team_object_permission=team_object.object_permission,
+        team_member=next(
+            (m for m in team_object.members_with_roles if user_id is not None and m.user_id == user_id),
+            None,
+        ),
+        team_member_spend=team_membership.spend if team_membership is not None else None,
+        team_member_tpm_limit=(
+            team_membership.safe_get_team_member_tpm_limit() if team_membership is not None else None
+        ),
+        team_member_rpm_limit=(
+            team_membership.safe_get_team_member_rpm_limit() if team_membership is not None else None
+        ),
+    )

litellm/proxy/auth/user_api_key_auth.py

@@ -82,6 +82,7 @@ from litellm.proxy.auth.oauth2_proxy_hook import handle_oauth2_proxy_request
 from litellm.proxy.auth.resolvers import CredentialRef, Principal
 from litellm.proxy.auth.resolvers.store import IdentityStore
 from litellm.proxy.auth.route_checks import RouteChecks
+from litellm.proxy.auth.team_grants import team_grants
 from litellm.proxy.auth.trusted_proxy_utils import get_trusted_proxy_cidrs
 from litellm.proxy.common_utils.cache_coordinator import EventDrivenCacheCoordinator
 from litellm.proxy.common_utils.http_parsing_utils import (
@@ -1476,24 +1477,16 @@ async def _user_api_key_auth_builder(
                             user_id=user_id,
                             user_email=user_email,
                             team_id=team_id,
-                            team_alias=(team_object.team_alias if team_object is not None else None),
-                            team_tpm_limit=(team_object.tpm_limit if team_object is not None else None),
-                            team_rpm_limit=(team_object.rpm_limit if team_object is not None else None),
-                            team_models=(team_object.models if team_object is not None else []),
-                            team_metadata=(team_object.metadata if team_object is not None else None),
                             org_id=org_id,
                             end_user_id=end_user_id,
                             parent_otel_span=parent_otel_span,
                             jwt_claims=jwt_claims,
+                            **team_grants(team_object=team_object, team_membership=team_membership, user_id=user_id),
                         )
 
                     valid_token = UserAPIKeyAuth(
                         api_key=None,
                         team_id=team_id,
-                        team_alias=(team_object.team_alias if team_object is not None else None),
-                        team_tpm_limit=(team_object.tpm_limit if team_object is not None else None),
-                        team_rpm_limit=(team_object.rpm_limit if team_object is not None else None),
-                        team_models=(team_object.models if team_object is not None else []),
                         user_role=(
                             LitellmUserRoles(user_object.user_role)
                             if user_object is not None and user_object.user_role is not None
@@ -1507,17 +1500,8 @@ async def _user_api_key_auth_builder(
                         user_tpm_limit=(user_object.tpm_limit if user_object is not None else None),
                         user_rpm_limit=(user_object.rpm_limit if user_object is not None else None),
                         user_model_max_budget=(user_object.model_max_budget if user_object is not None else None),
-                        team_member_rpm_limit=(
-                            team_membership.safe_get_team_member_rpm_limit() if team_membership is not None else None
-                        ),
-                        team_member_tpm_limit=(
-                            team_membership.safe_get_team_member_tpm_limit() if team_membership is not None else None
-                        ),
-                        team_metadata=(team_object.metadata if team_object is not None else None),
                         jwt_claims=jwt_claims,
-                    )
-                    valid_token.team_object_permission = (
-                        team_object.object_permission if team_object is not None else None
+                        **team_grants(team_object=team_object, team_membership=team_membership, user_id=user_id),
                     )
 
                     # AUTO_REGISTER deferred from _resolve_jwt_to_virtual_key.

litellm/proxy/client/cli/README.md

@@ -508,7 +508,7 @@ The credential is short-lived by design (default 24h, configurable via `LITELLM_
 
 ### Route Every Claude Code Session Through the Proxy
 
-`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true` and `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY` to `1` when those keys are missing, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
+`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true` and `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY` to `1` when those keys are missing, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` or `ANTHROPIC_AUTH_TOKEN` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
 
 Two things need to already be true: you've run `lite login` (or `lite login --pkce`, whose key the helper renews on its own), since the apiKeyHelper depends on that stored token, and the proxy is already reachable, since `lite up` does not start one for you.
 
@@ -532,12 +532,28 @@ Cursor is not supported: it has no equivalent file-based config to hot-patch thi
 lite --base-url https://your-proxy.example.com login --config-claude
 ```
 
-It writes the same settings `lite up` does, `env.ANTHROPIC_BASE_URL`, `env.ENABLE_TOOL_SEARCH`, `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY`, and `apiKeyHelper`, but persistently: there is no backup, nothing to restore, and no foreground process to keep alive. Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
+It writes the same settings `lite up` does, `env.ANTHROPIC_BASE_URL`, `env.ENABLE_TOOL_SEARCH`, `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY`, and `apiKeyHelper`, but persistently: no foreground process to keep alive, and `lite unconfigure claude` restores what it changed (see below). Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
 
 Because the credential is reached through `apiKeyHelper` rather than copied into the file, a later `lite login` refreshes it with no further action: Claude Code re-runs the helper on every request and picks up whatever token the most recent login stored. Nothing secret is written to `settings.json`.
 
 Run it again to point Claude Code at a different proxy; the base URL and the helper are both rewritten. `lite up` and `--config-claude` manage the same file, so the flag refuses to run while a `lite up` session holds a backup, and tells you to run `lite down` first, rather than writing settings that `lite up` would silently revert when it stops.
 
+#### Configuring Claude Code Once, With a Virtual Key or Your Login
+
+`lite configure claude` wires Claude Code up persistently and `lite unconfigure claude` puts things back. It is what `lite login --config-claude` does, plus a pinned model and an undo, and it also takes a long-lived virtual key when that is what you have:
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/BerriAI/litellm/main/scripts/install.sh | sh
+lite --base-url https://your-proxy.example.com configure claude --api-key sk-... --model claude-auto
+claude
+```
+
+With `--api-key` (or `lite --api-key` / `LITELLM_PROXY_API_KEY`) the key is written into `env.ANTHROPIC_AUTH_TOKEN`. Without one, your `lite login` credential is used the way `--config-claude` uses it, through `apiKeyHelper`, so a later `lite login` (or a `--pkce` renewal) picks up on its own and nothing secret lands in the file; a missing or stale login is refreshed first. Either way the command checks the key against `GET /v1/models`, then patches `~/.claude/settings.json`: `env.ANTHROPIC_BASE_URL`, the credential, and `env.ENABLE_TOOL_SEARCH` and `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY` when those are missing, so Claude Code's `/model` picker lists the proxy's models (the ones whose id contains `claude` or `anthropic`) and you pick between them as usual. Claude Code keeps its own default model until you switch, so that id has to exist on the proxy for the first message to go through; `--model` (or the interactive prompt below) sets the model Claude Code starts on instead, as the top-level `model` key, which has to be on `/v1/models` for the key. Nothing forces Claude Code's sub-agent or background tiers onto a proxy model, so those built-in ids need to exist on the proxy too; `lite autoroute up` is the mode that pins every tier to one group. Claude Code treats a name it does not know as an unknown model: it prints a one-line `unrecognized_model` note, assumes a 200k context window and sends no thinking parameters for it, so either name the group like a Claude model id or append `[1m]` to opt into the 1M window. The other credential slots (`env.ANTHROPIC_API_KEY`, a stale `env.ANTHROPIC_AUTH_TOKEN` or `apiKeyHelper`) are removed so they cannot fight the one written. Every other setting is preserved and the file is written atomically with owner-only permissions; if `settings.json` is a symlink into a dotfiles repository, the key is written through to that target and the command says so, so keep it out of version control
+
+Plain `lite configure`, with no agent named, asks the same things interactively: which agents to wire (Claude Code today) and which of the proxy's models to start on, picked from `/v1/models` with a type-to-filter prompt
+
+What the command changed is recorded in `~/.litellm/claude_configure_state.json` (previous values plus fingerprints of what was written, never a second copy of the key). `lite unconfigure claude` restores each of those keys only if it still holds what `configure` wrote, so anything you changed since is left alone and named in the output; a `settings.json` or `env` object that only existed because of `configure` is removed again. Ownership moves only by a write: running `configure` again (a re-login is one) refreshes the record only for the keys its merge changed, keeps the original snapshot of a key that still holds what it wrote, and snapshots afresh a key you changed in between, so `unconfigure` brings back whatever the repeat displaced and never adopts your edit as its own. A credential (`env.ANTHROPIC_API_KEY`, `env.ANTHROPIC_AUTH_TOKEN`, `apiKeyHelper`) is put back only when the restored file points at the `ANTHROPIC_BASE_URL` it was captured next to; otherwise it stays removed, the output says which server it belonged to, and the receipt is kept so pointing the URL back and running `unconfigure` again finishes the job. It also undoes `lite login --config-claude`, which writes through the same path. Like `--config-claude`, both refuse to run while a `lite up` or `lite autoroute up` session holds a backup, and that check comes before any login prompt or request
+
 ### QA Complexity-Based Auto-Routing Against Your Real Proxy
 
 `lite autoroute` lets you try LiteLLM's complexity-based auto-routing -- picking a cheaper or more expensive model depending on how complex a prompt looks -- against models your key already has access to on your real, running proxy, without editing that proxy's `config.yaml` and without any real request ever bypassing it. It builds a second, throwaway proxy locally that forwards every request back to your real proxy, and points Claude Code at that local proxy for the duration of the session.
@@ -584,7 +600,7 @@ An interactive wizard. It runs the same model-group discovery as above, splits t
 
 The wizard writes the result to `~/.litellm/autorouter/config.yaml` with `0600` permissions, since the file embeds your real proxy API key. Every model referenced anywhere in that config -- tier targets, the classifier model, the embedding model -- becomes its own `litellm_proxy/<model-name>` deployment whose `api_base` and `api_key` point back at your real proxy. That is the trick that keeps your real proxy's config untouched: every actual network call this generates, whether it is the routed completion, an LLM-classifier call, or an embedding call, forwards transparently through your real, already-running proxy with your real key.
 
-You do not need to tell Claude Code to request `autorouter` by name yourself: `lite autoroute up` also sets `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_DEFAULT_HAIKU_MODEL`, and `ANTHROPIC_DEFAULT_OPUS_MODEL` to `autorouter` in `~/.claude/settings.json`, so every one of Claude Code's own model tiers requests it directly regardless of `/model` or whatever it defaults to otherwise. (A bare `model_name: "*"` deployment looks like the obvious way to catch any request instead, but litellm's Router looks up auto-router deployments by the literal requested model string with no wildcard resolution, so a `"*"` entry would never actually match real traffic -- these env var overrides are what makes it work.)
+You do not need to tell Claude Code to request `autorouter` by name yourself: `lite autoroute up` also sets the top-level `model` and `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_DEFAULT_HAIKU_MODEL`, `ANTHROPIC_DEFAULT_OPUS_MODEL` and `ANTHROPIC_DEFAULT_FABLE_MODEL` to `autorouter` in `~/.claude/settings.json` (and `CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY` to `1` when missing, like every other wiring), so every one of Claude Code's own model tiers requests it directly regardless of `/model` or whatever it defaults to otherwise. (A bare `model_name: "*"` deployment looks like the obvious way to catch any request instead, but litellm's Router looks up auto-router deployments by the literal requested model string with no wildcard resolution, so a `"*"` entry would never actually match real traffic -- these env var overrides are what makes it work.)
 
 You must run `configure` at least once before `up`; running `up` first fails with a clear error telling you to configure first.
 

litellm/proxy/client/cli/commands/agents.py

@@ -4,6 +4,7 @@ import subprocess
 import sys
 from collections.abc import Callable, Mapping, Sequence
 from dataclasses import dataclass
+from pathlib import Path
 from types import MappingProxyType
 from typing import Final, TypeAlias
 
@@ -12,10 +13,12 @@ import requests
 from pydantic import BaseModel, TypeAdapter, ValidationError
 
 from .auth import CliContextObj, context_secret_vault, get_stored_api_key, login
+from .claude_settings import claude_settings_path, lite_api_key_helper_configured
 from .cmd_quoting import quote_for_cmd
 from .pi import (
     LITELLM_PROXY_API_KEY_ENV,
     PI_PROVIDER_NAME,
+    ListingFailure,
     PiSyncError,
     fetch_model_ids,
     fetch_model_limits,
@@ -83,6 +86,8 @@ def build_agent_env(
     base_url: str,
     api_key: str,
     profiles: frozenset[str],
+    *,
+    export_anthropic_token: bool = True,
 ) -> dict[str, str]:
     """Return a copy of base_env wired to route the agent through the proxy.
 
@@ -97,12 +102,19 @@ def build_agent_env(
     proxy's /v1/models; likewise left alone when already set.
     pi ignores both base URL variables and instead resolves $LITELLM_PROXY_API_KEY
     from its synced models.json provider entry.
+
+    With export_anthropic_token=False the bearer is left out (and any inherited
+    one dropped) so Claude Code asks its configured apiKeyHelper instead; Claude
+    Code prefers ANTHROPIC_AUTH_TOKEN over the helper and warns when both are set.
     """
     env: Final = dict(base_env)
     root: Final = base_url.rstrip("/")
     if PROFILE_ANTHROPIC in profiles:
         env[ANTHROPIC_BASE_URL_ENV] = root
-        env[ANTHROPIC_AUTH_TOKEN_ENV] = api_key
+        if export_anthropic_token:
+            env[ANTHROPIC_AUTH_TOKEN_ENV] = api_key
+        else:
+            env.pop(ANTHROPIC_AUTH_TOKEN_ENV, None)
         env.pop(ANTHROPIC_API_KEY_ENV, None)
         if ENABLE_TOOL_SEARCH_ENV not in env:
             env[ENABLE_TOOL_SEARCH_ENV] = ENABLE_TOOL_SEARCH_VALUE
@@ -165,7 +177,9 @@ def prepare_pi(
     """
     ids: Final = fetch_model_ids(base_url, api_key, get=get)
     if isinstance(ids, PiSyncError):
-        raise AgentRunError(ids.message)
+        raise AgentRunError(
+            f"{ids.message} pi would have nothing to run." if ids.kind is ListingFailure.EMPTY else ids.message
+        )
     limits: Final = fetch_model_limits(base_url, api_key, get=get)
     path: Final = models_json_path(base_env)
     error: Final = sync_models_json(path, base_url, ids, limits)
@@ -460,6 +474,7 @@ def run_agent(
     launcher: Callable[[str, Sequence[str], Mapping[str, str]], None] = _hand_off,
     reattach_terminal: Callable[[], None] | None = None,
     preparers: Mapping[str, _Preparer] = MappingProxyType(_PREPARERS),
+    export_anthropic_token: bool = True,
 ) -> None:
     """Validate, wire the environment, and hand off to the agent.
 
@@ -491,7 +506,9 @@ def run_agent(
 
     env: Final = MappingProxyType(
         {
-            **build_agent_env(env_before_sync, base_url, api_key, profiles),
+            **build_agent_env(
+                env_before_sync, base_url, api_key, profiles, export_anthropic_token=export_anthropic_token
+            ),
             **(_NO_EXTRA_ENV if isinstance(synced, ModelSyncSkipped) else synced),
         }
     )
@@ -529,14 +546,26 @@ def resolve_api_key(ctx: click.Context) -> str:
 _SKIP_VERIFY_HELP: Final = "Skip the pre-launch key check against the proxy."
 
 
+def _helper_supplies_token(
+    ctx_obj: CliContextObj, base_url: str, profiles: frozenset[str], settings_path: Path
+) -> bool:
+    if PROFILE_ANTHROPIC not in profiles or not ctx_obj.get("api_key_from_token_file"):
+        return False
+    return lite_api_key_helper_configured(base_url, settings_path)
+
+
 def _launch(ctx: click.Context, binary: str, args: Sequence[str], *, skip_verify: bool) -> None:
     ctx_obj: Final[CliContextObj] = ctx.obj
     base_url: Final = ctx_obj["base_url"]
     started_interactive: Final = _is_interactive()
     api_key: Final = resolve_api_key(ctx)
 
-    display_name, _ = agent_profile(binary)
+    display_name, profiles = agent_profile(binary)
+    settings_path: Final = claude_settings_path(os.environ)
+    helper_supplies_token: Final = _helper_supplies_token(ctx_obj, base_url, profiles, settings_path)
     click.echo(f"litellm: routing {display_name} through proxy at {base_url.rstrip('/')}")
+    if helper_supplies_token:
+        click.echo(f"litellm: {display_name} reads its key from the apiKeyHelper in {settings_path}")
 
     try:
         run_agent(
@@ -545,6 +574,7 @@ def _launch(ctx: click.Context, binary: str, args: Sequence[str], *, skip_verify
             [binary, *args],
             skip_verify=skip_verify,
             reattach_terminal=(_restore_controlling_terminal if started_interactive else None),
+            export_anthropic_token=not helper_supplies_token,
         )
     except AgentRunError as e:
         raise click.ClickException(str(e))

litellm/proxy/client/cli/commands/auth.py

@@ -1,3 +1,4 @@
+import os
 import sys
 import time
 import webbrowser
@@ -40,10 +41,16 @@ from litellm.litellm_core_utils.cli_token_utils import (
 )
 
 from .claude_settings import (
-    CLAUDE_SETTINGS_PATH,
-    SETTINGS_FILE_OWNERS,
+    STARTING_MODEL_ROLE,
+    ApiKeyHelper,
     ClaudeSettingsError,
-    write_claude_settings,
+    KeepModel,
+    claude_settings_path,
+    configure_claude_settings,
+    configure_state_path,
+    refuse_while_owned,
+    resolve_api_key_helper,
+    settings_file_owners,
 )
 from .pkce_login import (
     Http,
@@ -778,13 +785,24 @@ def _render_and_prompt_for_team_selection(teams: list[CliTeam]) -> str | None:
 
 
 def _configure_claude_code(base_url: str) -> None:
-    """Point Claude Code at base_url by patching ~/.claude/settings.json."""
+    """Point Claude Code at base_url by patching the settings.json it reads, undoable with `lite unconfigure claude`."""
+    settings_path: Final = claude_settings_path(os.environ)
     try:
-        write_claude_settings(base_url, CLAUDE_SETTINGS_PATH, SETTINGS_FILE_OWNERS)
+        configure_claude_settings(
+            base_url,
+            ApiKeyHelper(resolve_api_key_helper(base_url)),
+            KeepModel(),
+            settings_path,
+            configure_state_path(settings_path),
+            settings_file_owners(settings_path),
+        )
     except ClaudeSettingsError as e:
         raise click.ClickException(f"Logged in, but could not configure Claude Code: {e}")
-    click.echo(f"\nConfigured Claude Code: {CLAUDE_SETTINGS_PATH} now routes through {base_url.rstrip('/')}.")
-    click.echo("Your other Claude Code settings were left untouched. Restart Claude Code to pick this up.")
+    click.echo(f"\nConfigured Claude Code: {settings_path} now routes through {base_url.rstrip('/')}.")
+    click.echo(
+        "Your other Claude Code settings were left untouched. Restart Claude Code to pick this up. "
+        f"Undo with `lite unconfigure claude`; `lite configure claude --model` sets {STARTING_MODEL_ROLE}."
+    )
 
 
 def _finish_login(base_url: str, api_key: str, config_claude: bool, stored: SecretSave) -> None:
@@ -853,6 +871,12 @@ def login(ctx: click.Context, config_claude: bool, pkce: bool) -> None:
 
     ctx_obj: Final[CliContextObj] = ctx.obj
     base_url: Final = ctx_obj["base_url"]
+    if config_claude:
+        settings_path: Final = claude_settings_path(os.environ)
+        try:
+            refuse_while_owned(settings_path, settings_file_owners(settings_path))
+        except ClaudeSettingsError as e:
+            raise click.ClickException(f"Cannot configure Claude Code, so not logging in: {e}")
 
     try:
         if pkce:

litellm/proxy/client/cli/commands/autoroute/commands.py

@@ -14,11 +14,13 @@ from ..claude_settings import (
     AUTOROUTE_BACKUP_PATH,
     CLAUDE_SETTINGS_PATH,
     ClaudeSettingsError,
+    StaticToken,
     load_json_or_empty,
+    merge_claude_settings,
 )
 from ..up import BackupRecord as ClaudeBackupRecord
 from ..up import restore_claude_settings, write_backup
-from .config import master_key_from_config
+from .config import AUTOROUTER_MODEL_NAME, master_key_from_config
 from .process import (
     CONFIG_PATH,
     DEFAULT_AUTOROUTE_PORT,
@@ -37,7 +39,6 @@ from .process import (
     terminate,
     write_pid_record,
 )
-from .settings import merge_claude_settings_static_token
 from .wizard import run_configure_wizard
 
 _GENERATED_CONFIG_ADAPTER: Final = TypeAdapter(dict[str, JsonValue])
@@ -156,7 +157,9 @@ def up(port: int) -> None:
             ClaudeBackupRecord(existed=original_existed, content=original_settings if original_existed else None),
             AUTOROUTE_BACKUP_PATH,
         )
-        merged: Final = merge_claude_settings_static_token(original_settings, base_url, master_key)
+        merged: Final = merge_claude_settings(
+            original_settings, base_url, StaticToken(master_key), AUTOROUTER_MODEL_NAME, AUTOROUTER_MODEL_NAME
+        )
         CLAUDE_SETTINGS_PATH.parent.mkdir(parents=True, exist_ok=True)
         with secure_create(CLAUDE_SETTINGS_PATH) as f:
             json.dump(merged, f, indent=2)

litellm/proxy/client/cli/commands/autoroute/settings.py

@@ -1,51 +0,0 @@
-from typing import Final
-
-from pydantic import JsonValue
-
-from .config import AUTOROUTER_MODEL_NAME
-
-ENV_KEY: Final = "env"
-API_KEY_HELPER_KEY: Final = "apiKeyHelper"
-ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
-ANTHROPIC_AUTH_TOKEN_KEY: Final = "ANTHROPIC_AUTH_TOKEN"
-ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
-ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
-ENABLE_TOOL_SEARCH_VALUE: Final = "true"
-# Force every one of Claude Code's own model tiers to request the auto-router by name.
-# Router's auto-router registry is keyed by the literal requested model string
-# (litellm/router.py:10711-10717) with no wildcard/pattern resolution, so a bare "*"
-# model_name can never work as a catch-all -- these overrides are what actually makes
-# Claude Code send "autorouter" regardless of /model or its own version-specific defaults.
-ANTHROPIC_DEFAULT_MODEL_ENV_KEYS: Final = (
-    "ANTHROPIC_DEFAULT_SONNET_MODEL",
-    "ANTHROPIC_DEFAULT_HAIKU_MODEL",
-    "ANTHROPIC_DEFAULT_OPUS_MODEL",
-)
-
-
-def merge_claude_settings_static_token(
-    settings: dict[str, JsonValue], base_url: str, auth_token: str
-) -> dict[str, JsonValue]:
-    """Return a new settings dict wired to a local ephemeral proxy with a static token.
-
-    Unlike up.py's merge_claude_settings (which sets apiKeyHelper for a long-lived, real
-    remote proxy needing refreshable SSO tokens), this proxy is ephemeral and its key is the
-    locally persisted autoroute master key, so a plain env var is simpler and correct. Any
-    existing apiKeyHelper is cleared so it can't fight with the static token.
-    """
-    raw_env: Final = settings.get(ENV_KEY, {})
-    base_env: Final = raw_env if isinstance(raw_env, dict) else {}
-    env: Final[dict[str, JsonValue]] = {
-        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
-        **base_env,
-        ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
-        ANTHROPIC_AUTH_TOKEN_KEY: auth_token,
-        **{key: AUTOROUTER_MODEL_NAME for key in ANTHROPIC_DEFAULT_MODEL_ENV_KEYS},
-    }
-    env.pop(ANTHROPIC_API_KEY_KEY, None)
-    merged: Final[dict[str, JsonValue]] = {**settings, ENV_KEY: env}
-    merged.pop(API_KEY_HELPER_KEY, None)
-    return merged
-
-
-__all__ = ["merge_claude_settings_static_token"]

litellm/proxy/client/cli/commands/claude_settings.py

@@ -1,37 +1,71 @@
 """Shared handling of Claude Code's ~/.claude/settings.json.
 
-`lite up` patches this file temporarily and restores it on exit; `lite login
---config-claude` patches it persistently. Both need the same merge and the same
-apiKeyHelper command, and `up` already imports from `auth`, so the shared parts
-live here rather than in either command module.
+`lite up` and `lite autoroute up` patch this file temporarily and restore it on
+exit; `lite login --config-claude` and `lite configure claude` patch it
+persistently and record how to undo it. All of them need the same merge and the
+same apiKeyHelper command, and `up` already imports from `auth`, so the shared
+parts live here rather than in any one command module.
 """
 
+import hashlib
+import json
 import shlex
 import shutil
 import sys
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence
 from dataclasses import dataclass
+from functools import reduce
+from itertools import chain
 from pathlib import Path
-from typing import Final
+from types import MappingProxyType
+from typing import Final, TypeAlias
 
-from pydantic import JsonValue, TypeAdapter, ValidationError
+from pydantic import BaseModel, ConfigDict, JsonValue, TypeAdapter, ValidationError
 
-from litellm.litellm_core_utils.private_json import write_private_json
+from litellm.litellm_core_utils.private_json import (
+    commit_staged_json,
+    discard_staged_json,
+    ensure_private_dir,
+    stage_private_json,
+)
 
 from .cmd_quoting import quote_for_cmd
 
 ENV_KEY: Final = "env"
 API_KEY_HELPER_KEY: Final = "apiKeyHelper"
+MODEL_KEY: Final = "model"
 ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
+ANTHROPIC_AUTH_TOKEN_KEY: Final = "ANTHROPIC_AUTH_TOKEN"
 ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
 ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
 ENABLE_TOOL_SEARCH_VALUE: Final = "true"
 ENABLE_GATEWAY_MODEL_DISCOVERY_KEY: Final = "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY"
 ENABLE_GATEWAY_MODEL_DISCOVERY_VALUE: Final = "1"
+ANTHROPIC_DEFAULT_MODEL_ENV_KEYS: Final = (
+    "ANTHROPIC_DEFAULT_SONNET_MODEL",
+    "ANTHROPIC_DEFAULT_HAIKU_MODEL",
+    "ANTHROPIC_DEFAULT_OPUS_MODEL",
+    "ANTHROPIC_DEFAULT_FABLE_MODEL",
+)
+OWNED_ENV_KEYS: Final = (
+    ENABLE_TOOL_SEARCH_KEY,
+    ENABLE_GATEWAY_MODEL_DISCOVERY_KEY,
+    ANTHROPIC_BASE_URL_KEY,
+    ANTHROPIC_AUTH_TOKEN_KEY,
+    ANTHROPIC_API_KEY_KEY,
+)
+OWNED_TOP_LEVEL_KEYS: Final = (API_KEY_HELPER_KEY, MODEL_KEY)
+OWNED_PATHS: Final = (*(f"{ENV_KEY}.{key}" for key in OWNED_ENV_KEYS), *OWNED_TOP_LEVEL_KEYS)
+_CREDENTIAL_ENV_KEYS: Final = frozenset((ANTHROPIC_API_KEY_KEY, ANTHROPIC_AUTH_TOKEN_KEY))
+_CREDENTIAL_PATHS: Final = (*(f"{ENV_KEY}.{key}" for key in sorted(_CREDENTIAL_ENV_KEYS)), API_KEY_HELPER_KEY)
+_BASE_URL_PATH: Final = f"{ENV_KEY}.{ANTHROPIC_BASE_URL_KEY}"
+STARTING_MODEL_ROLE: Final = "the /model picker's default row, the model Claude Code starts on"
 
 CLAUDE_SETTINGS_PATH: Final = Path.home() / ".claude" / "settings.json"
+CLAUDE_CONFIG_DIR_ENV: Final = "CLAUDE_CONFIG_DIR"
 BACKUP_PATH: Final = Path.home() / ".litellm" / "claude_settings_backup.json"
 AUTOROUTE_BACKUP_PATH: Final = Path.home() / ".litellm" / "autorouter" / "claude_settings_backup.json"
+CONFIGURE_STATE_PATH: Final = Path.home() / ".litellm" / "claude_configure_state.json"
 
 
 @dataclass(frozen=True, slots=True)
@@ -55,6 +89,129 @@ class ClaudeSettingsError(Exception):
     """Raised for any user-actionable failure while reading or writing Claude Code settings."""
 
 
+def claude_settings_path(environ: Mapping[str, str]) -> Path:
+    """The settings.json Claude Code reads: under CLAUDE_CONFIG_DIR when set, else ~/.claude/settings.json."""
+    config_dir: Final = environ.get(CLAUDE_CONFIG_DIR_ENV, "")
+    if not config_dir:
+        return CLAUDE_SETTINGS_PATH
+    return Path(config_dir).expanduser() / "settings.json"
+
+
+def _is_default_settings_file(settings_path: Path) -> bool:
+    return settings_path.resolve() == CLAUDE_SETTINGS_PATH.resolve()
+
+
+def settings_file_owners(settings_path: Path) -> tuple[SettingsFileOwner, ...]:
+    """The commands whose backups guard settings_path: `lite up` and `lite autoroute up` only ever manage the default file."""
+    return SETTINGS_FILE_OWNERS if _is_default_settings_file(settings_path) else ()
+
+
+def configure_state_path(settings_path: Path) -> Path:
+    """The receipt describing settings_path: the default file keeps CONFIGURE_STATE_PATH, and any other file
+    (a CLAUDE_CONFIG_DIR) gets its own beside it, keyed by its resolved path, so two settings files never
+    share one undo record."""
+    if _is_default_settings_file(settings_path):
+        return CONFIGURE_STATE_PATH
+    digest: Final = hashlib.sha256(str(settings_path.resolve()).encode()).hexdigest()
+    return CONFIGURE_STATE_PATH.parent / CONFIGURE_STATE_PATH.stem / f"{digest}.json"
+
+
+@dataclass(frozen=True, slots=True)
+class StaticToken:
+    """A long-lived virtual key, written into env.ANTHROPIC_AUTH_TOKEN."""
+
+    token: str
+
+
+@dataclass(frozen=True, slots=True)
+class ApiKeyHelper:
+    """A `lite auth print-token` command Claude Code runs per request, so a login renews in place."""
+
+    command: str
+
+
+ClaudeCredential: TypeAlias = StaticToken | ApiKeyHelper
+
+
+@dataclass(frozen=True, slots=True)
+class KeepModel:
+    """Leave the top-level `model` as it is, the user's or an earlier configure's (a re-login)."""
+
+
+@dataclass(frozen=True, slots=True)
+class UnpinModel:
+    """Let go of a `model` an earlier configure pinned; one the user set themselves stays."""
+
+
+@dataclass(frozen=True, slots=True)
+class StartOn:
+    """Pin the top-level `model`, the row Claude Code starts on."""
+
+    model: str
+
+
+ModelChoice: TypeAlias = KeepModel | UnpinModel | StartOn
+
+
+class OwnedValue(BaseModel):
+    """What one key held at a moment in time; `present=False` is an absent key, not a null one."""
+
+    model_config = ConfigDict(frozen=True)
+
+    present: bool
+    value: JsonValue = None
+
+
+class ConfigureReceipt(BaseModel):
+    """What `lite configure claude` found and what it owns, keyed by dotted path (`env.X` or a top-level key).
+
+    Ownership moves only by a write: `written` fingerprints the keys some configure changed, at the
+    value it wrote; a repeat configure refreshes a fingerprint only for a key its merge changed and
+    carries the earlier one otherwise, so a key the user edited in between stops matching and is left
+    alone. `previous` is what each key held before configure took it over; a repeat keeps the earlier
+    snapshot while the key still holds our value and snapshots afresh otherwise, so whatever the
+    repeat displaces is what comes back. `endpoints` is the ANTHROPIC_BASE_URL each credential slot
+    was captured beside, so a credential is only ever put back next to the server it was issued for.
+    No fingerprint is a second copy of a token.
+    """
+
+    model_config = ConfigDict(frozen=True)
+
+    file_existed: bool
+    env_present: bool
+    env_was_object: bool
+    previous: Mapping[str, OwnedValue]
+    written: Mapping[str, str]
+    endpoints: Mapping[str, OwnedValue]
+
+
+@dataclass(frozen=True, slots=True)
+class WithheldCredential:
+    """A credential left removed: captured beside `endpoint`, while the restored file points elsewhere."""
+
+    key: str
+    endpoint: str
+
+
+@dataclass(frozen=True, slots=True)
+class UnconfigureOutcome:
+    """Keys whose value unconfigure changed back, keys the user changed since and so were left as they
+    are, credentials withheld (the receipt is kept for them, so a later unconfigure can finish once the
+    URL points back), and whether no settings file remains."""
+
+    restored: tuple[str, ...]
+    kept: tuple[str, ...]
+    withheld: tuple[WithheldCredential, ...] = ()
+    file_removed: bool = False
+
+
+@dataclass(frozen=True, slots=True)
+class _Claim:
+    previous: OwnedValue
+    written: str | None
+    endpoint: OwnedValue | None
+
+
 def load_json_or_empty(path: Path) -> dict[str, JsonValue]:
     try:
         content: Final = path.read_bytes() if path.exists() else b""
@@ -70,29 +227,104 @@ def load_json_or_empty(path: Path) -> dict[str, JsonValue]:
         )
 
 
-def merge_claude_settings(
-    settings: Mapping[str, JsonValue], base_url: str, api_key_helper: str
-) -> dict[str, JsonValue]:
-    """Return a new settings dict wired to route Claude Code through the proxy.
+def _env_object(settings: Mapping[str, JsonValue], path: Path) -> Mapping[str, JsonValue]:
+    raw_env: Final = settings.get(ENV_KEY)
+    if raw_env is None:
+        return MappingProxyType({})
+    if not isinstance(raw_env, dict):
+        raise ClaudeSettingsError(
+            f'{path} has a non-object "{ENV_KEY}" value, which this would discard. Fix or remove it, then retry.'
+        )
+    return raw_env
 
-    Only env.ANTHROPIC_BASE_URL and the top-level apiKeyHelper are overridden; a
-    stray env.ANTHROPIC_API_KEY is dropped so it cannot outrank the helper-issued
-    token (same reasoning as build_agent_env in agents.py). ENABLE_TOOL_SEARCH
-    defaults to true because Claude Code turns tool search off when
-    ANTHROPIC_BASE_URL is not a first-party Anthropic host, and
-    CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY defaults to 1 so the /model picker
-    is filled from the proxy's /v1/models; existing values of both are left
-    alone. Every other key is preserved untouched.
+
+def refuse_while_owned(settings_path: Path, owners: Sequence[SettingsFileOwner]) -> None:
+    """Refuse while `lite up` or `lite autoroute up` holds a backup it will restore over any write; a
+    purely local check, so commands run it before any login prompt or request."""
+    for owner in owners:
+        if owner.backup_path.exists():
+            raise ClaudeSettingsError(
+                f"`{owner.start_command}` is currently managing {settings_path} (backup at "
+                f"{owner.backup_path}) and will restore it when it stops. "
+                f"Run `{owner.stop_command}` first, then retry."
+            )
+
+
+def _write_target(settings_path: Path) -> Path:
+    """Write through a symlinked settings.json rather than replacing the link, which would silently
+    detach a file symlinked into a dotfiles repo."""
+    try:
+        return settings_path.resolve() if settings_path.is_symlink() else settings_path
+    except OSError as e:
+        raise ClaudeSettingsError(f"Could not resolve {settings_path}: {e}") from e
+
+
+def _stage(path: Path, document: Mapping[str, object]) -> str:
+    try:
+        return stage_private_json(str(path), document)
+    except OSError as e:
+        raise ClaudeSettingsError(f"Could not write {path}: {e}") from e
+
+
+def _land(
+    path: Path,
+    staged: str | None,
+    also_discard: Sequence[str | None] = (),
+    commit: Callable[[str, str], None] = commit_staged_json,
+) -> None:
+    """Commit a staged file to `path`, or remove `path` when nothing is staged for it. The one place a
+    filesystem error becomes a ClaudeSettingsError; on failure the operation's other staged files are
+    discarded, so no temp file holding a token is left behind."""
+    try:
+        if staged is None:
+            path.unlink(missing_ok=True)
+        else:
+            commit(staged, str(path))
+    except OSError as e:
+        for other in also_discard:
+            if other is not None:
+                discard_staged_json(other)
+        raise ClaudeSettingsError(f"Could not {'remove' if staged is None else 'write'} {path}: {e}") from e
+
+
+def merge_claude_settings(
+    settings: Mapping[str, JsonValue],
+    base_url: str,
+    credential: ClaudeCredential,
+    default_model: str | None = None,
+    tier_model: str | None = None,
+) -> Mapping[str, JsonValue]:
+    """Return a new settings mapping wired to route Claude Code through the proxy.
+
+    A StaticToken lands in env.ANTHROPIC_AUTH_TOKEN, an ApiKeyHelper in the top-level apiKeyHelper;
+    the other credential slots are removed either way, since Claude Code given two credentials may
+    send the wrong one. ENABLE_TOOL_SEARCH and CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY get their
+    defaults only when missing. `default_model` is the top-level `model`, the row Claude Code starts
+    on; `tier_model` is `lite autoroute up`'s knob that points every ANTHROPIC_DEFAULT_*_MODEL at one
+    group. Apart from those tier keys, exactly OWNED_PATHS are touched.
     """
     raw_env: Final = settings.get(ENV_KEY, {})
-    base_env: Final = raw_env if isinstance(raw_env, dict) else {}
-    env: Final = {
-        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
-        ENABLE_GATEWAY_MODEL_DISCOVERY_KEY: ENABLE_GATEWAY_MODEL_DISCOVERY_VALUE,
-        **{key: value for key, value in base_env.items() if key != ANTHROPIC_API_KEY_KEY},
-        ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
-    }
-    return {**settings, ENV_KEY: env, API_KEY_HELPER_KEY: api_key_helper}
+    current_env: Final = raw_env if isinstance(raw_env, dict) else {}
+    env: Final = dict(  # mutable-ok: JSON document handed to json.dump, which rejects a read-only mapping
+        chain(
+            (
+                (ENABLE_TOOL_SEARCH_KEY, ENABLE_TOOL_SEARCH_VALUE),
+                (ENABLE_GATEWAY_MODEL_DISCOVERY_KEY, ENABLE_GATEWAY_MODEL_DISCOVERY_VALUE),
+            ),
+            ((key, value) for key, value in current_env.items() if key not in _CREDENTIAL_ENV_KEYS),
+            ((ANTHROPIC_BASE_URL_KEY, base_url.rstrip("/")),),
+            ((ANTHROPIC_AUTH_TOKEN_KEY, credential.token),) if isinstance(credential, StaticToken) else (),
+            ((key, tier_model) for key in ANTHROPIC_DEFAULT_MODEL_ENV_KEYS if tier_model is not None),
+        )
+    )
+    return dict(  # mutable-ok: JSON document handed to json.dump, which rejects a read-only mapping
+        chain(
+            ((key, value) for key, value in settings.items() if key not in (API_KEY_HELPER_KEY, ENV_KEY)),
+            ((ENV_KEY, env),),
+            ((API_KEY_HELPER_KEY, credential.command),) if isinstance(credential, ApiKeyHelper) else (),
+            ((MODEL_KEY, default_model),) if default_model is not None else (),
+        )
+    )
 
 
 def resolve_api_key_helper(base_url: str, platform: str = sys.platform) -> str:
@@ -121,56 +353,273 @@ def resolve_api_key_helper(base_url: str, platform: str = sys.platform) -> str:
     return " ".join(quote(token) for token in (lite_path, "--base-url", base_url, "auth", "print-token"))
 
 
-def write_claude_settings(base_url: str, settings_path: Path, owners: Sequence[SettingsFileOwner]) -> None:
-    """Persistently point Claude Code at base_url, preserving every unrelated setting.
+def lite_api_key_helper_configured(base_url: str, settings_path: Path) -> bool:
+    """Whether settings_path already carries the apiKeyHelper `lite login --config-claude` writes for base_url.
 
-    Refuses while any owner holds a backup: each restores its backup when it
-    stops, which would silently undo this write.
+    Only an exact match counts: a helper for another proxy, a hand-written one, or
+    settings that cannot be read leave the caller on the env-token path.
     """
-    for owner in owners:
-        if owner.backup_path.exists():
-            raise ClaudeSettingsError(
-                f"`{owner.start_command}` is currently managing {settings_path} (backup at "
-                f"{owner.backup_path}) and will restore it when it stops. "
-                f"Run `{owner.stop_command}` first, then retry."
-            )
-    normalized_base_url: Final = base_url.rstrip("/")
-    api_key_helper: Final = resolve_api_key_helper(normalized_base_url)
-    existing: Final = load_json_or_empty(settings_path)
-    raw_env: Final = existing.get(ENV_KEY)
-    if raw_env is not None and not isinstance(raw_env, dict):
-        raise ClaudeSettingsError(
-            f'{settings_path} has a non-object "{ENV_KEY}" value, which this would discard. '
-            "Fix or remove it, then retry."
-        )
-    merged: Final = merge_claude_settings(existing, normalized_base_url, api_key_helper)
-    # os.replace() swaps the symlink itself for a regular file, silently detaching a
-    # settings.json that is symlinked into a dotfiles repo. There is no backup to undo
-    # that here, unlike `lite up`, so write through to the link's target instead.
-    target: Final = settings_path.resolve() if settings_path.is_symlink() else settings_path
     try:
-        write_private_json(str(target), merged)
+        configured_helper: Final = load_json_or_empty(settings_path).get(API_KEY_HELPER_KEY)
+        return configured_helper == resolve_api_key_helper(base_url.rstrip("/"))
+    except ClaudeSettingsError:
+        return False
+
+
+def _owned(container: Mapping[str, JsonValue], key: str) -> OwnedValue:
+    return OwnedValue(present=key in container, value=container.get(key))
+
+
+def _fingerprint(owned: OwnedValue) -> str:
+    return hashlib.sha256(json.dumps(owned.model_dump(mode="json"), sort_keys=True).encode()).hexdigest()
+
+
+def _env(settings: Mapping[str, JsonValue]) -> Mapping[str, JsonValue]:
+    raw_env: Final = settings.get(ENV_KEY)
+    return raw_env if isinstance(raw_env, dict) else MappingProxyType({})
+
+
+def _lookup(settings: Mapping[str, JsonValue], path: str) -> OwnedValue:
+    section, _, key = path.rpartition(".")
+    return _owned(_env(settings) if section else settings, key)
+
+
+def _with_key(container: Mapping[str, JsonValue], key: str, owned: OwnedValue) -> Mapping[str, JsonValue]:
+    return dict(  # mutable-ok: JSON document handed to json.dump, which rejects a read-only mapping
+        chain(((k, v) for k, v in container.items() if k != key), ((key, owned.value),) if owned.present else ())
+    )
+
+
+def _with(settings: Mapping[str, JsonValue], path: str, owned: OwnedValue) -> Mapping[str, JsonValue]:
+    """`settings` with the key at `path` set (or removed when `owned` is absent); nothing else changes."""
+    section, _, key = path.rpartition(".")
+    if not section:
+        return _with_key(settings, key, owned)
+    return _with_key(settings, section, OwnedValue(present=True, value=_with_key(_env(settings), key, owned)))
+
+
+def _with_all(settings: Mapping[str, JsonValue], updates: Mapping[str, OwnedValue]) -> Mapping[str, JsonValue]:
+    return reduce(lambda acc, item: _with(acc, *item), updates.items(), settings)
+
+
+def _ours(settings: Mapping[str, JsonValue], path: str, receipt: ConfigureReceipt) -> bool:
+    """Whether the key still holds what a configure wrote (a key no configure ever changed is never ours)."""
+    return receipt.written.get(path) == _fingerprint(_lookup(settings, path))
+
+
+def _claim(
+    path: str,
+    current: Mapping[str, JsonValue],
+    merged: Mapping[str, JsonValue],
+    earlier: ConfigureReceipt | None,
+    url_now: OwnedValue,
+) -> _Claim:
+    """What this configure records for one key; see ConfigureReceipt for the rules."""
+    before, after = _lookup(current, path), _lookup(merged, path)
+    carried: Final = earlier if earlier is not None and _ours(current, path, earlier) else None
+    return _Claim(
+        previous=before if carried is None else carried.previous.get(path, before),
+        written=_fingerprint(after) if before != after else (None if earlier is None else earlier.written.get(path)),
+        endpoint=None
+        if path not in _CREDENTIAL_PATHS
+        else (url_now if carried is None else carried.endpoints.get(path, url_now)),
+    )
+
+
+def _receipt(
+    current: Mapping[str, JsonValue],
+    merged: Mapping[str, JsonValue],
+    earlier: ConfigureReceipt | None,
+    file_exists: bool,
+) -> ConfigureReceipt:
+    url_now: Final = _lookup(current, _BASE_URL_PATH)
+    claims: Final = MappingProxyType({path: _claim(path, current, merged, earlier, url_now) for path in OWNED_PATHS})
+    return ConfigureReceipt(
+        file_existed=file_exists if earlier is None else earlier.file_existed,
+        env_present=ENV_KEY in current if earlier is None else earlier.env_present,
+        env_was_object=isinstance(current.get(ENV_KEY), dict) if earlier is None else earlier.env_was_object,
+        previous=MappingProxyType({path: claim.previous for path, claim in claims.items()}),
+        written=MappingProxyType({path: claim.written for path, claim in claims.items() if claim.written is not None}),
+        endpoints=MappingProxyType(
+            {path: claim.endpoint for path, claim in claims.items() if claim.endpoint is not None}
+        ),
+    )
+
+
+def read_configure_receipt(state_path: Path) -> ConfigureReceipt | None:
+    if not state_path.exists():
+        return None
+    try:
+        return ConfigureReceipt.model_validate_json(state_path.read_bytes())
+    except (OSError, ValidationError) as e:
+        raise ClaudeSettingsError(
+            f"{state_path} is not a readable `lite configure claude` receipt ({e}). "
+            "Remove it and edit Claude Code's settings by hand if they still point at the proxy."
+        ) from e
+
+
+def configure_claude_settings(
+    base_url: str,
+    credential: ClaudeCredential,
+    model: ModelChoice,
+    settings_path: Path,
+    state_path: Path,
+    owners: Sequence[SettingsFileOwner],
+    commit: Callable[[str, str], None] = commit_staged_json,
+) -> None:
+    """Persistently route Claude Code through base_url, recording how to undo it.
+
+    Both files are staged before either is committed, so a full disk or a read-only directory fails
+    before anything changes. The two commits are still two renames: a receipt rename that fails
+    discards the staged settings, and a settings rename that fails after the receipt landed puts the
+    earlier receipt back (or removes the new one), so the receipt on disk never describes settings
+    that were not written. `model`: StartOn pins the starting model, UnpinModel lets go of a pin an
+    earlier configure made (never of the user's own), KeepModel leaves it alone (a re-login).
+    """
+    refuse_while_owned(settings_path, owners)
+    current: Final = load_json_or_empty(settings_path)
+    _env_object(current, settings_path)
+    earlier: Final = read_configure_receipt(state_path)
+    existing: Final = (
+        _with(current, MODEL_KEY, earlier.previous[MODEL_KEY])
+        if isinstance(model, UnpinModel) and earlier is not None and _ours(current, MODEL_KEY, earlier)
+        else current
+    )
+    merged: Final = merge_claude_settings(
+        existing, base_url, credential, model.model if isinstance(model, StartOn) else None
+    )
+    receipt: Final = _receipt(current, merged, earlier, settings_path.exists())
+    target: Final = _write_target(settings_path)
+    try:
+        ensure_private_dir(state_path.parent)
     except OSError as e:
-        raise ClaudeSettingsError(f"Could not write {target}: {e}") from e
+        raise ClaudeSettingsError(f"Could not write {state_path}: {e}") from e
+    staged_receipt: Final = _stage(state_path, receipt.model_dump(mode="json"))
+    try:
+        staged_settings: Final = _stage(target, merged)
+    except ClaudeSettingsError:
+        discard_staged_json(staged_receipt)
+        raise
+    _land(state_path, staged_receipt, (staged_settings,), commit)
+    try:
+        _land(target, staged_settings, commit=commit)
+    except ClaudeSettingsError as settings_error:
+        try:
+            _land(state_path, None if earlier is None else _stage(state_path, earlier.model_dump(mode="json")))
+        except ClaudeSettingsError as receipt_error:
+            raise ClaudeSettingsError(
+                f"{settings_error} The receipt at {state_path} now describes settings that were not written and "
+                f"could not be put back either ({receipt_error}); remove it before retrying."
+            ) from settings_error
+        raise
+
+
+def _endpoint_text(endpoint: OwnedValue) -> str:
+    if not endpoint.present:
+        return f"no {ANTHROPIC_BASE_URL_KEY} (Anthropic's default endpoint)"
+    return endpoint.value if isinstance(endpoint.value, str) else json.dumps(endpoint.value)
+
+
+def unconfigure_claude_settings(
+    settings_path: Path, state_path: Path, owners: Sequence[SettingsFileOwner]
+) -> UnconfigureOutcome:
+    """Undo `lite configure claude`: put back every key still holding what configure wrote, leave the
+    rest alone, and withhold a credential the restored file would send to a different server than it
+    was issued for (the receipt stays, owning only those slots, so a later unconfigure can finish)."""
+    refuse_while_owned(settings_path, owners)
+    receipt: Final = read_configure_receipt(state_path)
+    if receipt is None:
+        raise ClaudeSettingsError(
+            f"Claude Code is not configured by `lite configure claude` (no receipt at {state_path}); nothing to undo."
+        )
+    current: Final = load_json_or_empty(settings_path)
+    _env_object(current, settings_path)
+    ours: Final = tuple(path for path in receipt.written if _ours(current, path, receipt))
+    kept: Final = tuple(path for path in receipt.written if path not in ours and _lookup(current, path).present)
+    put_back: Final = _with_all(current, MappingProxyType({path: receipt.previous[path] for path in ours}))
+    url_after: Final = _lookup(put_back, _BASE_URL_PATH)
+    withheld: Final = tuple(
+        WithheldCredential(path, _endpoint_text(receipt.endpoints[path]))
+        for path in _CREDENTIAL_PATHS
+        if path in ours and receipt.previous[path].present and receipt.endpoints[path] != url_after
+    )
+    absent: Final = OwnedValue(present=False)
+    trimmed: Final = _with_all(put_back, MappingProxyType({item.key: absent for item in withheld}))
+    settings: Final = (
+        trimmed
+        if _env(trimmed) or receipt.env_was_object
+        else _with_key(trimmed, ENV_KEY, OwnedValue(present=receipt.env_present, value=None))
+    )
+    target: Final = _write_target(settings_path)
+    file_removed: Final = not settings and not (receipt.file_existed and target.exists())
+    kept_receipt: Final = (  # mutable-ok: pydantic serializes the update as given and rejects a mappingproxy
+        receipt.model_copy(update={"written": {item.key: _fingerprint(absent) for item in withheld}})
+        if withheld
+        else None
+    )
+    staged_settings: Final = None if file_removed else _stage(target, settings)
+    try:
+        staged_receipt: Final = (
+            None if kept_receipt is None else _stage(state_path, kept_receipt.model_dump(mode="json"))
+        )
+    except ClaudeSettingsError:
+        if staged_settings is not None:
+            discard_staged_json(staged_settings)
+        raise
+    _land(target, staged_settings, (staged_receipt,))
+    _land(state_path, staged_receipt)
+    return UnconfigureOutcome(
+        restored=tuple(path for path in ours if _lookup(current, path) != _lookup(settings, path)),
+        kept=kept,
+        withheld=withheld,
+        file_removed=file_removed,
+    )
 
 
 __all__ = (
     "ANTHROPIC_API_KEY_KEY",
+    "ANTHROPIC_AUTH_TOKEN_KEY",
     "ANTHROPIC_BASE_URL_KEY",
+    "ANTHROPIC_DEFAULT_MODEL_ENV_KEYS",
     "API_KEY_HELPER_KEY",
     "AUTOROUTE_BACKUP_PATH",
     "BACKUP_PATH",
+    "CLAUDE_CONFIG_DIR_ENV",
     "CLAUDE_SETTINGS_PATH",
+    "CONFIGURE_STATE_PATH",
     "ENABLE_GATEWAY_MODEL_DISCOVERY_KEY",
     "ENABLE_GATEWAY_MODEL_DISCOVERY_VALUE",
     "ENABLE_TOOL_SEARCH_KEY",
     "ENABLE_TOOL_SEARCH_VALUE",
     "ENV_KEY",
+    "MODEL_KEY",
+    "OWNED_ENV_KEYS",
+    "OWNED_PATHS",
+    "OWNED_TOP_LEVEL_KEYS",
     "SETTINGS_FILE_OWNERS",
+    "STARTING_MODEL_ROLE",
+    "ApiKeyHelper",
+    "ClaudeCredential",
     "ClaudeSettingsError",
+    "ConfigureReceipt",
+    "KeepModel",
+    "ModelChoice",
+    "OwnedValue",
     "SettingsFileOwner",
+    "StartOn",
+    "StaticToken",
+    "UnconfigureOutcome",
+    "UnpinModel",
+    "WithheldCredential",
+    "claude_settings_path",
+    "configure_claude_settings",
+    "configure_state_path",
+    "lite_api_key_helper_configured",
     "load_json_or_empty",
     "merge_claude_settings",
+    "read_configure_receipt",
+    "refuse_while_owned",
     "resolve_api_key_helper",
-    "write_claude_settings",
+    "settings_file_owners",
+    "unconfigure_claude_settings",
 )

litellm/proxy/client/cli/commands/configure.py

@@ -0,0 +1,262 @@
+"""`lite configure claude` and `lite unconfigure claude`: persistent Claude Code wiring, undoable."""
+
+import os
+import re
+import sys
+from collections.abc import Callable, Sequence
+from pathlib import Path
+from typing import Final
+
+import click
+from InquirerPy import inquirer
+from InquirerPy.base.control import Choice
+
+from .auth import CliContextObj, context_secret_vault, get_stored_api_key
+from .claude_settings import (
+    STARTING_MODEL_ROLE,
+    ApiKeyHelper,
+    ClaudeCredential,
+    ClaudeSettingsError,
+    ModelChoice,
+    StartOn,
+    StaticToken,
+    UnconfigureOutcome,
+    UnpinModel,
+    claude_settings_path,
+    configure_claude_settings,
+    configure_state_path,
+    refuse_while_owned,
+    resolve_api_key_helper,
+    settings_file_owners,
+    unconfigure_claude_settings,
+)
+from .pi import ListingFailure, PiSyncError, fetch_model_ids
+from .up import ensure_fresh_login
+
+_LISTED_MODELS_SHOWN: Final = 20
+_CLAUDE_TARGET: Final = "claude"
+_TARGETS: Final = ((_CLAUDE_TARGET, "Claude Code (CLI)"),)
+_KEEP_DEFAULT_MODEL: Final = "Keep Claude Code's own default"
+_CLAUDE_CODE_PICKER_FILTER: Final = re.compile(r"claude|anthropic", re.IGNORECASE)
+_MODEL_OPTION_HELP: Final = (
+    f"Proxy model to set as {STARTING_MODEL_ROLE}. Must be listed on /v1/models for the key; without it, "
+    "Claude Code keeps its own default and a pin an earlier configure made is let go of. Nothing pins Claude "
+    "Code's sub-agent or background tiers; `lite autoroute up` is the mode that does."
+)
+
+
+def resolve_credential(ctx: click.Context, api_key: str | None) -> tuple[ClaudeCredential, str]:
+    """The credential to write and the key to check the proxy with.
+
+    An explicit key (--api-key, `lite --api-key`, LITELLM_PROXY_API_KEY) is long-lived and goes
+    into settings.json as a static token. Without one, the stored `lite login` credential is used
+    the way `lite login --config-claude` uses it, through apiKeyHelper, since it expires within a
+    day and renews in place there; a missing or stale login is refreshed first, as `lite up` does.
+    """
+    ctx_obj: Final[CliContextObj] = ctx.obj
+    explicit: Final = api_key or (None if ctx_obj.get("api_key_from_token_file") else ctx_obj.get("api_key"))
+    if explicit:
+        return StaticToken(explicit), explicit
+    base_url: Final = ctx_obj["base_url"]
+    ensure_fresh_login(ctx)
+    stored: Final = get_stored_api_key(expected_base_url=base_url, vault=context_secret_vault(ctx))
+    if not stored:
+        raise ClaudeSettingsError("Login did not produce a usable token.")
+    return ApiKeyHelper(resolve_api_key_helper(base_url)), stored
+
+
+def _start(ctx: click.Context, api_key: str | None) -> tuple[ClaudeCredential, tuple[str, ...]]:
+    """Every configure path begins the same way: the local ownership check first, so a `lite up`
+    session is refused before any login prompt or request, then the credential, then the listing."""
+    settings_path: Final = claude_settings_path(os.environ)
+    try:
+        refuse_while_owned(settings_path, settings_file_owners(settings_path))
+        credential, key = resolve_credential(ctx, api_key)
+    except ClaudeSettingsError as e:
+        raise click.ClickException(str(e))
+    return credential, _listed_models(ctx.obj["base_url"], key)
+
+
+def _listing_error(base_url: str, error: PiSyncError) -> str:
+    """The hint that fits how the listing failed: only an unreachable proxy gets the "is it running" question."""
+    if error.kind is ListingFailure.REJECTED:
+        return f"LiteLLM rejected your key (HTTP {error.status}). Run `lite login` to refresh it, or pass a valid --api-key."
+    if error.kind is ListingFailure.UNREACHABLE:
+        return f"{error.message} Is the proxy at {base_url} running, and is --base-url (or LITELLM_PROXY_URL) correct?"
+    if error.kind is ListingFailure.EMPTY:
+        return f"{error.message} Claude Code would have nothing to run; give the key access to at least one model."
+    return f"{error.message} The proxy at {base_url} answered, so check that it is a LiteLLM proxy and is healthy."
+
+
+def _listed_models(base_url: str, key: str) -> tuple[str, ...]:
+    listed: Final = fetch_model_ids(base_url, key)
+    if isinstance(listed, PiSyncError):
+        raise click.ClickException(_listing_error(base_url, listed))
+    return listed
+
+
+def _model_choice(model: str | None) -> ModelChoice:
+    return StartOn(model) if model is not None else UnpinModel()
+
+
+def _apply_claude(ctx: click.Context, credential: ClaudeCredential, listed: Sequence[str], model: str | None) -> None:
+    ctx_obj: Final[CliContextObj] = ctx.obj
+    base_url: Final = ctx_obj["base_url"]
+    if model is not None and model not in listed:
+        shown: Final = ", ".join(listed[:_LISTED_MODELS_SHOWN])
+        more: Final = f", and {len(listed) - _LISTED_MODELS_SHOWN} more" if len(listed) > _LISTED_MODELS_SHOWN else ""
+        raise click.ClickException(
+            f"{model!r} is not served by {base_url} for this key. /v1/models lists: {shown}{more}."
+        )
+    settings_path: Final = claude_settings_path(os.environ)
+    try:
+        configure_claude_settings(
+            base_url,
+            credential,
+            _model_choice(model),
+            settings_path,
+            configure_state_path(settings_path),
+            settings_file_owners(settings_path),
+        )
+    except ClaudeSettingsError as e:
+        raise click.ClickException(str(e))
+    in_picker: Final = sum(1 for listed_model in listed if _CLAUDE_CODE_PICKER_FILTER.search(listed_model))
+    click.echo(f"Configured Claude Code: {settings_path} now routes through {base_url}.")
+    click.echo(
+        "Credential: your virtual key, stored in the file as ANTHROPIC_AUTH_TOKEN."
+        if isinstance(credential, StaticToken)
+        else "Credential: your `lite login`, read through apiKeyHelper on every request, so a later login renews it."
+    )
+    click.echo(
+        f"Starting model: {model} ({STARTING_MODEL_ROLE}); switch any time with /model."
+        if model is not None
+        else "Starting model: not pinned (Claude Code's default, or a model you set yourself); switch with /model, or "
+        "pass --model to start on a proxy model."
+    )
+    click.echo(
+        f"/model will list {in_picker} of the proxy's {len(listed)} models (Claude Code shows only ids containing "
+        "'claude' or 'anthropic')."
+    )
+    click.echo("Start `claude` from any terminal. Undo with `lite unconfigure claude`.")
+    if isinstance(credential, StaticToken) and settings_path.is_symlink():
+        click.echo(
+            f"Note: {settings_path} is a symlink to {settings_path.resolve()}, so your key now lives in "
+            "that file; keep it out of version control.",
+            err=True,
+        )
+
+
+def _pick_targets() -> tuple[str, ...]:
+    picked: Final = inquirer.checkbox(
+        message="Which agents should route through LiteLLM?",
+        choices=[Choice(value, name=label, enabled=True) for value, label in _TARGETS],
+        validate=lambda chosen: len(chosen) > 0,
+        invalid_message="Pick at least one.",
+    ).execute()
+    return tuple(str(value) for value in picked)
+
+
+def _pick_model(listed: Sequence[str]) -> str | None:
+    picked: Final = inquirer.fuzzy(
+        message="Model Claude Code starts on (type to filter; /model switches any time):",
+        choices=[_KEEP_DEFAULT_MODEL, *listed],
+    ).execute()
+    return None if picked == _KEEP_DEFAULT_MODEL else str(picked)
+
+
+def interactive_configure(
+    ctx: click.Context,
+    pick_targets: Callable[[], tuple[str, ...]] = _pick_targets,
+    pick_model: Callable[[Sequence[str]], str | None] = _pick_model,
+) -> None:
+    """`lite configure` with no agent named: ask which agents to wire and which model to pin."""
+    targets: Final = pick_targets()
+    if _CLAUDE_TARGET not in targets:
+        return
+    credential, listed = _start(ctx, None)
+    _apply_claude(ctx, credential, listed, pick_model(listed))
+
+
+@click.group(name="configure", invoke_without_command=True)
+@click.pass_context
+def configure_group(ctx: click.Context) -> None:
+    """Persistently route a coding agent through your LiteLLM proxy.
+
+    With no agent named, asks which agents to wire and which proxy model to pin.
+    """
+    if ctx.invoked_subcommand is not None:
+        return
+    if not sys.stdin.isatty():
+        raise click.ClickException(
+            "`lite configure` asks questions, so it needs a terminal. Non-interactively, run "
+            "`lite configure claude --api-key <key> --model <model>`."
+        )
+    interactive_configure(ctx)
+
+
+@click.group(name="unconfigure")
+def unconfigure_group() -> None:
+    """Undo `lite configure` for a coding agent."""
+
+
+@configure_group.command(name="claude")
+@click.option(
+    "--api-key",
+    "api_key",
+    default=None,
+    help="Long-lived LiteLLM virtual key written into Claude Code's settings. Defaults to the `lite --api-key` / "
+    "LITELLM_PROXY_API_KEY value; with neither, your `lite login` credential is used through apiKeyHelper.",
+)
+@click.option("--model", default=None, help=_MODEL_OPTION_HELP)
+@click.pass_context
+def configure_claude(ctx: click.Context, api_key: str | None, model: str | None) -> None:
+    """Route every Claude Code session through your LiteLLM proxy until `lite unconfigure claude`.
+
+    Patches ~/.claude/settings.json in place: the proxy URL, your credential (a virtual key as a
+    static token, or your `lite login` through apiKeyHelper), and gateway model discovery so
+    /model lists the proxy's models; --model picks the one Claude Code starts on. Every other
+    setting is kept, and what changed is recorded so `lite unconfigure claude` can put it back.
+    Assumes the proxy is already running.
+    """
+    credential, listed = _start(ctx, api_key)
+    _apply_claude(ctx, credential, listed, model)
+
+
+@unconfigure_group.command(name="claude")
+def unconfigure_claude() -> None:
+    """Return Claude Code's settings to what they were before `lite configure claude`.
+
+    Also undoes `lite login --config-claude`. Only keys still holding what configure wrote are
+    put back; anything you changed since is left as it is and named in the output.
+    """
+    settings_path: Final = claude_settings_path(os.environ)
+    state_path: Final = configure_state_path(settings_path)
+    try:
+        outcome: Final = unconfigure_claude_settings(settings_path, state_path, settings_file_owners(settings_path))
+    except ClaudeSettingsError as e:
+        raise click.ClickException(str(e))
+    _report_unconfigure(settings_path, state_path, outcome)
+
+
+def _report_unconfigure(settings_path: Path, state_path: Path, outcome: UnconfigureOutcome) -> None:
+    """Say what unconfigure did, naming only keys whose value it changed."""
+    if outcome.file_removed:
+        click.echo(
+            f"No settings file remains at {settings_path}; it held nothing but `lite configure claude`'s own keys."
+        )
+    elif outcome.restored:
+        click.echo(f"Restored in {settings_path}: {', '.join(outcome.restored)}.")
+    else:
+        click.echo(f"Nothing in {settings_path} was still ours to restore.")
+    if outcome.kept:
+        click.echo(f"Left as you changed them since: {', '.join(outcome.kept)}.")
+    if outcome.withheld:
+        click.echo(
+            "Left removed, since the file now points at a different server than they were issued for: "
+            + "; ".join(f"{item.key} (captured with {item.endpoint})" for item in outcome.withheld)
+            + f". They stay in {state_path}: point env.ANTHROPIC_BASE_URL back and run `lite unconfigure claude` "
+            "again to put them back, or delete that file to drop them."
+        )
+
+
+__all__ = ("configure_group", "interactive_configure", "resolve_credential", "unconfigure_group")

litellm/proxy/client/cli/commands/pi.py

@@ -10,6 +10,7 @@ import os
 import tempfile
 from collections.abc import Callable, Mapping
 from dataclasses import dataclass
+from enum import StrEnum
 from pathlib import Path
 from types import MappingProxyType
 from typing import Final
@@ -20,11 +21,28 @@ from pydantic import BaseModel, JsonValue, TypeAdapter, ValidationError
 PI_CONFIG_DIR_ENV: Final = "PI_CODING_AGENT_DIR"
 PI_PROVIDER_NAME: Final = "litellm"
 LITELLM_PROXY_API_KEY_ENV: Final = "LITELLM_PROXY_API_KEY"
+_REJECTED_STATUSES: Final = frozenset((401, 403))
+
+
+class ListingFailure(StrEnum):
+    """Why a proxy could not be listed, decided once where the HTTP outcome is classified.
+
+    `unreachable` means no response at all; the other kinds prove the proxy answered, so callers
+    must not suggest checking whether it is running.
+    """
+
+    UNREACHABLE = "unreachable"
+    REJECTED = "rejected"
+    BAD_BODY = "bad_body"
+    EMPTY = "empty"
+    OTHER = "other"
 
 
 @dataclass(frozen=True, slots=True)
 class PiSyncError:
     message: str
+    status: int | None = None
+    kind: ListingFailure | None = None
 
 
 @dataclass(frozen=True, slots=True)
@@ -65,16 +83,20 @@ def fetch_model_ids(
             timeout=10,
         )
     except requests.RequestException as e:
-        return PiSyncError(f"Could not list models from the proxy: {e}")
+        return PiSyncError(f"Could not list models from the proxy: {e}", kind=ListingFailure.UNREACHABLE)
     if resp.status_code != 200:
-        return PiSyncError(f"The proxy returned HTTP {resp.status_code} for /v1/models; cannot build pi's model list.")
+        return PiSyncError(
+            f"The proxy returned HTTP {resp.status_code} for /v1/models; cannot list models.",
+            resp.status_code,
+            ListingFailure.REJECTED if resp.status_code in _REJECTED_STATUSES else ListingFailure.OTHER,
+        )
     try:
         listing: Final = _ModelList.model_validate(resp.json())
     except (ValueError, ValidationError) as e:
-        return PiSyncError(f"Unexpected /v1/models response from the proxy: {e}")
+        return PiSyncError(f"Unexpected /v1/models response from the proxy: {e}", kind=ListingFailure.BAD_BODY)
     ids: Final = tuple(dict.fromkeys(model.id for model in listing.data))
     if not ids:
-        return PiSyncError("The proxy returned no models for your key, so pi would have nothing to run.")
+        return PiSyncError("The proxy returned no models for your key.", kind=ListingFailure.EMPTY)
     return ids
 
 
@@ -200,6 +222,7 @@ __all__ = (
     "LITELLM_PROXY_API_KEY_ENV",
     "PI_CONFIG_DIR_ENV",
     "PI_PROVIDER_NAME",
+    "ListingFailure",
     "ModelLimits",
     "PiSyncError",
     "fetch_model_ids",

litellm/proxy/client/cli/commands/up.py

@@ -23,6 +23,7 @@ from .auth import CliContextObj, context_secret_vault, get_stored_api_key, load_
 from .claude_settings import (
     BACKUP_PATH,
     CLAUDE_SETTINGS_PATH,
+    ApiKeyHelper,
     ClaudeSettingsError,
     load_json_or_empty,
     merge_claude_settings,
@@ -123,7 +124,7 @@ def _stored_login_is_pkce(vault: SecretVault) -> bool:
     return token_data is not None and token_data.get("refresh_token") is not None
 
 
-def _ensure_fresh_login(ctx: click.Context) -> None:
+def ensure_fresh_login(ctx: click.Context) -> None:
     ctx_obj: Final[CliContextObj] = ctx.obj
     base_url: Final = ctx_obj["base_url"].rstrip("/")
     vault: Final = context_secret_vault(ctx)
@@ -141,7 +142,7 @@ def _ensure_fresh_login(ctx: click.Context) -> None:
     click.echo("No fresh LiteLLM login found for this proxy; starting login...")
     ctx.invoke(login, pkce=pkce)
     if not _usable_login(get_stored_api_key(expected_base_url=base_url, vault=vault), vault):
-        raise UpError("Login did not produce a usable token; cannot start `lite up`.")
+        raise UpError("Login did not produce a usable token.")
 
 
 def _restore_and_report() -> None:
@@ -169,7 +170,7 @@ def up(ctx: click.Context) -> None:
     base_url: Final = ctx.obj["base_url"]
 
     try:
-        _ensure_fresh_login(ctx)
+        ensure_fresh_login(ctx)
         api_key: Final = resolve_api_key(ctx)
         verify_proxy_key(base_url, api_key)
 
@@ -190,7 +191,7 @@ def up(ctx: click.Context) -> None:
         )
 
         CLAUDE_SETTINGS_PATH.parent.mkdir(exist_ok=True)
-        merged: Final = merge_claude_settings(original_settings, base_url, api_key_helper)
+        merged: Final = merge_claude_settings(original_settings, base_url, ApiKeyHelper(api_key_helper))
         with open(CLAUDE_SETTINGS_PATH, "w") as f:
             json.dump(merged, f, indent=2)
     except (AgentRunError, ClaudeSettingsError) as e:

litellm/proxy/client/cli/main.py

@@ -13,6 +13,7 @@ from .commands.auth import auth_group, context_secret_vault, get_stored_api_key,
 from .commands.autoroute.commands import autoroute_group
 from .commands.chat import chat
 from .commands.config import config_commands, get_config_value, hidden_command_names
+from .commands.configure import configure_group, unconfigure_group
 from .commands.credentials import credentials
 from .commands.debug import debug
 from .commands.encryption import encryption
@@ -162,6 +163,9 @@ cli.add_command(model_groups)
 # Add the autoroute command group (QA auto-routing against your real proxy)
 cli.add_command(autoroute_group, name="autoroute")
 cli.add_command(config_commands)
+# Add configure/unconfigure (persistently wire a coding agent to the proxy with a virtual key)
+cli.add_command(configure_group)
+cli.add_command(unconfigure_group)
 
 
 if __name__ == "__main__":

litellm/proxy/common_request_processing.py

@@ -190,6 +190,7 @@ from litellm.proxy.litellm_pre_call_utils import (
     refresh_proxy_server_request_body_snapshot,
     reject_url_valued_destination,
 )
+from litellm.proxy.policy_engine.response_retrieval import attach_post_call_pipelines_to_retrieval
 from litellm.types.utils import (
     ModelResponse,
     ModelResponseStream,
@@ -1849,7 +1850,6 @@ class ProxyBaseLLMRequestProcessing:
                 data=self.data,
                 user_api_key_dict=user_api_key_dict,
             )
-
         # Calculate request queue time after add_litellm_data_to_request
         # which sets arrival_time in proxy_server_request. Ends at start_time
         # (not a freshly captured time.time() here) so this window is exactly
@@ -1997,6 +1997,12 @@ class ProxyBaseLLMRequestProcessing:
             data=self.data,
             call_type=route_type,
         )
+        if route_type == "aget_responses":
+            attach_post_call_pipelines_to_retrieval(
+                data=self.data,
+                user_api_key_dict=user_api_key_dict,
+                llm_router=llm_router,
+            )
 
         # Refresh AFTER pre_call_hook: guardrails (e.g. Presidio PII masking) may
         # have mutated `self.data` in place, and the audit-trail snapshot taken in
@@ -3294,9 +3300,10 @@ class ProxyBaseLLMRequestProcessing:
         has completed.
 
         Guardrails routed through unified_guardrail are skipped, since they already ran
-        via its streaming iterator.  Guardrails that override
-        async_post_call_success_hook directly run here, including those that implement
-        apply_guardrail but keep their native lifecycle hooks.
+        via its streaming iterator, and so are guardrails a post_call policy pipeline
+        manages, since the pipeline ran them against the buffered stream.  Guardrails
+        that override async_post_call_success_hook directly run here, including those
+        that implement apply_guardrail but keep their native lifecycle hooks.
 
         This is audit-only — content has already been delivered to the client.
 
@@ -3306,12 +3313,18 @@ class ProxyBaseLLMRequestProcessing:
         _response = assembled_response
         try:
             from litellm.proxy.proxy_server import llm_router as _global_llm_router
-            from litellm.proxy.utils import _check_and_merge_model_level_guardrails
+            from litellm.proxy.utils import (
+                _check_and_merge_model_level_guardrails,
+                stream_gated_guardrail_names,
+            )
 
             guardrail_data = _check_and_merge_model_level_guardrails(data=captured_data, llm_router=_global_llm_router)
+            stream_gated: Final = stream_gated_guardrail_names(captured_data, captured_user_api_key_dict)
             for cb in litellm.callbacks:
                 if not isinstance(cb, CustomGuardrail):
                     continue
+                if cb.guardrail_name in stream_gated:
+                    continue
                 if not cb.should_run_guardrail(
                     data=guardrail_data,
                     event_type=GuardrailEventHooks.post_call,

litellm/proxy/common_utils/callback_utils.py

@@ -1,10 +1,12 @@
 import copy
+import json
 import os
 from collections.abc import Callable, Iterable, Mapping
 from dataclasses import dataclass
+from itertools import accumulate
 from typing import TYPE_CHECKING, Any, Final, Literal, NoReturn, Optional, TypeAlias
 
-from typing_extensions import assert_never
+from typing_extensions import ReadOnly, TypedDict, assert_never
 
 import litellm
 from litellm import get_secret
@@ -12,6 +14,7 @@ from litellm._logging import verbose_proxy_logger
 from litellm.constants import (
     CLIENT_OUTPUT_CEILING_METADATA_KEY,
     CONSUMED_REQUEST_TAGS_METADATA_KEY,
+    MAX_GUARDRAIL_SCAN_METADATA_HEADER_LENGTH,
     PRE_CALL_EXECUTED_GUARDRAILS_KEY,
     ROUTING_REQUEST_TAGS_METADATA_KEY,
     SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
@@ -28,6 +31,7 @@ from litellm.proxy.common_utils.encrypt_decrypt_utils import (
     encrypt_value_helper,
 )
 from litellm.proxy.types_utils.utils import get_instance_fn
+from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.utils import (
     StandardLoggingGuardrailInformation,
     StandardLoggingPayload,
@@ -52,6 +56,15 @@ reset_color_code: Final = "\033[0m"
 TRUSTED_PILLAR_RESPONSE_HEADERS_METADATA_KEY: Final = "_pillar_response_headers_trusted"
 
 GUARDRAIL_SCAN_IDS_METADATA_KEY: Final = "guardrail_scan_ids"
+GUARDRAIL_SCAN_METADATA_METADATA_KEY: Final = "guardrail_scan_metadata"
+
+
+class GuardrailScanMetadata(TypedDict):
+    guardrail: ReadOnly[str | None]
+    stage: ReadOnly[str]
+    provider: ReadOnly[str]
+    scan_id: ReadOnly[str]
+
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
@@ -450,6 +463,16 @@ def get_remaining_tokens_and_requests_from_request_data(data: dict) -> dict[str,
     return headers
 
 
+def _serialize_scan_metadata_header(entries: Iterable[object], *, max_length: int) -> str | None:
+    """Compact JSON list of scan metadata entries, dropping trailing entries so the header fits in max_length."""
+    encoded: Final = tuple(json.dumps(entry, separators=(",", ":")) for entry in entries)
+    lengths: Final = tuple(accumulate(len(item) + 1 for item in encoded))
+    kept: Final = sum(1 for length in lengths if length + 1 <= max_length)
+    if kept == 0:
+        return None
+    return f"[{','.join(encoded[:kept])}]"
+
+
 def get_logging_caching_headers(request_data: dict) -> dict | None:
     _metadata: Final[dict] = {}
     metadata_bucket: Final = request_data.get("metadata")
@@ -468,6 +491,15 @@ def get_logging_caching_headers(request_data: dict) -> dict | None:
     if scan_ids:
         headers["x-litellm-guardrail-scan-id"] = ",".join(scan_ids)
 
+    scan_metadata: Final = _metadata.get(GUARDRAIL_SCAN_METADATA_METADATA_KEY)
+    scan_metadata_header: Final = (
+        _serialize_scan_metadata_header(scan_metadata, max_length=MAX_GUARDRAIL_SCAN_METADATA_HEADER_LENGTH)
+        if isinstance(scan_metadata, (list, tuple))
+        else None
+    )
+    if scan_metadata_header:
+        headers["x-litellm-guardrail-scan-metadata"] = scan_metadata_header
+
     if "applied_policies" in _metadata:
         headers["x-litellm-applied-policies"] = ",".join(_metadata["applied_policies"])
 
@@ -501,6 +533,7 @@ LITELLM_PROXY_INTERNAL_METADATA_KEYS: Final = frozenset(
         "applied_policies",
         "applied_guardrails",
         GUARDRAIL_SCAN_IDS_METADATA_KEY,
+        GUARDRAIL_SCAN_METADATA_METADATA_KEY,
         "policy_sources",
         "guardrails",
         "guardrail_config",
@@ -565,21 +598,40 @@ def add_guardrail_to_applied_guardrails_header(request_data: dict, guardrail_nam
         _metadata["applied_guardrails"] = [guardrail_name]
 
 
-def add_guardrail_scan_id(request_data: dict, scan_id: str | None) -> None:
+def add_guardrail_scan_id(
+    request_data: dict[str, object],
+    scan_id: str | None,
+    *,
+    guardrail_name: str | None,
+    provider: str,
+    stage: GuardrailEventHooks,
+) -> None:
     """
-    Record a provider scan id so it can be surfaced to the caller.
+    Record a provider scan id, keyed to the guardrail execution that produced it, so it can be surfaced to the caller.
 
     Guardrails only return scan details to the client when they block, so allowed requests carry no
-    audit trail. Ids recorded here become the x-litellm-guardrail-scan-id response header.
+    audit trail. Ids recorded here become the x-litellm-guardrail-scan-id response header, and the
+    (guardrail, stage, provider, scan_id) entries become the x-litellm-guardrail-scan-metadata header.
     """
     if not scan_id:
         return
     _, _metadata = get_or_create_metadata_bucket(request_data)
     existing: Final = _metadata.get(GUARDRAIL_SCAN_IDS_METADATA_KEY)
-    scan_ids: Final = tuple(existing) if isinstance(existing, (list, tuple)) else ()
+    scan_ids: Final[tuple[object, ...]] = tuple(existing) if isinstance(existing, (list, tuple)) else ()
     if scan_id not in scan_ids:
         _metadata[GUARDRAIL_SCAN_IDS_METADATA_KEY] = (*scan_ids, scan_id)
 
+    entry: Final[GuardrailScanMetadata] = {
+        "guardrail": guardrail_name,
+        "stage": stage.value,
+        "provider": provider,
+        "scan_id": scan_id,
+    }
+    existing_entries: Final = _metadata.get(GUARDRAIL_SCAN_METADATA_METADATA_KEY)
+    entries: Final[tuple[object, ...]] = tuple(existing_entries) if isinstance(existing_entries, (list, tuple)) else ()
+    if entry not in entries:
+        _metadata[GUARDRAIL_SCAN_METADATA_METADATA_KEY] = (*entries, entry)
+
 
 def add_policy_to_applied_policies_header(request_data: dict, policy_name: str | None):
     """

litellm/proxy/db/db_url_settings.py

@@ -34,12 +34,20 @@ writer's connection params (pool size, timeouts, pgbouncer mode) for the
 ones the reader URL does not pin itself.
 """
 
+import _ssl
+import hashlib
 import os
+import socket
+import ssl
+import struct
+import sys
+import tempfile
 import urllib.parse
-from collections.abc import Mapping
+from collections.abc import Callable, Mapping, Sequence
 from functools import partial
+from pathlib import Path
 from types import MappingProxyType
-from typing import Annotated, Final, cast
+from typing import Annotated, Final, Protocol, TypeAlias, cast
 
 from pydantic import AliasChoices, BeforeValidator, Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
@@ -126,21 +134,100 @@ def add_missing_query_params(url: str, params: Mapping[str, str | int | float])
 
 
 LIBPQ_VERIFY_SSLMODES: Final[frozenset[str]] = frozenset({"verify-ca", "verify-full"})
+PEM_CERT_HEADER: Final = b"-----BEGIN CERTIFICATE-----"
+PG_SSL_REQUEST: Final = struct.pack("!ii", 8, 80877103)
+TLS_PROBE_TIMEOUT_SECONDS: Final = 10.0
+
+RootCertResolver: TypeAlias = Callable[[str, str, int], str]  # mutable-ok: Callable parameter syntax
 
 
-def translate_libpq_ssl_params(url: str) -> str:
+class _VerifiedChainSource(Protocol):
+    def get_verified_chain(self) -> Sequence[_ssl.Certificate] | None: ...
+
+
+def _verified_chain_der(tls: ssl.SSLSocket) -> tuple[bytes, ...]:
+    if sys.version_info >= (3, 13):
+        return tuple(tls.get_verified_chain())
+    legacy: Final = cast(  # cast-ok: the stub omits _sslobj, the C object has get_verified_chain since 3.10
+        "_VerifiedChainSource | None",
+        tls._sslobj,  # pyright: ignore[reportAttributeAccessIssue, reportUnknownMemberType]  # public API only from 3.13
+    )
+    chain: Final = () if legacy is None else legacy.get_verified_chain() or ()
+    return tuple(cert.public_bytes(_ssl.ENCODING_DER) for cert in chain)
+
+
+def _server_trust_anchor(cafile: str, host: str, port: int) -> bytes | None:
+    try:
+        context: Final = ssl.create_default_context(cafile=cafile)
+        with socket.create_connection((host, port), timeout=TLS_PROBE_TIMEOUT_SECONDS) as raw:
+            raw.sendall(PG_SSL_REQUEST)
+            if raw.recv(1) != b"S":
+                return None
+            with context.wrap_socket(raw, server_hostname=host) as tls:
+                chain: Final = _verified_chain_der(tls)
+    except (OSError, ValueError):
+        return None
+    return chain[-1] if chain else None
+
+
+def pin_bundle_root(cert_path: str, host: str, port: int) -> str:
+    """Reduce a multi-root CA bundle to the one root that verifies ``host``.
+
+    Prisma's ``sslcert`` loads a single PEM certificate (native-tls
+    ``Certificate::from_pem``), so pointing it at a bundle such as the AWS RDS
+    global bundle trusts only the first of its 108 regional roots and the
+    handshake fails with "unable to get local issuer certificate" for every
+    other region. A single-certificate file is returned as is. For a bundle,
+    one verifying handshake (chain and hostname, whole bundle as trust store)
+    identifies the trust anchor the server actually chains to, which is
+    written to a single-certificate file for Prisma. If the probe fails the
+    bundle path is returned unchanged, so Prisma fails closed exactly as
+    before rather than trusting anything the bundle would not.
+    """
+    try:
+        if Path(cert_path).read_bytes().count(PEM_CERT_HEADER) < 2:
+            return cert_path
+    except OSError:
+        return cert_path
+    root: Final = _server_trust_anchor(cert_path, host, port)
+    if root is None:
+        return cert_path
+    pinned: Final = Path(tempfile.gettempdir()) / f"litellm-sslcert-{hashlib.sha256(root).hexdigest()[:16]}.pem"
+    return str(pinned) if _replace_file(pinned, ssl.DER_cert_to_PEM_cert(root)) else cert_path
+
+
+def _replace_file(target: Path, content: str) -> bool:
+    """Write ``content`` to a private temp file and rename it over ``target``, so
+    readers never see a partial file and a symlink planted at ``target`` is
+    replaced rather than followed."""
+    try:
+        fd, staged = tempfile.mkstemp(dir=target.parent, prefix=f"{target.name}.")
+    except OSError:
+        return False
+    try:
+        with os.fdopen(fd, "w") as handle:
+            handle.write(content)
+        os.replace(staged, target)
+    except OSError:
+        Path(staged).unlink(missing_ok=True)
+        return False
+    return True
+
+
+def translate_libpq_ssl_params(url: str, resolve_root_cert: RootCertResolver = pin_bundle_root) -> str:
     """Rewrite libpq's certificate-verification params into Prisma's dialect.
 
     Prisma's engine only knows ``sslmode=disable|prefer|require``, ``sslcert``
-    (the CA bundle) and ``sslaccept=strict``. It silently discards
+    (a single CA certificate) and ``sslaccept=strict``. It silently discards
     ``sslrootcert`` and downgrades ``sslmode=verify-ca`` / ``verify-full`` to
     ``prefer``, so a URL copied from libpq / RDS docs connects over TLS with no
     certificate check at all. ``verify-ca`` and ``verify-full`` both become
     ``require`` (Prisma has no CA-only mode), ``sslrootcert`` becomes
-    ``sslcert``, and either one turns on ``sslaccept=strict`` (chain and
-    hostname), matching libpq where a root cert makes ``require`` verify.
-    Prisma params the operator pinned themselves win; anything else is left
-    untouched.
+    ``sslcert`` (run through ``resolve_root_cert``, which pins a multi-root
+    bundle down to the server's root), and either one turns on
+    ``sslaccept=strict`` (chain and hostname), matching libpq where a root
+    cert makes ``require`` verify. Prisma params the operator pinned
+    themselves win; anything else is left untouched.
     """
     parsed: Final = urllib.parse.urlsplit(url)
     pairs: Final = tuple(urllib.parse.parse_qsl(parsed.query, keep_blank_values=True))
@@ -154,7 +241,9 @@ def translate_libpq_ssl_params(url: str) -> str:
         if key != "sslrootcert"
     )
     root_cert: Final = tuple(
-        ("sslcert", value) for key, value in pairs if key == "sslrootcert" and "sslcert" not in keys
+        ("sslcert", resolve_root_cert(value, parsed.hostname or "", parsed.port or int(DEFAULT_POSTGRES_PORT)))
+        for key, value in pairs
+        if key == "sslrootcert" and "sslcert" not in keys
     )
     strict: Final = () if "sslaccept" in keys else (("sslaccept", "strict"),)
     query: Final = urllib.parse.urlencode(translated + root_cert + strict)

litellm/proxy/db/gateway_request_tracking.py

@@ -10,13 +10,28 @@ strings rather than passing the raw path through. Nothing a caller sends can
 add a key, so the fold and the table it commits to are bounded by (days x
 routes) however much traffic arrives, and the response path carries no
 unbounded queue that would block once full.
+
+A flush commits its whole snapshot as one multi-row ``INSERT ... ON CONFLICT DO
+UPDATE`` rather than one upsert per key, so a worker costs the primary one
+statement per interval however many routes it served. With
+``use_redis_transaction_buffer`` on, workers instead push their snapshot to a
+Redis list and one lock-holding pod folds every entry and writes the table, so
+the deployment as a whole costs the primary one statement per interval.
 """
 
-from dataclasses import asdict
+import json
+from collections.abc import AsyncIterator, Iterable
 from datetime import datetime, timezone
-from typing import TYPE_CHECKING, Final
+from itertools import chain
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final, TypeAlias
+
+from pydantic import TypeAdapter
 
 from litellm._logging import verbose_proxy_logger
+from litellm.caching import RedisCache
+from litellm.constants import MAX_REDIS_BUFFER_DEQUEUE_COUNT, REDIS_GATEWAY_REQUESTS_BUFFER_KEY
+from litellm.proxy.db.db_transaction_queue.pod_lock_manager import PodLockManager
 from litellm.proxy.middleware.billable_request_metrics_middleware import BillableCategory
 from litellm.types.proxy.gateway_requests import (
     GatewayRequestCounts,
@@ -28,6 +43,15 @@ if TYPE_CHECKING:
     from litellm.proxy.utils import PrismaClient
 
 _EMPTY: Final = GatewayRequestCounts(successful_requests=0, failed_requests=0)
+_TABLE: Final = '"LiteLLM_DailyGatewayRequests"'
+_COLUMNS_PER_ROW: Final = 5
+_UTC_NOW: Final = "(NOW() AT TIME ZONE 'UTC')"
+GATEWAY_REQUESTS_JOB_NAME: Final = "update_gateway_requests_job"
+
+_BufferedRows: TypeAlias = tuple[tuple[str, str, str, int, int], ...]
+_BUFFERED_ROWS: Final = TypeAdapter(_BufferedRows)
+_BUFFERED_ENTRIES: Final = TypeAdapter(tuple[str | bytes, ...])
+_NO_COUNTS: Final[GatewayRequestSnapshot] = MappingProxyType({})
 
 
 def _utc_date() -> str:
@@ -59,20 +83,54 @@ class GatewayRequestAccumulator:
         route) however long the database is unreachable.
 
         This buys at-least-once, not exactly-once, and the cost is worth stating.
-        The batch commits inside its context manager's ``__aexit__``, so a failure
-        raised after the transaction committed (a connection dropped while reading
-        the acknowledgement) restores counts that are already persisted, and the
-        next flush increments them a second time. Exactly-once would need a dedup
-        key the upserts could ignore on replay. For a traffic-volume metric a rare
+        The statement commits on the server before its acknowledgement is read, so
+        a failure raised after the commit (a connection dropped while reading the
+        acknowledgement) restores counts that are already persisted, and the next
+        flush increments them a second time. Exactly-once would need a dedup key
+        the upsert could ignore on replay. For a traffic-volume metric a rare
         overcount on a dropped acknowledgement beats losing a whole interval to
         every database blip, so the trade is deliberate.
         """
-        for key, counts in snapshot.items():
-            existing = self._counts.get(key, _EMPTY)
-            self._counts[key] = GatewayRequestCounts(
-                successful_requests=existing.successful_requests + counts.successful_requests,
-                failed_requests=existing.failed_requests + counts.failed_requests,
-            )
+        self._counts = dict(fold_counts(chain(self._counts.items(), snapshot.items())))  # mutable-ok: fold replaced
+
+
+def fold_counts(items: Iterable[tuple[GatewayRequestKey, GatewayRequestCounts]]) -> GatewayRequestSnapshot:
+    """Sum counts key-wise; the result stays bounded by (date x category x route)."""
+    folded: Final[dict[GatewayRequestKey, GatewayRequestCounts]] = {}  # mutable-ok: local fold returned once
+    for key, counts in items:
+        existing = folded.get(key, _EMPTY)
+        folded[key] = GatewayRequestCounts(
+            successful_requests=existing.successful_requests + counts.successful_requests,
+            failed_requests=existing.failed_requests + counts.failed_requests,
+        )
+    return folded
+
+
+def build_gateway_requests_upsert(snapshot: GatewayRequestSnapshot) -> tuple[str, tuple[str | int, ...]]:
+    """
+    One ``INSERT ... ON CONFLICT DO UPDATE`` that increments every (date, category,
+    route) in the snapshot. Rows are ordered by the conflict key so concurrent
+    writers lock rows in the same order and cannot deadlock.
+    """
+    ordered: Final = sorted(snapshot.items(), key=lambda item: (item[0].date, item[0].category, item[0].route))
+    rows: Final = ", ".join(
+        f"(${base + 1}::text, ${base + 2}::text, ${base + 3}::text, ${base + 4}::bigint, ${base + 5}::bigint, {_UTC_NOW})"
+        for base in range(0, len(ordered) * _COLUMNS_PER_ROW, _COLUMNS_PER_ROW)
+    )
+    sql: Final = (
+        f'INSERT INTO {_TABLE} ("date", "category", "route", "successful_requests", "failed_requests", "updated_at")\n'
+        f"VALUES {rows}\n"
+        'ON CONFLICT ("date", "category", "route") DO UPDATE SET\n'
+        f'  "successful_requests" = {_TABLE}."successful_requests" + EXCLUDED."successful_requests",\n'
+        f'  "failed_requests" = {_TABLE}."failed_requests" + EXCLUDED."failed_requests",\n'
+        f'  "updated_at" = {_UTC_NOW}'
+    )
+    params: Final[tuple[str | int, ...]] = tuple(
+        value
+        for key, counts in ordered
+        for value in (key.date, key.category, key.route, counts.successful_requests, counts.failed_requests)
+    )
+    return sql, params
 
 
 async def commit_gateway_requests_to_db(
@@ -80,50 +138,130 @@ async def commit_gateway_requests_to_db(
     prisma_client: "PrismaClient",
     snapshot: GatewayRequestSnapshot,
 ) -> None:
-    """Upsert one incrementing row per (date, category, route)."""
+    """Increment every (date, category, route) in the snapshot with a single statement."""
     if not snapshot:
         return
 
-    ordered: Final = sorted(snapshot.items(), key=lambda item: (item[0].date, item[0].category, item[0].route))
+    sql, params = build_gateway_requests_upsert(snapshot)
+    await prisma_client.db.execute_raw(sql, *params)  # pyright: ignore[reportAny]  # untyped prisma client
 
-    # pyright: ignore[reportAny] on both lines -- prisma's generated client is untyped,
-    # so .db and every table action off it resolve to Any at this boundary. The dict
-    # literals below are the shape prisma's generated inputs require.
-    async with prisma_client.db.batch_() as batcher:  # pyright: ignore[reportAny]  # untyped prisma client
-        for key, counts in ordered:
-            columns = asdict(key)
-            batcher.litellm_dailygatewayrequests.upsert(  # pyright: ignore[reportAny]  # untyped prisma client
-                where={"date_category_route": columns},  # mutable-ok: prisma input is dict-shaped
-                data={  # mutable-ok: prisma input is dict-shaped
-                    "create": {  # mutable-ok: prisma input is dict-shaped
-                        **columns,
-                        "successful_requests": counts.successful_requests,
-                        "failed_requests": counts.failed_requests,
-                    },
-                    "update": {  # mutable-ok: prisma input is dict-shaped
-                        "successful_requests": {"increment": counts.successful_requests},  # mutable-ok: as above
-                        "failed_requests": {"increment": counts.failed_requests},  # mutable-ok: as above
-                    },
-                },
+    verbose_proxy_logger.debug(
+        "Gateway request tracking - committed %d aggregated rows in one statement", len(snapshot)
+    )
+
+
+class GatewayRequestRedisBuffer:
+    """
+    Folds every worker's snapshot through one Redis list so a single pod per
+    interval writes the table, mirroring the spend writer's transaction buffer.
+
+    Each entry is one worker's snapshot as JSON rows; the lock holder pops them,
+    sums them, and commits one statement. A commit failure pushes the summed
+    rows back so the next holder retries, keeping the at-least-once guarantee.
+    If that push fails too, the rows go back to the holder's own accumulator so
+    they ride along with its next flush instead of vanishing with the pop.
+    """
+
+    def __init__(self, *, redis_cache: RedisCache, pod_lock_manager: PodLockManager) -> None:
+        self._redis_cache: Final = redis_cache
+        self._pod_lock_manager: Final = pod_lock_manager
+
+    async def push(self, snapshot: GatewayRequestSnapshot) -> None:
+        if not snapshot:
+            return
+        rows: Final[_BufferedRows] = tuple(
+            (key.date, key.category, key.route, counts.successful_requests, counts.failed_requests)
+            for key, counts in snapshot.items()
+        )
+        await self._redis_cache.async_rpush(key=REDIS_GATEWAY_REQUESTS_BUFFER_KEY, values=(json.dumps(rows),))
+
+    async def _pop_batch(self) -> tuple[str | bytes, ...]:
+        popped: Final[object] = await self._redis_cache.async_lpop(  # pyright: ignore[reportAny]  # redis returns Any
+            key=REDIS_GATEWAY_REQUESTS_BUFFER_KEY, count=MAX_REDIS_BUFFER_DEQUEUE_COUNT
+        )
+        if not popped:
+            return ()
+        return _BUFFERED_ENTRIES.validate_python(popped if isinstance(popped, list) else (popped,))
+
+    async def _pop_all(self) -> AsyncIterator[str | bytes]:
+        while True:
+            batch = await self._pop_batch()
+            for entry in batch:
+                yield entry
+            if len(batch) < MAX_REDIS_BUFFER_DEQUEUE_COUNT:
+                return
+
+    async def pop(self) -> GatewayRequestSnapshot:
+        entries: Final = tuple([entry async for entry in self._pop_all()])
+        return fold_counts(
+            (
+                GatewayRequestKey(date=date, category=category, route=route),
+                GatewayRequestCounts(successful_requests=succeeded, failed_requests=failed),
             )
+            for entry in entries
+            for date, category, route, succeeded, failed in _BUFFERED_ROWS.validate_json(entry)
+        )
 
-    verbose_proxy_logger.debug("Gateway request tracking - committed %d aggregated rows", len(ordered))
+    async def commit_if_leader(self, prisma_client: "PrismaClient") -> GatewayRequestSnapshot:
+        """
+        Drain the list and write it as one statement, but only on the pod holding the job lock.
+
+        The lock is a lease, never released: the holder re-enters it on every flush and
+        keeps committing alone until the TTL lapses, so the primary sees one statement
+        per flush interval deployment-wide instead of one per worker.
+
+        Returns the popped rows that could be neither committed nor re-queued, for the
+        caller to keep in memory. Empty on success.
+        """
+        if not await self._pod_lock_manager.acquire_lock(cronjob_id=GATEWAY_REQUESTS_JOB_NAME):
+            return _NO_COUNTS
+        buffered: Final = await self.pop()
+        try:
+            await commit_gateway_requests_to_db(prisma_client=prisma_client, snapshot=buffered)
+        except Exception:  # noqa: BLE001 -- a failed commit must not stop the scheduler
+            verbose_proxy_logger.warning(
+                "Gateway request tracking - failed to commit %d buffered rows, re-queuing to Redis for the next flush",
+                len(buffered),
+                exc_info=True,
+            )
+            return await self._requeue(buffered)
+        return _NO_COUNTS
+
+    async def _requeue(self, snapshot: GatewayRequestSnapshot) -> GatewayRequestSnapshot:
+        try:
+            await self.push(snapshot)
+        except Exception:  # noqa: BLE001 -- the rows go back to the caller's accumulator instead
+            verbose_proxy_logger.warning(
+                "Gateway request tracking - Redis re-queue failed, keeping %d rows in memory for the next flush",
+                len(snapshot),
+                exc_info=True,
+            )
+            return snapshot
+        return _NO_COUNTS
 
 
 async def flush_gateway_requests(
     prisma_client: "PrismaClient",
     accumulator: GatewayRequestAccumulator,
+    redis_buffer: GatewayRequestRedisBuffer | None = None,
 ) -> None:
     """
     Scheduler entrypoint. Never raises: a metering failure must not kill the job.
 
+    With ``redis_buffer`` the snapshot goes to Redis and only the lease holder
+    writes to Postgres. Shutdown passes no buffer so a departing worker writes its
+    own counts directly instead of parking them behind a lease it may not hold.
+
     ``CancelledError`` is deliberately not caught, so a flush cancelled during
     shutdown drops its snapshot rather than restoring counts onto an accumulator
     the process is about to discard.
     """
     snapshot: Final = accumulator.drain()
     try:
-        await commit_gateway_requests_to_db(prisma_client=prisma_client, snapshot=snapshot)
+        if redis_buffer is None:
+            await commit_gateway_requests_to_db(prisma_client=prisma_client, snapshot=snapshot)
+        else:
+            await redis_buffer.push(snapshot)
     except Exception:  # noqa: BLE001 -- a failed flush must not stop the scheduler
         accumulator.restore(snapshot)
         verbose_proxy_logger.warning(
@@ -131,3 +269,13 @@ async def flush_gateway_requests(
             len(snapshot),
             exc_info=True,
         )
+        return
+    if redis_buffer is None:
+        return
+    try:
+        accumulator.restore(await redis_buffer.commit_if_leader(prisma_client))
+    except Exception:  # noqa: BLE001 -- entries still in Redis are drained by the next flush
+        verbose_proxy_logger.warning(
+            "Gateway request tracking - leader drain failed, buffered rows stay in Redis for the next flush",
+            exc_info=True,
+        )

litellm/proxy/guardrails/auto_router_compression.py

@@ -80,17 +80,19 @@ def policy_from_litellm_params(litellm_params: Mapping[str, object]) -> AutoRout
 def policy_for_model(
     llm_router: "Router | None",
     model_alias: str,
-    team_id: str | None,
+    request_kwargs: Mapping[str, object],
     request_tags: Sequence[str],
 ) -> AutoRouterCompressionPolicy | None:
-    """The compression policy of the auto router marker `model_alias` resolves to.
+    """The compression policy of the auto router marker `model_alias` resolves to for this caller.
 
-    Pre-call arming and the routing hook both resolve through here, so an alias with
-    several tag-scoped markers cannot suppress under one and then route under another.
+    Pre-call arming and the routing hook both resolve through here, and here resolves through the
+    router's own request-scoped deployment lookup, so an alias with several tag-scoped markers
+    cannot suppress under one and then route under another, and a team router reached by its
+    public name carries its policy for every principal that can reach it.
     """
     if llm_router is None:
         return None
-    deployments: Final = llm_router.get_model_list(model_name=model_alias, team_id=team_id) or ()
+    deployments: Final = llm_router.deployments_for_request(model_alias, request_kwargs)
     markers: Final = tuple(
         litellm_params
         for deployment in deployments
@@ -108,17 +110,6 @@ def policy_for_model(
     return next((policy for policy in candidates if policy is not None), None)
 
 
-def team_id_from_request(request_kwargs: Mapping[str, object]) -> str | None:
-    """The caller's team id, from whichever metadata bucket this surface writes to."""
-    for meta_key in ("metadata", "litellm_metadata"):
-        meta = request_kwargs.get(meta_key)
-        if isinstance(meta, Mapping):
-            team_id = meta.get("user_api_key_team_id")
-            if isinstance(team_id, str):
-                return team_id
-    return None
-
-
 def _compression_guardrail_classes() -> tuple[type, ...]:
     """The registered guardrail classes whose provider compresses prompts."""
     from litellm.proxy.guardrails.guardrail_registry import guardrail_class_registry
@@ -172,7 +163,7 @@ async def arm_pre_call(
     policy: Final = policy_for_model(
         llm_router=llm_router,
         model_alias=model_alias,
-        team_id=team_id_from_request(data),
+        request_kwargs=data,
         request_tags=_get_tags_from_request_kwargs(data),
     )
     if policy is None:

litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py

@@ -44,7 +44,7 @@ from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicM
 from litellm.llms.base_llm.guardrail_translation.utils import (
     effective_scan_only_tool_results_for_guardrail,
 )
-from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, bedrock_bearer_token
+from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, bedrock_bearer_token, run_aws_signing
 from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
@@ -917,7 +917,9 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
                 source,
             )
             return BedrockGuardrailResponse()
-        credentials, aws_region_name = self._load_credentials(bearer_token=bedrock_bearer_token(api_key))
+        credentials, aws_region_name = await run_aws_signing(
+            self._load_credentials, bearer_token=bedrock_bearer_token(api_key)
+        )
         allow_chunking: Final = not self._content_uses_contextual_grounding(content)
 
         completed_chunk_usages: Final[list[BedrockGuardrailUsage]] = []  # mutable-ok: billed-chunk usage accumulator
@@ -1178,7 +1180,8 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
             **base_request_data,
             "content": content,
         }  # mutable-ok: outbound JSON request body
-        prepared_request: Final = self._prepare_request(
+        prepared_request: Final = await run_aws_signing(
+            self._prepare_request,
             credentials=credentials,
             data=bedrock_request_data,
             optional_params=self.optional_params,
@@ -1875,10 +1878,13 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
             return BedrockGuardrailResponse()
 
         api_key: Final[str | None] = request_data.get("api_key") if request_data else None
-        credentials, aws_region_name = self._load_credentials(bearer_token=bedrock_bearer_token(api_key))
+        credentials, aws_region_name = await run_aws_signing(
+            self._load_credentials, bearer_token=bedrock_bearer_token(api_key)
+        )
         body: Final[dict[str, object]] = {"messages": checks_messages, "checks": self.checks}
 
-        prepared_request: Final = self._prepare_request(
+        prepared_request: Final = await run_aws_signing(
+            self._prepare_request,
             credentials=credentials,
             data=body,
             optional_params=self.optional_params,

litellm/proxy/guardrails/guardrail_hooks/openai/moderations.py

@@ -17,7 +17,8 @@ from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
 )
-from litellm.types.guardrails import GuardrailEventHooks
+from litellm.proxy.common_utils.callback_utils import add_guardrail_scan_id
+from litellm.types.guardrails import GuardrailEventHooks, SupportedGuardrailIntegrations
 from litellm.types.utils import (
     GenericGuardrailAPIInputs,
     GuardrailStatus,
@@ -218,6 +219,13 @@ class OpenAIModerationGuardrail(OpenAIGuardrailBase, CustomGuardrail):
             metadata: Final = request_data.get("metadata") or {}
             request_data["metadata"] = metadata
             metadata["_openai_moderation_response"] = moderation_response.model_dump()
+            add_guardrail_scan_id(
+                request_data=request_data,
+                scan_id=moderation_response.id,
+                guardrail_name=self.guardrail_name,
+                provider=SupportedGuardrailIntegrations.OPENAI_MODERATION.value,
+                stage=GuardrailEventHooks.post_call if input_type == "response" else GuardrailEventHooks.pre_call,
+            )
 
         # Check if content is flagged and raise exception if needed
         self._check_moderation_result(moderation_response)

litellm/proxy/guardrails/guardrail_hooks/panw_prisma_airs/panw_prisma_airs.py

@@ -721,10 +721,18 @@ class PanwPrismaAirsHandler(CustomGuardrail):
             }
         }
 
-    def _record_scan_id(self, request_data: dict[str, object], scan_result: Mapping[str, object]) -> None:
+    def _record_scan_id(
+        self, request_data: dict[str, object], scan_result: Mapping[str, object], stage: GuardrailEventHooks
+    ) -> None:
         """Surface the AIRS scan id on the response, so allowed calls are auditable too."""
         scan_id: Final = scan_result.get("scan_id")
-        add_guardrail_scan_id(request_data=request_data, scan_id=str(scan_id) if scan_id else None)
+        add_guardrail_scan_id(
+            request_data=request_data,
+            scan_id=str(scan_id) if scan_id else None,
+            guardrail_name=self.guardrail_name,
+            provider=self._PROVIDER_NAME,
+            stage=stage,
+        )
 
     def _handle_api_error_with_logging(
         self,
@@ -948,7 +956,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
             event_type=GuardrailEventHooks.post_call,
         )
         add_guardrail_to_applied_guardrails_header(request_data=request_data, guardrail_name=self.guardrail_name)
-        self._record_scan_id(request_data, scan_result)
+        self._record_scan_id(request_data, scan_result, GuardrailEventHooks.post_call)
 
     def _check_and_mark_scanned(self, data: dict, scan_type: str) -> bool:
         """
@@ -1078,7 +1086,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                 duration=(end_time - start_time).total_seconds(),
                 event_type=GuardrailEventHooks.pre_call,
             )
-            self._record_scan_id(data, scan_result)
+            self._record_scan_id(data, scan_result, GuardrailEventHooks.pre_call)
 
             action: Final = scan_result.get("action", "block")
             category: Final = scan_result.get("category", "unknown")
@@ -1199,7 +1207,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                 duration=(end_time - start_time).total_seconds(),
                 event_type=GuardrailEventHooks.post_call,
             )
-            self._record_scan_id(data, scan_result)
+            self._record_scan_id(data, scan_result, GuardrailEventHooks.post_call)
 
             action: Final = scan_result.get("action", "block")
             category: Final = scan_result.get("category", "unknown")
@@ -1401,7 +1409,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                     duration=(end_time - start_time).total_seconds(),
                     event_type=GuardrailEventHooks.post_call,
                 )
-                self._record_scan_id(request_data, scan_result)
+                self._record_scan_id(request_data, scan_result, GuardrailEventHooks.post_call)
 
                 # Add guardrail to applied guardrails header for observability
                 add_guardrail_to_applied_guardrails_header(
@@ -1475,7 +1483,11 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                 )
                 continue
 
-            self._record_scan_id(request_data, scan_result)
+            self._record_scan_id(
+                request_data,
+                scan_result,
+                GuardrailEventHooks.post_call if is_response else GuardrailEventHooks.pre_call,
+            )
 
             action = scan_result.get("action", "block")
             masked_args = self._masked_tool_call_arguments(
@@ -1829,7 +1841,11 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                 new_texts.append(text)
                 continue
 
-            self._record_scan_id(request_data, scan_result)
+            self._record_scan_id(
+                request_data,
+                scan_result,
+                GuardrailEventHooks.post_call if is_response else GuardrailEventHooks.pre_call,
+            )
 
             action = scan_result.get("action", "block")
             masked_text = self._get_masked_text(scan_result, is_response=is_response)
@@ -1901,7 +1917,7 @@ class PanwPrismaAirsHandler(CustomGuardrail):
                 )
                 # If we reach here, fallback_on_error="allow"
             else:
-                self._record_scan_id(request_data, mcp_scan_result)
+                self._record_scan_id(request_data, mcp_scan_result, GuardrailEventHooks.pre_call)
                 action = mcp_scan_result.get("action", "block")
                 masked_text = self._get_masked_text(mcp_scan_result, is_response=False)
                 if action == "allow":

litellm/proxy/health_endpoints/_health_endpoints.py

@@ -1582,6 +1582,13 @@ async def _show_no_redis_warning() -> bool:
     return await count_live_proxy_workers(prisma_client) != 1
 
 
+def _show_env_credential_login_warning() -> bool:
+    from litellm.proxy.auth.login_utils import is_env_credential_login_enabled
+    from litellm.proxy.proxy_server import general_settings
+
+    return is_env_credential_login_enabled(general_settings)
+
+
 async def _get_health_readiness_details(
     response: Response | None = None,
 ) -> dict[str, Any]:
@@ -1623,6 +1630,7 @@ async def _get_health_readiness_details(
         log_level_name: Final = logging.getLevelName(verbose_logger.getEffectiveLevel())
         is_detailed_debug: Final = verbose_logger.isEnabledFor(logging.DEBUG)
         show_no_redis_warning: Final = await _show_no_redis_warning()
+        show_env_credential_login_warning: Final = _show_env_credential_login_warning()
 
         # check DB
         if prisma_client is not None:  # if db passed in, check if it's connected
@@ -1650,6 +1658,7 @@ async def _get_health_readiness_details(
                 "log_level": log_level_name,
                 "is_detailed_debug": is_detailed_debug,
                 "show_no_redis_warning": show_no_redis_warning,
+                "show_env_credential_login_warning": show_env_credential_login_warning,
             }
         else:
             return {
@@ -1662,6 +1671,7 @@ async def _get_health_readiness_details(
                 "log_level": log_level_name,
                 "is_detailed_debug": is_detailed_debug,
                 "show_no_redis_warning": show_no_redis_warning,
+                "show_env_credential_login_warning": show_env_credential_login_warning,
             }
     except Exception as e:
         raise HTTPException(status_code=503, detail=f"Service Unhealthy ({e})")

litellm/proxy/hooks/dynamic_rate_limiter_v3.py

@@ -32,6 +32,10 @@ from litellm.proxy.hooks.rate_limiter_utils import (
     resolve_llm_provider_for_rate_limit,
 )
 from litellm.proxy.utils import InternalUsageCache
+from litellm.router_utils.add_retry_fallback_headers import (
+    ensure_response_additional_headers,
+    response_has_hidden_params,
+)
 from litellm.types.router import ModelGroupInfo
 from litellm.types.utils import CallTypesLiteral
 
@@ -659,22 +663,12 @@ class _PROXY_DynamicRateLimitHandlerV3(CustomLogger):
                 data=data, user_api_key_dict=user_api_key_dict, response=response
             )
 
-            # Add additional priority-specific headers
-            if isinstance(response, ModelResponse):
+            if response_has_hidden_params(response):
                 priority: Final = self._get_priority_from_user_api_key_dict(user_api_key_dict=user_api_key_dict)
-
-                # Get existing additional headers
-                additional_headers: Final = getattr(response, "_hidden_params", {}).get("additional_headers", {}) or {}
-
-                # Add priority information
+                additional_headers: Final = ensure_response_additional_headers(response)
                 additional_headers["x-litellm-priority"] = priority or "default"
                 additional_headers["x-litellm-rate-limiter-version"] = "v3"
 
-                # Update response
-                if not hasattr(response, "_hidden_params"):
-                    response._hidden_params = {}
-                response._hidden_params["additional_headers"] = additional_headers
-
             return response
 
         except Exception as e:

litellm/proxy/hooks/parallel_request_limiter_v3.py

@@ -31,6 +31,7 @@ from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     get_str_from_messages,
 )
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.auth.auth_utils import (
     ESTIMATED_OUTPUT_TOKENS_FIELD,
@@ -52,6 +53,10 @@ from litellm.proxy.hooks.batch_enqueued_tokens import (
     canonical_provider_batch_id,
 )
 from litellm.proxy.hooks.rate_limiter_utils import resolve_llm_provider_for_rate_limit
+from litellm.router_utils.add_retry_fallback_headers import (
+    ensure_response_additional_headers,
+    response_has_hidden_params,
+)
 from litellm.types.caching import RedisPipelineIncrementOperation
 from litellm.types.llms.openai import BaseLiteLLMOpenAIResponseObject, ResponseAPIUsage
 from litellm.types.utils import (
@@ -3303,7 +3308,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
             min_configured_tpm_limit=min_configured_otpm_limit,
             call_type=call_type,
         )
-        raw_estimated_input_tokens: Final = self._estimate_precise_input_tokens(
+        raw_estimated_input_tokens: Final = await offload_token_count(self._estimate_precise_input_tokens)(
             data=data, model=requested_model, call_type=call_type
         )
         estimated_input_tokens: Final = max(raw_estimated_input_tokens, 1)
@@ -4677,34 +4682,17 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger):
         Post-call hook to update rate limit headers in the response.
         """
         try:
-            from pydantic import BaseModel
-
             stash: Final = get_request_stash()
             litellm_proxy_rate_limit_response: Final = stash.rate_limit_response if stash is not None else None
 
-            if litellm_proxy_rate_limit_response is not None:
-                # Update response headers
-                if hasattr(response, "_hidden_params"):
-                    _hidden_params = getattr(response, "_hidden_params")
-                else:
-                    _hidden_params = None
-
-                if _hidden_params is not None and (
-                    isinstance(_hidden_params, BaseModel) or isinstance(_hidden_params, dict)
-                ):
-                    if isinstance(_hidden_params, BaseModel):
-                        _hidden_params = _hidden_params.model_dump()
-
-                    _additional_headers: Final = self._merge_ratelimit_statuses_into_additional_headers(
-                        additional_headers=_hidden_params.get("additional_headers", {}) or {},
+            if litellm_proxy_rate_limit_response is not None and response_has_hidden_params(response):
+                additional_headers: Final = ensure_response_additional_headers(response)
+                additional_headers.update(
+                    self._merge_ratelimit_statuses_into_additional_headers(
+                        additional_headers={},
                         statuses=litellm_proxy_rate_limit_response["statuses"],
                     )
-
-                    setattr(
-                        response,
-                        "_hidden_params",
-                        {**_hidden_params, "additional_headers": _additional_headers},
-                    )
+                )
 
         except Exception as e:
             verbose_proxy_logger.exception("Error in rate limit post-call hook: %s", e)

litellm/proxy/litellm_pre_call_utils.py

@@ -235,6 +235,7 @@ _UNTRUSTED_ROOT_CONTROL_FIELDS: Final = (
     "applied_policies",
     "policy_sources",
     "guardrail_scan_ids",
+    "guardrail_scan_metadata",
     "routing_decision",
     GATEWAY_INJECTED_CACHE_METADATA_KEY,
     "pillar_response_headers",
@@ -291,6 +292,7 @@ _UNTRUSTED_METADATA_CONTROL_FIELDS: Final = (
     "applied_policies",
     "policy_sources",
     "guardrail_scan_ids",
+    "guardrail_scan_metadata",
     "routing_decision",
     GATEWAY_INJECTED_CACHE_METADATA_KEY,
     SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
@@ -771,6 +773,16 @@ def apply_missing_session_id_policy(
         return
     if policy == "omit":
         metadata[SESSION_ID_OMITTED_METADATA_KEY] = True
+        requester_metadata: Final = data.get("metadata")
+        requester_session_id: Final = (
+            requester_metadata.get("session_id") if isinstance(requester_metadata, dict) else None
+        )
+        if (
+            (body_session_id := data.get("litellm_session_id"))
+            and not metadata.get("session_id")
+            and not requester_session_id
+        ):
+            metadata["session_id"] = body_session_id
         return
     if data.get("litellm_session_id") or metadata.get("session_id"):
         return
@@ -1748,7 +1760,9 @@ class LiteLLMProxyRequestSetup:
         callback_vars_dict.pop("success_callback", None)
         callback_vars_dict.pop("failure_callback", None)
         callback_vars_dict = {
-            key: (litellm.utils.get_secret(value, default_value=value) or value if isinstance(value, str) else value)
+            key: (
+                litellm.utils.get_secret(value, default_value=value) or value if isinstance(value, str) else str(value)
+            )
             for key, value in callback_vars_dict.items()
         }
 

litellm/proxy/management_endpoints/common_daily_activity.py

@@ -14,6 +14,7 @@ from litellm.proxy._types import CommonProxyErrors
 from litellm.proxy.spend_tracking.key_metadata_recovery import (
     attach_user_emails,
     recover_double_hashed_key_metadata,
+    recover_key_metadata_from_spend_logs,
 )
 from litellm.proxy.spend_tracking.ptu_feature_flag import is_ptu_cost_attribution_enabled
 from litellm.proxy.utils import PrismaClient
@@ -433,9 +434,29 @@ def update_breakdown_metrics(
     return breakdown
 
 
+def _spend_logs_window(dates: AbstractSet[str | None]) -> tuple[datetime, datetime] | None:
+    parsed: Final = sorted(day for day in (_parse_spend_date(raw) for raw in dates) if day is not None)
+    if not parsed:
+        return None
+    return (parsed[0] - timedelta(days=1), parsed[-1] + timedelta(days=2))
+
+
+def _parse_spend_date(raw: str | None) -> datetime | None:
+    if not isinstance(raw, str):
+        return None
+    try:
+        return datetime.fromisoformat(raw)
+    except ValueError:
+        return None
+
+
+_EMPTY_KEY_METADATA: Final[Mapping[str, _KeyMetadataDict]] = MappingProxyType({})
+
+
 async def get_api_key_metadata(
     prisma_client: PrismaClient,
     api_keys: AbstractSet[str],
+    spend_logs_window: tuple[datetime, datetime] | None = None,
 ) -> Mapping[str, _KeyMetadataDict]:
     """Get api key metadata, falling back to deleted keys table for keys not found in active table.
 
@@ -481,11 +502,17 @@ async def get_api_key_metadata(
             )
 
     still_missing: Final = api_keys - frozenset(result)
-    combined: Final = (
-        result
-        if not still_missing
-        else MappingProxyType({**result, **(await recover_double_hashed_key_metadata(prisma_client, still_missing))})
+    from_reverse_hash: Final = (
+        await recover_double_hashed_key_metadata(prisma_client, still_missing) if still_missing else _EMPTY_KEY_METADATA
     )
+    after_token_recovery: Final = MappingProxyType({**result, **from_reverse_hash})
+    unresolved: Final = api_keys - frozenset(after_token_recovery)
+    from_spend_logs: Final = (
+        await recover_key_metadata_from_spend_logs(prisma_client, unresolved, spend_logs_window)
+        if unresolved and spend_logs_window is not None
+        else _EMPTY_KEY_METADATA
+    )
+    combined: Final = MappingProxyType({**after_token_recovery, **from_spend_logs})
     return await attach_user_emails(prisma_client, combined)
 
 
@@ -898,7 +925,9 @@ async def _aggregate_spend_records(
 
     api_key_metadata: dict[str, _KeyMetadataDict] = {}
     if api_keys:
-        api_key_metadata = await get_api_key_metadata(prisma_client, api_keys)
+        api_key_metadata = await get_api_key_metadata(
+            prisma_client, api_keys, _spend_logs_window(frozenset(record.date for record in records))
+        )
 
     return await asyncio.to_thread(
         _aggregate_spend_records_sync,
@@ -1094,7 +1123,9 @@ async def _aggregate_grouping_sets_records(
 
     api_key_metadata: dict[str, _KeyMetadataDict] = {}
     if api_keys:
-        api_key_metadata = await get_api_key_metadata(prisma_client, api_keys)
+        api_key_metadata = await get_api_key_metadata(
+            prisma_client, api_keys, _spend_logs_window(frozenset(r.date for r in records))
+        )
 
     return await asyncio.to_thread(
         _aggregate_grouping_sets_records_sync,
@@ -1357,7 +1388,9 @@ async def get_daily_activity_aggregated(
                 r.api_key for r in entity_records if r.api_key and r.api_key != PTU_SENTINEL_API_KEY
             )
             entity_key_metadata: Final = (
-                await get_api_key_metadata(prisma_client, entity_api_keys)
+                await get_api_key_metadata(
+                    prisma_client, entity_api_keys, _spend_logs_window(frozenset(r.date for r in entity_records))
+                )
                 if entity_api_keys
                 else {}  # mutable-ok: matches the helper's dict return
             )

litellm/proxy/management_endpoints/cost_tracking_settings.py

@@ -18,6 +18,7 @@ from fastapi import APIRouter, Depends, HTTPException
 from pydantic import BaseModel
 
 import litellm
+from litellm._internal_context import current_billing_time, pinned_billing_time
 from litellm._logging import verbose_proxy_logger
 from litellm.cost_calculator import completion_cost
 from litellm.proxy._types import (
@@ -27,7 +28,15 @@ from litellm.proxy._types import (
     UserAPIKeyAuth,
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
-from litellm.types.utils import CostPerToken, LlmProvidersSet, ModelInfo
+from litellm.types.utils import (
+    CostBreakdown,
+    CostPerToken,
+    LlmProvidersSet,
+    ModelInfo,
+    ModelResponse,
+    PromptTokensDetailsWrapper,
+    Usage,
+)
 
 router: Final = APIRouter()
 
@@ -46,13 +55,15 @@ def _configured_price(key: str, sources: tuple[Mapping[str, object], ...]) -> fl
 
 
 def _extract_custom_pricing(
-    litellm_params: Mapping[str, object], model_info: Mapping[str, object]
+    litellm_params: Mapping[str, object], model_info: Mapping[str, object], builtin: ModelInfo | None
 ) -> CostPerToken | None:
     """
     Pull per-token pricing configured on a deployment so on-prem / self-hosted
     models (absent from the public cost map) still estimate a real cost.
     Pricing may live on ``litellm_params`` or ``model_info``; ``litellm_params``
-    wins, matching the router's cost-map registration precedence.
+    wins, matching the router's cost-map registration precedence. Cache rates the
+    deployment leaves unset come from the backend model's built-in entry, then its
+    own input rate, again matching what the router registers for live billing.
     """
     sources: Final = (litellm_params, model_info)
     input_price: Final = _configured_price("input_cost_per_token", sources)
@@ -61,15 +72,21 @@ def _extract_custom_pricing(
     if input_price is None and output_price is None:
         return None
 
+    input_rate: Final = input_price or 0.0
+    cache_sources: Final = sources if builtin is None else (*sources, builtin)
+    cache_read_price: Final = _configured_price("cache_read_input_token_cost", cache_sources)
+    cache_creation_price: Final = _configured_price("cache_creation_input_token_cost", cache_sources)
     return CostPerToken(
-        input_cost_per_token=input_price or 0.0,
+        input_cost_per_token=input_rate,
         output_cost_per_token=output_price or 0.0,
+        cache_read_input_token_cost=input_rate if cache_read_price is None else cache_read_price,
+        cache_creation_input_token_cost=input_rate if cache_creation_price is None else cache_creation_price,
     )
 
 
-def _lookup_model_info(model: str) -> ModelInfo | None:
+def _lookup_model_info(model: str, custom_llm_provider: str | None = None) -> ModelInfo | None:
     try:
-        return litellm.get_model_info(model=model)
+        return litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider)
     except Exception:
         return None
 
@@ -98,17 +115,14 @@ def _resolve_model_for_cost_lookup(model: str) -> ResolvedCostModel:
                 model_info: Final = first_deployment.get("model_info", {})
                 custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
                 provider: Final = str(custom_llm_provider) if custom_llm_provider is not None else None
-                custom_cost_per_token: Final = _extract_custom_pricing(litellm_params, model_info)
-
-                # Check base_model first (needed for Azure custom deployment names)
+                # base_model wins (needed for Azure custom deployment names)
                 base_model: Final = model_info.get("base_model") or litellm_params.get("base_model")
-                if base_model:
-                    verbose_proxy_logger.debug("Resolved model '%s' to base_model '%s' from router", model, base_model)
-                    return ResolvedCostModel(str(base_model), provider, custom_cost_per_token)
-
-                resolved_model: Final = litellm_params.get("model")
+                resolved_model: Final = base_model or litellm_params.get("model")
                 if resolved_model:
                     verbose_proxy_logger.debug("Resolved model '%s' to '%s' from router", model, resolved_model)
+                    custom_cost_per_token: Final = _extract_custom_pricing(
+                        litellm_params, model_info, _lookup_model_info(str(resolved_model), provider)
+                    )
                     return ResolvedCostModel(str(resolved_model), provider, custom_cost_per_token)
         except Exception as e:
             verbose_proxy_logger.debug("Could not resolve model '%s' from router: %s", model, e)
@@ -117,19 +131,59 @@ def _resolve_model_for_cost_lookup(model: str) -> ResolvedCostModel:
     return ResolvedCostModel(model, None, None)
 
 
-def _calculate_period_costs(num_requests, cost_per_request, input_cost, output_cost, margin_cost):
-    """
-    Calculate costs for a given number of requests.
+@dataclass(frozen=True, slots=True)
+class CostLines:
+    """Cost of one request split the way the spend logs split it: the cache lines are
+    shares of input_cost and the reasoning line is a share of output_cost."""
 
-    Returns tuple of (total_cost, input_cost, output_cost, margin_cost) or all None if num_requests is None/0.
-    """
-    if not num_requests:
-        return None, None, None, None
-    return (
-        cost_per_request * num_requests,
-        input_cost * num_requests,
-        output_cost * num_requests,
-        margin_cost * num_requests,
+    total_cost: float
+    input_cost: float
+    output_cost: float
+    margin_cost: float
+    cache_read_cost: float
+    cache_creation_cost: float
+    reasoning_cost: float
+
+    def times(self, num_requests: int | None) -> "CostLines | None":
+        if not num_requests:
+            return None
+        return CostLines(
+            total_cost=self.total_cost * num_requests,
+            input_cost=self.input_cost * num_requests,
+            output_cost=self.output_cost * num_requests,
+            margin_cost=self.margin_cost * num_requests,
+            cache_read_cost=self.cache_read_cost * num_requests,
+            cache_creation_cost=self.cache_creation_cost * num_requests,
+            reasoning_cost=self.reasoning_cost * num_requests,
+        )
+
+
+def _cost_lines(cost_per_request: float, cost_breakdown: CostBreakdown | None) -> CostLines:
+    breakdown: Final = cost_breakdown if cost_breakdown is not None else CostBreakdown()
+    return CostLines(
+        total_cost=cost_per_request,
+        input_cost=breakdown.get("input_cost", 0.0),
+        output_cost=breakdown.get("output_cost", 0.0),
+        margin_cost=breakdown.get("margin_total_amount", 0.0),
+        cache_read_cost=breakdown.get("cache_read_cost", 0.0),
+        cache_creation_cost=breakdown.get("cache_creation_cost", 0.0),
+        reasoning_cost=breakdown.get("reasoning_cost", 0.0),
+    )
+
+
+def _usage_for_estimate(request: CostEstimateRequest) -> Usage:
+    cache_tokens: Final = request.cache_read_input_tokens + request.cache_creation_input_tokens
+    return Usage(
+        prompt_tokens=request.input_tokens,
+        completion_tokens=request.output_tokens,
+        total_tokens=request.input_tokens + request.output_tokens,
+        reasoning_tokens=request.reasoning_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            cached_tokens=request.cache_read_input_tokens,
+            cache_creation_tokens=request.cache_creation_input_tokens,
+        )
+        if cache_tokens
+        else None,
     )
 
 
@@ -530,11 +584,14 @@ async def estimate_cost(
     - model: Model name (e.g., "gpt-4", "claude-3-opus")
     - input_tokens: Expected input tokens per request
     - output_tokens: Expected output tokens per request
+    - cache_read_input_tokens: Cache-read tokens per request, counted within input_tokens (optional)
+    - cache_creation_input_tokens: Cache-write tokens per request, counted within input_tokens (optional)
+    - reasoning_tokens: Reasoning tokens per request, counted within output_tokens (optional)
     - num_requests_per_day: Number of requests per day (optional)
     - num_requests_per_month: Number of requests per month (optional)
 
     Returns cost breakdown including:
-    - Per-request costs (input, output, margin)
+    - Per-request costs (input, output, margin, plus the cache-read, cache-write and reasoning shares)
     - Daily costs (if num_requests_per_day provided)
     - Monthly costs (if num_requests_per_month provided)
 
@@ -543,14 +600,15 @@ async def estimate_cost(
     {
         "model": "gpt-4",
         "input_tokens": 1000,
+        "cache_read_input_tokens": 800,
         "output_tokens": 500,
+        "reasoning_tokens": 200,
         "num_requests_per_day": 100,
         "num_requests_per_month": 3000
     }
     ```
     """
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
-    from litellm.types.utils import ModelResponse, Usage
 
     # Resolve model name (handles router aliases like 'e-model-router' -> 'azure_ai/gpt-4')
     resolved: Final = _resolve_model_for_cost_lookup(request.model)
@@ -559,15 +617,8 @@ async def estimate_cost(
 
     verbose_proxy_logger.debug("Cost estimate: request.model='%s' resolved to '%s'", request.model, resolved_model)
 
-    # Create a mock response with usage for completion_cost
-    mock_response: Final = ModelResponse(
-        model=resolved_model,
-        usage=Usage(
-            prompt_tokens=request.input_tokens,
-            completion_tokens=request.output_tokens,
-            total_tokens=request.input_tokens + request.output_tokens,
-        ),
-    )
+    usage: Final = _usage_for_estimate(request)
+    mock_response: Final = ModelResponse(model=resolved_model, usage=usage)
 
     # Create a logging object to capture cost breakdown
     litellm_logging_obj: Final = LiteLLMLoggingObj(
@@ -580,92 +631,73 @@ async def estimate_cost(
         function_id="cost-estimate",
     )
 
-    # Use completion_cost which handles all the logic including margins/discounts
-    try:
-        cost_per_request: Final = completion_cost(
-            completion_response=mock_response,
-            model=resolved_model,
-            custom_llm_provider=resolved_provider,
-            custom_cost_per_token=resolved.custom_cost_per_token,
-            litellm_logging_obj=litellm_logging_obj,
-        )
-    except Exception as e:
-        raise HTTPException(
-            status_code=404,
-            detail={
-                "error": f"Could not calculate cost for model '{request.model}' (resolved to '{resolved_model}'): {e}"
-            },
-        )
+    # Pinning one moment keeps an off-peak window that opens mid-quote from pricing the totals on
+    # one side of it and the reported rates on the other.
+    with pinned_billing_time(current_billing_time()):
+        # Use completion_cost which handles all the logic including margins/discounts
+        try:
+            cost_per_request: Final = completion_cost(
+                completion_response=mock_response,
+                model=resolved_model,
+                custom_llm_provider=resolved_provider,
+                custom_cost_per_token=resolved.custom_cost_per_token,
+                litellm_logging_obj=litellm_logging_obj,
+            )
+        except Exception as e:  # noqa: BLE001  # completion_cost raises a bare Exception for an unpriceable model
+            raise HTTPException(
+                status_code=404,
+                detail={
+                    "error": f"Could not calculate cost for model '{request.model}' (resolved to '{resolved_model}'): {e}"
+                },
+            )
 
-    # Get cost breakdown from the logging object
-    cost_breakdown: Final = litellm_logging_obj.cost_breakdown
+    # The rates come back from the pricing call itself rather than a second lookup, so they are the
+    # ones the cost lines above billed at even when completion_cost infers a provider this endpoint
+    # never resolved (an unrouted "xai/grok-4" prices on xai's inclusive tier thresholds; a lookup
+    # here without that provider would report the sub-200k rate for a line billed above it).
+    rates: Final = litellm_logging_obj.billed_token_rates
+    per_request: Final = _cost_lines(cost_per_request, litellm_logging_obj.cost_breakdown)
+    daily: Final = per_request.times(request.num_requests_per_day)
+    monthly: Final = per_request.times(request.num_requests_per_month)
 
-    input_cost: Final = cost_breakdown.get("input_cost", 0.0) if cost_breakdown else 0.0
-    output_cost: Final = cost_breakdown.get("output_cost", 0.0) if cost_breakdown else 0.0
-    margin_cost: Final = cost_breakdown.get("margin_total_amount", 0.0) if cost_breakdown else 0.0
-
-    model_info: Final = _lookup_model_info(resolved_model)
-    mapped_input_price: Final = model_info.get("input_cost_per_token") if model_info is not None else None
-    mapped_output_price: Final = model_info.get("output_cost_per_token") if model_info is not None else None
+    model_info: Final = _lookup_model_info(resolved_model, resolved_provider)
     mapped_provider: Final = model_info.get("litellm_provider") if model_info is not None else None
-
-    input_cost_per_token: Final = (
-        resolved.custom_cost_per_token["input_cost_per_token"]
-        if resolved.custom_cost_per_token is not None
-        else mapped_input_price
-    )
-    output_cost_per_token: Final = (
-        resolved.custom_cost_per_token["output_cost_per_token"]
-        if resolved.custom_cost_per_token is not None
-        else mapped_output_price
-    )
     custom_llm_provider: Final = mapped_provider if mapped_provider is not None else resolved_provider
 
-    # Calculate daily and monthly costs
-    (
-        daily_cost,
-        daily_input_cost,
-        daily_output_cost,
-        daily_margin_cost,
-    ) = _calculate_period_costs(
-        num_requests=request.num_requests_per_day,
-        cost_per_request=cost_per_request,
-        input_cost=input_cost,
-        output_cost=output_cost,
-        margin_cost=margin_cost,
-    )
-    (
-        monthly_cost,
-        monthly_input_cost,
-        monthly_output_cost,
-        monthly_margin_cost,
-    ) = _calculate_period_costs(
-        num_requests=request.num_requests_per_month,
-        cost_per_request=cost_per_request,
-        input_cost=input_cost,
-        output_cost=output_cost,
-        margin_cost=margin_cost,
-    )
-
     return CostEstimateResponse(
         model=request.model,
         input_tokens=request.input_tokens,
         output_tokens=request.output_tokens,
+        cache_read_input_tokens=request.cache_read_input_tokens,
+        cache_creation_input_tokens=request.cache_creation_input_tokens,
+        reasoning_tokens=request.reasoning_tokens,
         num_requests_per_day=request.num_requests_per_day,
         num_requests_per_month=request.num_requests_per_month,
-        cost_per_request=cost_per_request,
-        input_cost_per_request=input_cost,
-        output_cost_per_request=output_cost,
-        margin_cost_per_request=margin_cost,
-        daily_cost=daily_cost,
-        daily_input_cost=daily_input_cost,
-        daily_output_cost=daily_output_cost,
-        daily_margin_cost=daily_margin_cost,
-        monthly_cost=monthly_cost,
-        monthly_input_cost=monthly_input_cost,
-        monthly_output_cost=monthly_output_cost,
-        monthly_margin_cost=monthly_margin_cost,
-        input_cost_per_token=input_cost_per_token,
-        output_cost_per_token=output_cost_per_token,
+        cost_per_request=per_request.total_cost,
+        input_cost_per_request=per_request.input_cost,
+        output_cost_per_request=per_request.output_cost,
+        margin_cost_per_request=per_request.margin_cost,
+        cache_read_cost_per_request=per_request.cache_read_cost,
+        cache_creation_cost_per_request=per_request.cache_creation_cost,
+        reasoning_cost_per_request=per_request.reasoning_cost,
+        daily_cost=daily.total_cost if daily is not None else None,
+        daily_input_cost=daily.input_cost if daily is not None else None,
+        daily_output_cost=daily.output_cost if daily is not None else None,
+        daily_margin_cost=daily.margin_cost if daily is not None else None,
+        daily_cache_read_cost=daily.cache_read_cost if daily is not None else None,
+        daily_cache_creation_cost=daily.cache_creation_cost if daily is not None else None,
+        daily_reasoning_cost=daily.reasoning_cost if daily is not None else None,
+        monthly_cost=monthly.total_cost if monthly is not None else None,
+        monthly_input_cost=monthly.input_cost if monthly is not None else None,
+        monthly_output_cost=monthly.output_cost if monthly is not None else None,
+        monthly_margin_cost=monthly.margin_cost if monthly is not None else None,
+        monthly_cache_read_cost=monthly.cache_read_cost if monthly is not None else None,
+        monthly_cache_creation_cost=monthly.cache_creation_cost if monthly is not None else None,
+        monthly_reasoning_cost=monthly.reasoning_cost if monthly is not None else None,
+        input_cost_per_token=rates.input_cost_per_token if rates is not None else None,
+        output_cost_per_token=rates.output_cost_per_token if rates is not None else None,
+        cache_read_input_token_cost=rates.cache_read_input_token_cost if rates is not None else None,
+        cache_creation_input_token_cost=rates.cache_creation_input_token_cost if rates is not None else None,
+        output_cost_per_reasoning_token=rates.output_cost_per_reasoning_token if rates is not None else None,
         provider=custom_llm_provider,
     )

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -4559,6 +4559,23 @@ async def delete_verification_tokens(
                 litellm_changed_by=litellm_changed_by,
             )
 
+            # Snapshot before the delete: the FK cascade drops the mapping rows, but their
+            # cached jwt_key_mapping entries still resolve to the now-dead token (LIT-5380).
+            jwt_mapping_cache_keys: Final[tuple[str, ...]] = tuple(
+                cache_key
+                for keys_for_token in await asyncio.gather(
+                    *(
+                        get_jwt_key_mapping_cache_keys_for_token(
+                            hashed_token=key.token,
+                            prisma_client=prisma_client,
+                        )
+                        for key in authorized_keys
+                        if key.token is not None
+                    )
+                )
+                for cache_key in keys_for_token
+            )
+
             if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN.value:
                 deleted_tokens = await prisma_client.delete_data(tokens=tokens)
                 if deleted_tokens is not None and len(deleted_tokens) != len(tokens):
@@ -4571,6 +4588,8 @@ async def delete_verification_tokens(
                 if len(deleted_tokens) != len(tokens):
                     failed_tokens = [token for token in tokens if token not in deleted_tokens]
 
+            await evict_and_broadcast(cache_keys=jwt_mapping_cache_keys, user_api_key_cache=user_api_key_cache)
+
         else:
             raise Exception("DB not connected. prisma_client is None")
     except Exception as e:

litellm/proxy/management_endpoints/mcp_management_endpoints.py

@@ -2673,6 +2673,8 @@ if MCP_AVAILABLE:
         """
         Updates the MCP Server in the db.
 
+        Partial update: a field left out of the payload keeps its stored value, and a field sent as null is cleared.
+
         Parameters:
         - payload: UpdateMCPServerRequest - Required. The updated mcp server data.
         ```
@@ -3098,6 +3100,8 @@ if MCP_AVAILABLE:
         user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth),
         litellm_changed_by: str | None = Header(None),
     ):
+        """Partial update: a field left out keeps its stored value, and a field sent as null is cleared, except
+        ``toolset_name`` and ``tools``, which a toolset always has; empty the tool selection with an explicit []."""
         prisma_client: Final = get_prisma_client_or_throw("Database not connected. Connect a database to your proxy")
         if LitellmUserRoles.PROXY_ADMIN != user_api_key_dict.user_role:
             raise HTTPException(

litellm/proxy/management_endpoints/team_endpoints.py

@@ -5601,7 +5601,7 @@ async def team_model_add(
     updated_team: Final = await _team_db(prisma_client).update(
         where={"team_id": data.team_id},
         data={"updated_at": datetime.now(timezone.utc)},
-        include={"object_permission": True},
+        include={"litellm_model_table": True, "object_permission": True},
     )
     if updated_team is None:
         raise HTTPException(
@@ -5688,7 +5688,7 @@ async def team_model_delete(
     updated_team: Final = await _team_db(prisma_client).update(
         where={"team_id": data.team_id},
         data={"models": updated_models},
-        include={"object_permission": True},
+        include={"litellm_model_table": True, "object_permission": True},
     )
     if updated_team is None:
         raise HTTPException(

litellm/proxy/management_endpoints/ui_sso.py

@@ -22,7 +22,6 @@ from html import escape
 from types import MappingProxyType
 from typing import (
     TYPE_CHECKING,
-    Annotated,
     Any,
     Final,
     Literal,
@@ -42,7 +41,7 @@ if TYPE_CHECKING:
 import jwt
 from fastapi import APIRouter, Depends, Header, HTTPException, Request, Response, status
 from fastapi.responses import RedirectResponse
-from pydantic import BaseModel, BeforeValidator, ConfigDict, TypeAdapter, ValidationError
+from pydantic import BaseModel, TypeAdapter, ValidationError
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -95,6 +94,7 @@ from litellm.proxy.auth.auth_utils import (
 )
 from litellm.proxy.auth.handle_jwt import JWTHandler
 from litellm.proxy.auth.ip_address_utils import IPAddressUtils
+from litellm.proxy.auth.team_grants import TeamModelAliasTable
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_utils.admin_ui_utils import (
     admin_ui_disabled,
@@ -209,31 +209,14 @@ def _team_detail_db(repo: TeamRepository) -> "TableActions[_TeamDetailRow]":
     return repo.table
 
 
-_MODEL_ALIASES_ADAPTER: Final = TypeAdapter(dict[str, str])
 _SSO_TOKEN_CLAIMS_ADAPTER: Final = TypeAdapter(Mapping[str, object])
 
 
-def _decode_model_aliases(value: object) -> object:
-    """``/team/new`` stores team model aliases as a JSON-encoded string in the Json column."""
-    if not isinstance(value, str):
-        return value
-    try:
-        return _MODEL_ALIASES_ADAPTER.validate_json(value)
-    except ValidationError:
-        return None
-
-
-class _TeamModelAliasTable(BaseModel):
-    model_config = ConfigDict(protected_namespaces=())
-
-    model_aliases: Annotated[Mapping[str, str] | None, BeforeValidator(_decode_model_aliases)] = None
-
-
 class _TeamRowGrants(BaseModel):
     team_id: str
     team_alias: str | None = None
     models: tuple[str, ...] = ()
-    litellm_model_table: _TeamModelAliasTable | None = None
+    litellm_model_table: TeamModelAliasTable | None = None
 
 
 class CliSsoTeamDetail(BaseModel):

litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py

@@ -15,6 +15,7 @@ import os
 import re
 from collections.abc import AsyncGenerator, Callable, Mapping, Sequence
 from dataclasses import dataclass
+from functools import partial
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Annotated, Final, Literal, Protocol, cast
 
@@ -1099,13 +1100,6 @@ async def bedrock_proxy_route(
     """
     create_request_copy(request)
 
-    try:
-        from botocore.auth import SigV4Auth
-        from botocore.awsrequest import AWSRequest
-        from botocore.credentials import Credentials
-    except ImportError:
-        raise ImportError("Missing boto3 to call bedrock. Run 'pip install boto3'.")
-
     aws_region_name: Final = get_secret_str(secret_name="AWS_REGION_NAME")
     if not _is_bedrock_agent_runtime_route(endpoint=endpoint):
         return await bedrock_llm_proxy_route(
@@ -1136,20 +1130,24 @@ async def bedrock_proxy_route(
     )
 
     # Add or update query parameters
+    from litellm.llms.bedrock.base_aws_llm import run_aws_signing, sign_aws_json_post
     from litellm.llms.bedrock.chat import BedrockConverseLLM
 
     bedrock_llm: Final = BedrockConverseLLM()
-    credentials: Final[Credentials] = bedrock_llm.get_credentials()
-    sigv4: Final = SigV4Auth(credentials, "bedrock", aws_region_name)
-    headers: Final = {"Content-Type": "application/json"}
     # Assuming the body contains JSON data, parse it
     try:
         data: Final = await _json_request_body(request)
     except Exception as e:
         raise HTTPException(status_code=400, detail={"error": e})
-    _request: Final = AWSRequest(method="POST", url=str(updated_url), data=json.dumps(data), headers=headers)
-    sigv4.add_auth(_request)
-    prepped: Final = _request.prepare()
+    prepped: Final = await run_aws_signing(
+        sign_aws_json_post,
+        get_credentials=bedrock_llm.get_credentials,
+        service_name="bedrock",
+        aws_region_name=aws_region_name,
+        url=str(updated_url),
+        body=json.dumps(data),
+        headers=MappingProxyType({"Content-Type": "application/json"}),
+    )
 
     ## check for streaming
     is_streaming_request = False
@@ -1207,13 +1205,6 @@ async def comprehend_medical_proxy_route(
 
     [Docs](https://docs.litellm.ai/docs/pass_through/comprehend_medical)
     """
-    try:
-        from botocore.auth import SigV4Auth
-        from botocore.awsrequest import AWSRequest
-        from botocore.credentials import Credentials
-    except ImportError:
-        raise ImportError("Missing boto3 to call comprehendmedical. Run 'pip install boto3'.")
-
     from .llm_provider_handlers.comprehend_medical_passthrough_logging_handler import (
         COMPREHEND_MEDICAL_SUPPORTED_OPERATIONS,
     )
@@ -1244,20 +1235,23 @@ async def comprehend_medical_proxy_route(
     if "stream" in data:
         raise HTTPException(status_code=400, detail="'stream' is not a Comprehend Medical request member")
 
-    from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
+    from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM, run_aws_signing, sign_aws_json_post
 
-    credentials: Final[Credentials] = BaseAWSLLM().get_credentials(aws_region_name=aws_region_name)
-    sigv4: Final = SigV4Auth(credentials, "comprehendmedical", aws_region_name)
-    headers: Final = MappingProxyType(
-        {
-            "Content-Type": "application/x-amz-json-1.1",
-            "X-Amz-Target": f"{COMPREHEND_MEDICAL_TARGET_PREFIX}.{operation}",
-        }
-    )
     target_url: Final = f"https://comprehendmedical.{aws_region_name}.{get_aws_dns_suffix(aws_region_name)}/"
-    _request: Final = AWSRequest(method="POST", url=target_url, data=json.dumps(data), headers=headers)
-    sigv4.add_auth(_request)
-    prepped: Final = _request.prepare()
+    prepped: Final = await run_aws_signing(
+        sign_aws_json_post,
+        get_credentials=partial(BaseAWSLLM().get_credentials, aws_region_name=aws_region_name),
+        service_name="comprehendmedical",
+        aws_region_name=aws_region_name,
+        url=target_url,
+        body=json.dumps(data),
+        headers=MappingProxyType(
+            {
+                "Content-Type": "application/x-amz-json-1.1",
+                "X-Amz-Target": f"{COMPREHEND_MEDICAL_TARGET_PREFIX}.{operation}",
+            }
+        ),
+    )
 
     endpoint_func: Final = create_pass_through_route(
         endpoint=operation,

litellm/proxy/policy_engine/pipeline_executor.py

@@ -70,6 +70,10 @@ def _text_snapshot(texts: Sequence[str] | None) -> tuple[str, ...] | None:
     return None if texts is None else tuple(texts)
 
 
+def _scanned_texts(texts: Sequence[str] | None) -> tuple[str, ...]:
+    return tuple(texts or ())
+
+
 def _tool_call_shapes(tool_calls: Sequence[object] | None) -> tuple[tuple[object, object], ...] | None:
     return None if tool_calls is None else tuple(_tool_call_shape(tool_call) for tool_call in tool_calls)
 
@@ -78,6 +82,10 @@ def _rewrote(sent: tuple[object, ...] | None, returned: tuple[object, ...] | Non
     return sent is not None and returned is not None and returned != sent
 
 
+def _changed_count(sent: tuple[object, ...] | None, returned: tuple[object, ...] | None) -> bool:
+    return sent is not None and returned is not None and len(returned) != len(sent)
+
+
 _GuardrailMethodT = TypeVar("_GuardrailMethodT", bound=Callable[..., object])
 
 
@@ -89,10 +97,11 @@ def _logged_by_inner_guardrail(method: _GuardrailMethodT) -> _GuardrailMethodT:
 class _StreamRewriteObserver(CustomGuardrail):
     """Stand-in handed to the endpoint translation in place of a streaming pipeline step's
     guardrail. It records whether the guardrail returned different output than it was given,
-    which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text
-    rewrites are deliverable on translations that write them back across the buffered chunks
-    (``delivers_ended_stream_text_rewrites``); tool-call rewrites and text rewrites on any
-    other translation are discarded by the executor, which releases the original chunks.
+    which for guardrails like Bedrock's ANONYMIZED action is only known at runtime. Text and
+    tool-call rewrites are deliverable on translations that write them back across the
+    buffered chunks (``delivers_ended_stream_rewrites``); rewrites on any other translation,
+    and a rewrite that drops or adds a tool call on any translation, are discarded by the
+    executor, which releases the original chunks.
     The inner guardrail's ``apply_guardrail`` already records the guardrail information
     and span, so the observer's stays out of ``log_guardrail_information``."""
 
@@ -101,6 +110,7 @@ class _StreamRewriteObserver(CustomGuardrail):
         self.inner: Final = inner
         self.rewrote_texts = False
         self.rewrote_tool_calls = False
+        self.changed_tool_call_count = False
 
     def structured_messages_cover_full_request(self) -> bool:
         return self.inner.structured_messages_cover_full_request()
@@ -118,13 +128,103 @@ class _StreamRewriteObserver(CustomGuardrail):
         outputs: Final = await self.inner.apply_guardrail(
             inputs=inputs, request_data=request_data, input_type=input_type, logging_obj=logging_obj
         )
+        returned_tool_shapes: Final = _tool_call_shapes(outputs.get("tool_calls"))
         self.rewrote_texts = self.rewrote_texts or _rewrote(sent_texts, _text_snapshot(outputs.get("texts")))
-        self.rewrote_tool_calls = self.rewrote_tool_calls or _rewrote(
-            sent_tool_shapes, _tool_call_shapes(outputs.get("tool_calls"))
+        self.rewrote_tool_calls = self.rewrote_tool_calls or _rewrote(sent_tool_shapes, returned_tool_shapes)
+        self.changed_tool_call_count = self.changed_tool_call_count or _changed_count(
+            sent_tool_shapes, returned_tool_shapes
         )
         return outputs
 
 
+class _ScannedTextRecorder(CustomGuardrail):
+    def __init__(self, guardrail_name: str) -> None:
+        super().__init__(guardrail_name=guardrail_name)
+        self.inputs: GenericGuardrailAPIInputs | None = None
+
+    @_logged_by_inner_guardrail
+    async def apply_guardrail(
+        self,
+        inputs: GenericGuardrailAPIInputs,
+        request_data: dict,  # mutable-ok: matches CustomGuardrail.apply_guardrail
+        input_type: Literal["request", "response"],
+        logging_obj: "LiteLLMLoggingObj | None" = None,
+    ) -> GenericGuardrailAPIInputs:
+        self.inputs = inputs
+        return inputs
+
+
+class _LegacyHookStreamAdapter(CustomGuardrail):
+    """Runs a guardrail that only implements the legacy post-call hook (no unified
+    ``apply_guardrail``, or ``use_native_lifecycle_hooks``) as a streaming pipeline step. The
+    endpoint translation hands it the texts it scanned plus the assembled response under
+    ``request_data["response"]``; the hook gets that response in the shape its route gives
+    non-streaming hooks, an exception it raises ends the stream through the executor's
+    fail/error classification, and the response it hands back, or the one it changed in place
+    and returned ``None`` for, is re-scanned by the same translation so its texts reach the
+    client through the translation's ended-stream write-back. A
+    replacement whose scanned texts do not line up with the originals, or whose tool calls
+    differ from them, is undeliverable, so the executor releases the original chunks. A stream
+    that carried no text to scan, such as a tool-only Anthropic message, stays deliverable as
+    long as the hook left the tool calls alone."""
+
+    def __init__(
+        self,
+        inner: CustomGuardrail,
+        endpoint_translation: "BaseTranslation",
+        user_api_key_dict: "UserAPIKeyAuth",
+    ) -> None:
+        super().__init__(guardrail_name=inner.guardrail_name)
+        self.inner: Final = inner
+        self.endpoint_translation: Final = endpoint_translation
+        self.user_api_key_dict: Final = user_api_key_dict
+
+    def structured_messages_cover_full_request(self) -> bool:
+        return self.inner.structured_messages_cover_full_request()
+
+    @_logged_by_inner_guardrail
+    async def apply_guardrail(
+        self,
+        inputs: GenericGuardrailAPIInputs,
+        request_data: dict,  # mutable-ok: matches CustomGuardrail.apply_guardrail
+        input_type: Literal["request", "response"],
+        logging_obj: "LiteLLMLoggingObj | None" = None,
+    ) -> GenericGuardrailAPIInputs:
+        hooked: Final = self.endpoint_translation.post_call_hook_response(request_data.get("response"))
+        replacement: Final = await self.inner.async_post_call_success_hook(
+            data=request_data,
+            user_api_key_dict=self.user_api_key_dict,
+            response=hooked,
+        )
+        rewrite: Final = hooked if replacement is None else replacement
+        if rewrite is None:
+            return inputs
+        rescanned: Final = await self._rescan(rewrite, logging_obj)
+        if rescanned is None:
+            raise UndeliverableStreamRewrite(self.guardrail_name or "unknown")
+        rewritten: Final = rescanned.get("texts")
+        if len(_scanned_texts(rewritten)) != len(_scanned_texts(inputs.get("texts"))):
+            raise UndeliverableStreamRewrite(self.guardrail_name or "unknown")
+        if _tool_call_shapes(rescanned.get("tool_calls")) != _tool_call_shapes(inputs.get("tool_calls")):
+            raise UndeliverableStreamRewrite(self.guardrail_name or "unknown")
+        if not rewritten:
+            return inputs
+        rewritten_inputs: Final[GenericGuardrailAPIInputs] = {**inputs, "texts": rewritten}
+        return rewritten_inputs
+
+    async def _rescan(
+        self, response: object, logging_obj: "LiteLLMLoggingObj | None"
+    ) -> GenericGuardrailAPIInputs | None:
+        recorder: Final = _ScannedTextRecorder(self.guardrail_name or "unknown")
+        await self.endpoint_translation.process_output_response(
+            response=response,
+            guardrail_to_apply=recorder,
+            litellm_logging_obj=logging_obj,
+            user_api_key_dict=self.user_api_key_dict,
+        )
+        return recorder.inputs
+
+
 def _prepare_hook_input(
     step: PipelineStep,
     callback: CustomGuardrail,
@@ -292,18 +392,29 @@ class PipelineExecutor:
         endpoint_translation: "BaseTranslation",
         streaming_chunks: list[object],  # mutable-ok: shared buffered-stream chunks the translation rewrites in place
         hook_input: dict[str, object],  # mutable-ok: same request-payload shape as data
-        user_api_key_dict: "UserAPIKeyAuth | None",
+        user_api_key_dict: "UserAPIKeyAuth",
         litellm_logging_obj: "LiteLLMLoggingObj | None",
     ) -> None:
         """Run one streaming post_call step through the endpoint translation, delivering
-        text rewrites on translations that support ended-stream write-back. A rewrite that
-        cannot reach the client yet (a tool-call rewrite, a text rewrite on a translation
-        without write-back, or one the translation refused with
-        ``UndeliverableStreamRewrite``) is discarded: the buffered chunks go back to the
-        originals and the step passes, so the client gets the stream the merge base sent."""
-        observer: Final = _StreamRewriteObserver(callback)
-        deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_text_rewrites
+        text and tool-call rewrites on translations that support ended-stream write-back. A
+        guardrail without the unified interface runs its legacy post-call hook against the
+        assembled response through ``_LegacyHookStreamAdapter``. A rewrite that cannot reach the
+        client yet (one on a translation without write-back, one that drops or adds a tool call,
+        or one the translation or adapter refused with ``UndeliverableStreamRewrite``) is
+        discarded: the buffered chunks go back to the originals and the step passes, so the
+        client gets the stream the merge base sent, and the guardrail stays out of the
+        applied-guardrails header since its output never reached the client. The response an
+        earlier step's translation stored under ``request_data["response"]`` is dropped first,
+        so this step's hook sees the stream as the steps before it left it."""
+        scanner: Final = (
+            callback
+            if PipelineExecutor.supports_unified_execution(callback)
+            else _LegacyHookStreamAdapter(callback, endpoint_translation, user_api_key_dict)
+        )
+        observer: Final = _StreamRewriteObserver(scanner)
+        deliver_rewrites: Final = type(endpoint_translation).delivers_ended_stream_rewrites
         originals: Final = copy.deepcopy(streaming_chunks)
+        hook_input.pop("response", None)  # rebind-ok: an earlier step's stored response goes so this step's is stored
         try:
             if deliver_rewrites:
                 await endpoint_translation.process_output_streaming_response(
@@ -324,9 +435,12 @@ class PipelineExecutor:
                 )
         except UndeliverableStreamRewrite:
             _release_original_chunks(step.guardrail, streaming_chunks, originals)
-        else:
-            if observer.rewrote_tool_calls or (observer.rewrote_texts and not deliver_rewrites):
-                _release_original_chunks(step.guardrail, streaming_chunks, originals)
+            return
+        if observer.changed_tool_call_count or (
+            not deliver_rewrites and (observer.rewrote_texts or observer.rewrote_tool_calls)
+        ):
+            _release_original_chunks(step.guardrail, streaming_chunks, originals)
+            return
         if not callback.records_own_guardrail_information:
             add_guardrail_to_applied_guardrails_header(request_data=hook_input, guardrail_name=step.guardrail)
 
@@ -386,11 +500,11 @@ class PipelineExecutor:
                     if isinstance(response, dict):
                         callback.mark_pre_call_hook_ran(response)
             elif mode == "post_call" and streaming_chunks is not None:
-                if not use_unified or endpoint_translation is None:
+                if endpoint_translation is None:
                     return (
                         "error",
                         None,
-                        f"Guardrail '{step.guardrail}' does not support streaming pipeline execution",
+                        f"Guardrail '{step.guardrail}' cannot run on a stream without an endpoint translation",
                         None,
                     )
                 await PipelineExecutor._run_streaming_step(
@@ -446,10 +560,22 @@ class PipelineExecutor:
 
     @staticmethod
     def supports_unified_execution(callback: CustomGuardrail) -> bool:
-        """Whether this guardrail runs through the unified apply_guardrail path,
-        the interface streaming pipeline execution requires."""
+        """Whether this guardrail runs through the unified apply_guardrail path."""
         return "apply_guardrail" in type(callback).__dict__ and not callback.use_native_lifecycle_hooks
 
+    @staticmethod
+    def supports_streaming_execution(callback: CustomGuardrail) -> bool:
+        """Whether a streaming pipeline step can run this guardrail against the buffered
+        stream: through the unified path, or through its post-call hook on the assembled
+        response when that hook is its only streaming path. A guardrail with its own
+        streaming iterator hook, or with neither hook, keeps running on its own."""
+        callback_type: Final = type(callback)
+        return PipelineExecutor.supports_unified_execution(callback) or (
+            callback_type.async_post_call_success_hook is not CustomLogger.async_post_call_success_hook
+            and callback_type.async_post_call_streaming_iterator_hook
+            is CustomLogger.async_post_call_streaming_iterator_hook
+        )
+
     @staticmethod
     def find_guardrail_callback(guardrail_name: str) -> CustomGuardrail | None:
         """Look up an initialized guardrail callback by name from litellm.callbacks."""

litellm/proxy/policy_engine/response_retrieval.py

@@ -0,0 +1,152 @@
+from collections.abc import Mapping
+from dataclasses import dataclass
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final, Literal, TypeAlias
+
+from pydantic import TypeAdapter
+
+from litellm._logging import verbose_proxy_logger
+from litellm.litellm_core_utils.core_helpers import get_or_create_metadata_bucket
+from litellm.proxy.common_utils.callback_utils import (
+    add_guardrail_to_applied_guardrails_header,
+    add_policy_sources_to_metadata,
+    add_policy_to_applied_policies_header,
+)
+from litellm.proxy.common_utils.http_parsing_utils import get_tags_from_request_body
+from litellm.proxy.policy_engine.attachment_registry import get_attachment_registry
+from litellm.proxy.policy_engine.policy_matcher import PolicyMatcher
+from litellm.proxy.policy_engine.policy_registry import get_policy_registry
+from litellm.proxy.policy_engine.policy_resolver import PolicyResolver
+from litellm.responses.utils import ResponsesAPIRequestUtils
+from litellm.router_utils.common_utils import resolve_model_group_alias
+from litellm.types.proxy.policy_engine import PolicyMatchContext
+from litellm.types.proxy.policy_engine.pipeline_types import GuardrailPipeline
+
+if TYPE_CHECKING:
+    from litellm.proxy._types import UserAPIKeyAuth
+    from litellm.router import Router
+
+PolicyPipelines: TypeAlias = tuple[tuple[str, GuardrailPipeline], ...]
+
+_POLICY_PIPELINES_ADAPTER: Final = TypeAdapter(PolicyPipelines)
+
+
+@dataclass(frozen=True, slots=True)
+class UngovernedRetrieval:
+    reason: Literal["no router", "response id names no deployment", "deployment no longer in the router"]
+
+
+def _model_group_for_response_id(response_id: object, llm_router: "Router | None") -> str | UngovernedRetrieval:
+    if llm_router is None:
+        return UngovernedRetrieval("no router")
+    model_id: Final = (
+        ResponsesAPIRequestUtils.get_model_id_from_response_id(response_id) if isinstance(response_id, str) else None
+    )
+    if model_id is None:
+        return UngovernedRetrieval("response id names no deployment")
+    deployment: Final = llm_router.get_deployment(model_id)
+    if deployment is None:
+        return UngovernedRetrieval("deployment no longer in the router")
+    hidden_by: Final = _submit_model_hidden_by(deployment.model_name, llm_router.model_group_alias)
+    if hidden_by is not None:
+        verbose_proxy_logger.warning(
+            "Policy engine: background response %s re-matches policies on retrieval as model group %s (%s), "
+            "so a policy attached to the model name it was submitted as does not run on it",
+            response_id,
+            deployment.model_name,
+            hidden_by,
+        )
+    return deployment.model_name
+
+
+def _submit_model_hidden_by(model_group: str, model_group_alias: Mapping[str, object]) -> str | None:
+    if "*" in model_group:
+        return "a wildcard deployment"
+    aliases: Final = tuple(
+        alias for alias in model_group_alias if resolve_model_group_alias(model_group_alias, alias) == model_group
+    )
+    if not aliases:
+        return None
+    return f"the target of model_group_alias {', '.join(aliases)}"
+
+
+def _retrieval_context(
+    data: Mapping[str, object], user_api_key_dict: "UserAPIKeyAuth", model_group: str
+) -> PolicyMatchContext:
+    team_alias: Final = user_api_key_dict.team_alias
+    key_alias: Final = user_api_key_dict.key_alias
+    return PolicyMatchContext(
+        team_alias=team_alias if isinstance(team_alias, str) else None,
+        key_alias=key_alias if isinstance(key_alias, str) else None,
+        model=model_group,
+        tags=get_tags_from_request_body(data) or None,
+    )
+
+
+def _post_call_pipelines_for_context(context: PolicyMatchContext) -> tuple[PolicyPipelines, Mapping[str, str]]:
+    matches: Final = get_attachment_registry().get_attached_policies_with_reasons(context)
+    if not matches:
+        return (), MappingProxyType({})
+    applied_policy_names: Final = PolicyMatcher.get_policies_with_matching_conditions(
+        policy_names=[match["policy_name"] for match in matches],  # mutable-ok: the matcher takes a list
+        context=context,
+    )
+    post_call_pipelines: Final = tuple(
+        (policy_name, pipeline)
+        for policy_name, pipeline in PolicyResolver.resolve_pipelines_for_context(
+            context=context, policy_names=applied_policy_names
+        )
+        if pipeline.mode == "post_call"
+    )
+    return post_call_pipelines, MappingProxyType({match["policy_name"]: match["matched_via"] for match in matches})
+
+
+def attach_post_call_pipelines_to_retrieval(
+    data: dict[str, object],  # mutable-ok: request-state dict the policy engine hooks all write in place
+    user_api_key_dict: "UserAPIKeyAuth",
+    llm_router: "Router | None",
+) -> None:
+    if not get_policy_registry().is_initialized():
+        return
+    model_group: Final = _model_group_for_response_id(data.get("response_id"), llm_router)
+    if isinstance(model_group, UngovernedRetrieval):
+        verbose_proxy_logger.warning(
+            "Policy engine: background response %s is retrieved without its post_call policy pipelines (%s)",
+            data.get("response_id"),
+            model_group.reason,
+        )
+        return
+    context: Final = _retrieval_context(data, user_api_key_dict, model_group)
+    post_call_pipelines, policy_sources = _post_call_pipelines_for_context(context)
+    _, bucket = get_or_create_metadata_bucket(data)
+    already_attached: Final = _POLICY_PIPELINES_ADAPTER.validate_python(bucket.get("_guardrail_pipelines") or ())
+    attached_policy_names: Final = frozenset(policy_name for policy_name, _pipeline in already_attached)
+    added: Final = tuple(
+        (policy_name, pipeline)
+        for policy_name, pipeline in post_call_pipelines
+        if policy_name not in attached_policy_names
+    )
+    if not added:
+        return
+    pipelines: Final = (*already_attached, *added)
+    bucket["_guardrail_pipelines"] = pipelines
+    bucket["_pipeline_managed_guardrails"] = frozenset(
+        step.guardrail for _policy_name, pipeline in pipelines for step in pipeline.steps
+    )
+    for policy_name, _pipeline in added:
+        add_policy_to_applied_policies_header(request_data=data, policy_name=policy_name)
+    for _policy_name, pipeline in added:
+        for step in pipeline.steps:
+            add_guardrail_to_applied_guardrails_header(request_data=data, guardrail_name=step.guardrail)
+    add_policy_sources_to_metadata(
+        request_data=data,
+        policy_sources={  # mutable-ok: add_policy_sources_to_metadata takes a dict
+            policy_name: policy_sources[policy_name] for policy_name, _pipeline in added
+        },
+    )
+    verbose_proxy_logger.debug(
+        "Policy engine: attached post_call pipelines to the retrieval of background response %s (model group %s): %s",
+        data.get("response_id"),
+        model_group,
+        ", ".join(policy_name for policy_name, _pipeline in added),
+    )

litellm/proxy/proxy_server.py

@@ -17,6 +17,7 @@ import time
 import traceback
 import warnings
 from collections.abc import AsyncGenerator, AsyncIterator, Callable, Collection, Mapping, MutableMapping, Sequence
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
 from types import MappingProxyType, UnionType
 from typing import (
@@ -62,11 +63,13 @@ from litellm.constants import (
     LITELLM_UI_SESSION_DURATION,
     RUNTIME_UPDATABLE_ROUTER_SETTINGS,
 )
+from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.litellm_logging import (
     _init_custom_logger_compatible_class,
 )
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.litellm_core_utils.safe_json_loads import safe_json_loads
+from litellm.litellm_core_utils.token_counter import offload_token_count
 from litellm.proxy._types import (
     UI_TEAM_ID,
     CallbackDelete,
@@ -131,6 +134,7 @@ from litellm.router_utils.auto_router_tuning_baseline import (
     snapshot_tuning_baselines,
     tuning_limit_violation,
 )
+from litellm.types.caching import RedisPipelineIncrementOperation
 from litellm.types.utils import (
     ModelResponse,
     ModelResponseStream,
@@ -138,11 +142,7 @@ from litellm.types.utils import (
     TextCompletionResponse,
     TokenCountResponse,
 )
-from litellm.utils import (
-    _invalidate_model_cost_lowercase_map,
-    load_credentials_from_list,
-    reapply_runtime_model_cost_registrations,
-)
+from litellm.utils import load_credentials_from_list
 
 if TYPE_CHECKING:
     from aiohttp import ClientSession
@@ -274,7 +274,6 @@ from litellm.integrations.SlackAlerting.slack_alerting import SlackAlerting
 from litellm.litellm_core_utils.agentic_loop_settings import (
     validated_max_agentic_loops,
 )
-from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.audio_utils.utils import resolve_speech_media_type
 from litellm.litellm_core_utils.core_helpers import (
     _get_parent_otel_span_from_kwargs,
@@ -426,6 +425,7 @@ from litellm.proxy.db.exception_handler import (
 )
 from litellm.proxy.db.gateway_request_tracking import (
     GatewayRequestAccumulator,
+    GatewayRequestRedisBuffer,
     flush_gateway_requests,
 )
 from litellm.proxy.db.proxy_worker_heartbeat import (
@@ -2357,6 +2357,17 @@ open_telemetry_logger: OpenTelemetry | None = None
 gateway_request_accumulator: Final = GatewayRequestAccumulator()
 ### INITIALIZE GLOBAL LOGGING OBJECT ###
 proxy_logging_obj: ProxyLogging = ProxyLogging(user_api_key_cache=user_api_key_cache, premium_user=premium_user)
+
+
+def _gateway_request_redis_buffer() -> GatewayRequestRedisBuffer | None:
+    """Shares the spend writer's transaction-buffer Redis and pod lock when use_redis_transaction_buffer is on."""
+    writer: Final = proxy_logging_obj.db_spend_update_writer
+    redis_cache: Final = writer.redis_update_buffer.redis_cache
+    if redis_cache is None or not writer.redis_update_buffer._should_commit_spend_updates_to_redis():
+        return None
+    return GatewayRequestRedisBuffer(redis_cache=redis_cache, pod_lock_manager=writer.pod_lock_manager)
+
+
 ### REDIS QUEUE ###
 async_result: Final = None
 celery_app_conn: Final = None
@@ -2707,6 +2718,12 @@ async def _read_spend_counter_estimate(counter_key: str, fallback_spend: float)
     return fallback_spend, False
 
 
+@dataclass(frozen=True, slots=True)
+class _PendingSpendIncrement:
+    counter_key: str
+    increment: float
+
+
 async def increment_spend_counters(
     token: str | None,
     team_id: str | None,
@@ -2741,7 +2758,7 @@ async def increment_spend_counters(
 
     cost: Final[float] = response_cost
 
-    async def _key_scope(key_token: str) -> None:
+    async def _key_scope(key_token: str) -> tuple[_PendingSpendIncrement | BaseException, ...]:
         # key_token arrives pre-hashed from metadata["user_api_key"] (auth flow
         # hashes raw "sk-..." keys before they reach the callback). The
         # startswith("sk-") check is a safety net matching update_cache —
@@ -2752,30 +2769,29 @@ async def increment_spend_counters(
             hash_token(token=key_token) if isinstance(key_token, str) and key_token.startswith("sk-") else key_token
         )
         key_counter_key: Final = f"spend:key:{hashed_token}"
-        if key_counter_key not in reserved_counter_keys:
-            await _init_and_increment_spend_counter(
-                counter_key=key_counter_key,
-                source_cache_key=hashed_token,
-                increment=cost,
+        key_pending: Final[tuple[_PendingSpendIncrement, ...]] = (
+            ()
+            if key_counter_key in reserved_counter_keys
+            else (
+                await _prepare_spend_counter_increment(
+                    counter_key=key_counter_key,
+                    source_cache_key=hashed_token,
+                    increment=cost,
+                ),
             )
-
-        key_obj: Final[object] = await user_api_key_cache.async_get_cache(key=hashed_token)
-        if key_obj is None:
-            return
-        key_budget_limits = getattr(key_obj, "budget_limits", None) or (
-            key_obj.get("budget_limits") if isinstance(key_obj, dict) else None
         )
-        if isinstance(key_budget_limits, str):
-            key_budget_limits = json.loads(key_budget_limits)
-        if not isinstance(key_budget_limits, list):
-            return
-        for window in key_budget_limits:
-            duration = window["budget_duration"] if isinstance(window, dict) else window.budget_duration
-            key_window_reset_at = window.get("reset_at") if isinstance(window, dict) else window.reset_at
-            key_window_counter = f"spend:key:{hashed_token}:window:{duration}"
+
+        async def _key_window_increment(window: object) -> _PendingSpendIncrement | None:
+            duration = (
+                window["budget_duration"] if isinstance(window, dict) else getattr(window, "budget_duration", None)
+            )
+            key_window_reset_at = (
+                window.get("reset_at") if isinstance(window, dict) else getattr(window, "reset_at", None)
+            )
+            key_window_counter: Final = f"spend:key:{hashed_token}:window:{duration}"
             key_window_start = get_budget_window_start(window)
-            if key_window_counter not in reserved_counter_keys:
-                await _init_and_increment_window_spend_counter(
+            pending_window: Final = (
+                await _prepare_window_spend_counter_increment(
                     counter_key=key_window_counter,
                     entity_type="Key",
                     entity_id=hashed_token,
@@ -2783,6 +2799,9 @@ async def increment_spend_counters(
                     window_start=key_window_start,
                     increment=cost,
                 )
+                if key_window_counter not in reserved_counter_keys
+                else None
+            )
             await _enqueue_window_spend_row_update(
                 entity_type=Litellm_EntityType.KEY,
                 entity_id=hashed_token,
@@ -2792,33 +2811,48 @@ async def increment_spend_counters(
                 increment=cost,
                 request_started_at=request_started_at,
             )
+            return pending_window
 
-    async def _team_scope(scope_team_id: str) -> None:
-        team_counter_key: Final = f"spend:team:{scope_team_id}"
-        if team_counter_key not in reserved_counter_keys:
-            await _init_and_increment_spend_counter(
-                counter_key=team_counter_key,
-                source_cache_key=f"team_id:{scope_team_id}",
-                increment=cost,
-            )
-
-        team_obj: Final[object] = await user_api_key_cache.async_get_cache(key=f"team_id:{scope_team_id}")
-        if team_obj is None:
-            return
-        team_budget_limits = getattr(team_obj, "budget_limits", None) or (
-            team_obj.get("budget_limits") if isinstance(team_obj, dict) else None
+        key_obj: Final[object] = await user_api_key_cache.async_get_cache(key=hashed_token)
+        if key_obj is None:
+            return key_pending
+        key_budget_limits = getattr(key_obj, "budget_limits", None) or (
+            key_obj.get("budget_limits") if isinstance(key_obj, dict) else None
         )
-        if isinstance(team_budget_limits, str):
-            team_budget_limits = json.loads(team_budget_limits)
-        if not isinstance(team_budget_limits, list):
-            return
-        for window in team_budget_limits:
-            duration = window["budget_duration"] if isinstance(window, dict) else window.budget_duration
-            team_window_reset_at = window.get("reset_at") if isinstance(window, dict) else window.reset_at
-            team_window_counter = f"spend:team:{scope_team_id}:window:{duration}"
+        if isinstance(key_budget_limits, str):
+            key_budget_limits = json.loads(key_budget_limits)
+        if not isinstance(key_budget_limits, list):
+            return key_pending
+        window_pending: Final = await asyncio.gather(
+            *(_key_window_increment(window) for window in key_budget_limits), return_exceptions=True
+        )
+        return key_pending + tuple(item for item in window_pending if item is not None)
+
+    async def _team_scope(scope_team_id: str) -> tuple[_PendingSpendIncrement | BaseException, ...]:
+        team_counter_key: Final = f"spend:team:{scope_team_id}"
+        team_pending: Final[tuple[_PendingSpendIncrement, ...]] = (
+            ()
+            if team_counter_key in reserved_counter_keys
+            else (
+                await _prepare_spend_counter_increment(
+                    counter_key=team_counter_key,
+                    source_cache_key=f"team_id:{scope_team_id}",
+                    increment=cost,
+                ),
+            )
+        )
+
+        async def _team_window_increment(window: object) -> _PendingSpendIncrement | None:
+            duration = (
+                window["budget_duration"] if isinstance(window, dict) else getattr(window, "budget_duration", None)
+            )
+            team_window_reset_at = (
+                window.get("reset_at") if isinstance(window, dict) else getattr(window, "reset_at", None)
+            )
+            team_window_counter: Final = f"spend:team:{scope_team_id}:window:{duration}"
             team_window_start = get_budget_window_start(window)
-            if team_window_counter not in reserved_counter_keys:
-                await _init_and_increment_window_spend_counter(
+            pending_window: Final = (
+                await _prepare_window_spend_counter_increment(
                     counter_key=team_window_counter,
                     entity_type="Team",
                     entity_id=scope_team_id,
@@ -2826,6 +2860,9 @@ async def increment_spend_counters(
                     window_start=team_window_start,
                     increment=cost,
                 )
+                if team_window_counter not in reserved_counter_keys
+                else None
+            )
             await _enqueue_window_spend_row_update(
                 entity_type=Litellm_EntityType.TEAM,
                 entity_id=scope_team_id,
@@ -2835,25 +2872,47 @@ async def increment_spend_counters(
                 increment=cost,
                 request_started_at=request_started_at,
             )
+            return pending_window
 
-    async def _team_member_scope(scope_user_id: str, scope_team_id: str) -> None:
+        team_obj: Final[object] = await user_api_key_cache.async_get_cache(key=f"team_id:{scope_team_id}")
+        if team_obj is None:
+            return team_pending
+        team_budget_limits = getattr(team_obj, "budget_limits", None) or (
+            team_obj.get("budget_limits") if isinstance(team_obj, dict) else None
+        )
+        if isinstance(team_budget_limits, str):
+            team_budget_limits = json.loads(team_budget_limits)
+        if not isinstance(team_budget_limits, list):
+            return team_pending
+        window_pending: Final = await asyncio.gather(
+            *(_team_window_increment(window) for window in team_budget_limits), return_exceptions=True
+        )
+        return team_pending + tuple(item for item in window_pending if item is not None)
+
+    async def _team_member_scope(
+        scope_user_id: str, scope_team_id: str
+    ) -> tuple[_PendingSpendIncrement | BaseException, ...]:
         team_member_counter_key: Final = f"spend:team_member:{scope_user_id}:{scope_team_id}"
         if team_member_counter_key in reserved_counter_keys:
-            return
-        await _init_and_increment_spend_counter(
-            counter_key=team_member_counter_key,
-            source_cache_key=f"team_membership:{scope_user_id}:{scope_team_id}",
-            increment=cost,
+            return ()
+        return (
+            await _prepare_spend_counter_increment(
+                counter_key=team_member_counter_key,
+                source_cache_key=f"team_membership:{scope_user_id}:{scope_team_id}",
+                increment=cost,
+            ),
         )
 
-    async def _user_scope(scope_user_id: str) -> None:
+    async def _user_scope(scope_user_id: str) -> tuple[_PendingSpendIncrement | BaseException, ...]:
         user_counter_key: Final = f"spend:user:{scope_user_id}"
         if user_counter_key in reserved_counter_keys:
-            return
-        await _init_and_increment_spend_counter(
-            counter_key=user_counter_key,
-            source_cache_key=scope_user_id,
-            increment=cost,
+            return ()
+        return (
+            await _prepare_spend_counter_increment(
+                counter_key=user_counter_key,
+                source_cache_key=scope_user_id,
+                increment=cost,
+            ),
         )
 
     scope_coros: Final = tuple(
@@ -2863,7 +2922,7 @@ async def increment_spend_counters(
             _team_scope(team_id) if team_id is not None else None,
             _team_member_scope(user_id, team_id) if user_id is not None and team_id is not None else None,
             _user_scope(user_id) if user_id is not None else None,
-            _increment_end_user_and_tag_spend_counters(
+            _prepare_end_user_and_tag_spend_increments(
                 end_user_id=end_user_id,
                 tags=tags,
                 response_cost=cost,
@@ -2871,14 +2930,14 @@ async def increment_spend_counters(
             )
             if end_user_id is not None or tags is not None
             else None,
-            _increment_model_access_group_spend_counters(
+            _prepare_model_access_group_spend_increments(
                 model_access_groups=model_access_groups,
                 response_cost=cost,
                 reserved_counter_keys=reserved_counter_keys,
             )
             if model_access_groups
             else None,
-            _increment_org_spend_counter(
+            _prepare_org_spend_increment(
                 org_id=org_id,
                 response_cost=cost,
                 reserved_counter_keys=reserved_counter_keys,
@@ -2893,7 +2952,20 @@ async def increment_spend_counters(
     # as orphaned tasks that race the caller's reservation-counter invalidation;
     # all scopes settle, then the first error propagates as before.
     scope_results: Final = await asyncio.gather(*scope_coros, return_exceptions=True)
-    scope_errors: Final = [r for r in scope_results if isinstance(r, BaseException)]
+    scope_errors: Final = tuple(
+        item
+        for scope in scope_results
+        for item in (scope if isinstance(scope, tuple) else (scope,))
+        if isinstance(item, BaseException)
+    )
+    pending: Final = tuple(
+        item
+        for scope in scope_results
+        if not isinstance(scope, BaseException)
+        for item in scope
+        if not isinstance(item, BaseException)
+    )
+    await _apply_spend_counter_increments(pending=pending)
     if scope_errors:
         raise scope_errors[0]
 
@@ -2936,41 +3008,49 @@ async def _reconcile_budget_reservation_for_counter_update(
     return reserved_counter_keys
 
 
-async def _increment_end_user_and_tag_spend_counters(
+async def _prepare_end_user_and_tag_spend_increments(
     end_user_id: str | None,
     tags: list[str] | None,
     response_cost: float,
     reserved_counter_keys: set[str],
-) -> None:
-    if end_user_id is not None:
-        await _init_and_increment_unreserved_spend_counter(
-            counter_key=f"spend:end_user:{end_user_id}",
-            source_cache_key=end_user_cache_key(end_user_id),
-            increment=response_cost,
-            reserved_counter_keys=reserved_counter_keys,
-        )
-
-    if tags is None:
-        return
-
-    seen_tags: Final[set[str]] = set()
-    for tag_name in tags:
-        if not tag_name or not isinstance(tag_name, str) or tag_name in seen_tags:
-            continue
-        seen_tags.add(tag_name)
-        await _init_and_increment_unreserved_spend_counter(
-            counter_key=f"spend:tag:{tag_name}",
-            source_cache_key=tag_cache_key(tag_name),
-            increment=response_cost,
-            reserved_counter_keys=reserved_counter_keys,
-        )
+) -> tuple[_PendingSpendIncrement | BaseException, ...]:
+    unique_tags: Final = (
+        tuple(dict.fromkeys(tag for tag in tags if tag and isinstance(tag, str))) if tags is not None else ()
+    )
+    results: Final = await asyncio.gather(
+        *(
+            coro
+            for coro in (
+                _prepare_unreserved_spend_counter_increment(
+                    counter_key=f"spend:end_user:{end_user_id}",
+                    source_cache_key=end_user_cache_key(end_user_id),
+                    increment=response_cost,
+                    reserved_counter_keys=reserved_counter_keys,
+                )
+                if end_user_id is not None
+                else None,
+                *(
+                    _prepare_unreserved_spend_counter_increment(
+                        counter_key=f"spend:tag:{tag_name}",
+                        source_cache_key=tag_cache_key(tag_name),
+                        increment=response_cost,
+                        reserved_counter_keys=reserved_counter_keys,
+                    )
+                    for tag_name in unique_tags
+                ),
+            )
+            if coro is not None
+        ),
+        return_exceptions=True,
+    )
+    return tuple(item for item in results if item is not None)
 
 
-async def _increment_model_access_group_spend_counters(
+async def _prepare_model_access_group_spend_increments(
     model_access_groups: Sequence[object],
     response_cost: float,
     reserved_counter_keys: set[str],
-) -> None:
+) -> tuple[_PendingSpendIncrement | BaseException, ...]:
     """Charge the model access groups that authorized this request.
 
     Without this the counter auth reads is written only by the reservation path, so
@@ -2984,55 +3064,63 @@ async def _increment_model_access_group_spend_counters(
     unique_groups: Final = tuple(
         dict.fromkeys(group for group in model_access_groups if group and isinstance(group, str))
     )
-    for group in unique_groups:
-        await _init_and_increment_unreserved_spend_counter(
-            counter_key=model_access_group_spend_counter_key(group),
-            source_cache_key=model_access_group_cache_key(group),
-            increment=response_cost,
-            reserved_counter_keys=reserved_counter_keys,
-        )
+    results: Final = await asyncio.gather(
+        *(
+            _prepare_unreserved_spend_counter_increment(
+                counter_key=model_access_group_spend_counter_key(group),
+                source_cache_key=model_access_group_cache_key(group),
+                increment=response_cost,
+                reserved_counter_keys=reserved_counter_keys,
+            )
+            for group in unique_groups
+        ),
+        return_exceptions=True,
+    )
+    return tuple(item for item in results if item is not None)
 
 
-async def _increment_org_spend_counter(
+async def _prepare_org_spend_increment(
     org_id: str | None,
     response_cost: float,
     reserved_counter_keys: set[str],
-) -> None:
+) -> tuple[_PendingSpendIncrement, ...]:
     if org_id is None:
-        return
+        return ()
 
-    await _init_and_increment_unreserved_spend_counter(
+    pending: Final = await _prepare_unreserved_spend_counter_increment(
         counter_key=f"spend:org:{org_id}",
         source_cache_key=[f"org_id:{org_id}:with_budget", f"org_id:{org_id}"],
         increment=response_cost,
         reserved_counter_keys=reserved_counter_keys,
     )
+    return (pending,) if pending is not None else ()
 
 
-async def _init_and_increment_unreserved_spend_counter(
+async def _prepare_unreserved_spend_counter_increment(
     counter_key: str,
     source_cache_key: str | list[str],
     increment: float,
     reserved_counter_keys: set[str],
-) -> None:
+) -> _PendingSpendIncrement | None:
     if counter_key in reserved_counter_keys:
-        return
+        return None
 
-    await _init_and_increment_spend_counter(
+    return await _prepare_spend_counter_increment(
         counter_key=counter_key,
         source_cache_key=source_cache_key,
         increment=increment,
     )
 
 
-async def _init_and_increment_spend_counter(
+async def _prepare_spend_counter_increment(
     counter_key: str,
     source_cache_key: str | list[str],
     increment: float,
-):
+) -> _PendingSpendIncrement:
     """
     Initialize counter from the authoritative DB spend value if not yet
-    set, then atomically increment in both in-memory and Redis.
+    set, then return the pending increment for the caller to apply in one
+    pipelined Redis call.
 
     On first access per pod:
     1. Check spend_counter_cache (in-memory -> Redis via DualCache)
@@ -3044,13 +3132,13 @@ async def _init_and_increment_spend_counter(
        the counter as absent and seed it. Using increment means the worst case
        is over-counting (conservative, blocks slightly early) rather than
        under-counting (would allow overspend).
-    4. Increment atomically (both in-memory + Redis)
+    4. Increment is returned for the caller to apply via pipeline
     """
     await _ensure_spend_counter_initialized(
         counter_key=counter_key,
         source_cache_key=source_cache_key,
     )
-    await _increment_spend_counter_cache(counter_key=counter_key, increment=increment)
+    return _PendingSpendIncrement(counter_key=counter_key, increment=increment)
 
 
 async def _enqueue_window_spend_row_update(
@@ -3102,20 +3190,20 @@ async def _enqueue_window_spend_row_update(
         )
 
 
-async def _init_and_increment_window_spend_counter(
+async def _prepare_window_spend_counter_increment(
     counter_key: str,
     entity_type: str,
     entity_id: str,
     window_duration: str | None,
     window_start: datetime | None,
     increment: float,
-):
+) -> _PendingSpendIncrement | None:
     if window_start is None:
         verbose_proxy_logger.warning(
             "Skipping spend counter increment for invalid budget window %s",
             counter_key,
         )
-        return
+        return None
 
     initialized: Final = await _ensure_window_spend_counter_initialized(
         counter_key=counter_key,
@@ -3125,8 +3213,8 @@ async def _init_and_increment_window_spend_counter(
         window_start=window_start,
     )
     if initialized is False:
-        return
-    await _increment_spend_counter_cache(counter_key=counter_key, increment=increment)
+        return None
+    return _PendingSpendIncrement(counter_key=counter_key, increment=increment)
 
 
 async def _ensure_spend_counter_initialized(
@@ -3259,6 +3347,32 @@ async def _invalidate_spend_counter(counter_key: str):
             )
 
 
+async def _apply_spend_counter_increments(pending: Sequence[_PendingSpendIncrement]) -> None:
+    if not pending:
+        return
+    redis_cache: Final = spend_counter_cache.redis_cache
+    if redis_cache is None:
+        for item in pending:
+            await spend_counter_cache.async_increment_cache(
+                key=item.counter_key,
+                value=item.increment,
+                refresh_ttl=True,
+            )
+        return
+    ttl: Final = redis_cache.get_ttl()
+    increment_list: Final = [  # mutable-ok: async_increment_pipeline signature requires list[RedisPipelineIncrementOperation]
+        RedisPipelineIncrementOperation(key=item.counter_key, increment_value=item.increment, ttl=ttl)
+        for item in pending
+    ]
+    try:
+        results: Final = await redis_cache.async_increment_pipeline(increment_list=increment_list)
+    except Exception:
+        await asyncio.gather(*(_invalidate_spend_counter(counter_key=item.counter_key) for item in pending))
+        raise
+    for item, current_value in zip(pending, results or ()):
+        spend_counter_cache.in_memory_cache.set_cache(key=item.counter_key, value=current_value)
+
+
 async def update_cache(
     token: str | None,
     user_id: str | None,
@@ -4436,20 +4550,9 @@ def resolve_classifier_plugin(
 
 
 def _swap_in_model_cost_map(new_model_cost_map: dict) -> int:
-    """Adopt a freshly fetched cost map into this process's litellm state, return the model count"""
-    litellm.model_cost = new_model_cost_map
-    # Invalidate case-insensitive lookup map since model_cost was replaced
-    _invalidate_model_cost_lowercase_map()
-    # Repopulate provider model sets (e.g. litellm.anthropic_models) so that
-    # wildcard patterns like "anthropic/*" include any newly added models.
-    litellm.add_known_models(model_cost_map=new_model_cost_map)
-    # Counted before the re-apply below, which writes into this same dict, so the
-    # number reported describes the fetched price data alone.
-    fetched_model_count: Final = len(new_model_cost_map) if new_model_cost_map else 0
-    # The swap discards everything registered at runtime (deployment model_info,
-    # register_model overrides), so put it back on top of the fresh catalog.
-    reapply_runtime_model_cost_registrations()
-    return fetched_model_count
+    from litellm.litellm_core_utils.get_model_cost_map import adopt_model_cost_map
+
+    return adopt_model_cost_map(new_model_cost_map)
 
 
 def should_load_db_object(object_type: str | SupportedDBObjectType) -> bool:
@@ -9543,7 +9646,7 @@ class ProxyStartupEvent:
             flush_gateway_requests,
             "interval",
             seconds=batch_writing_interval,
-            args=(prisma_client, gateway_request_accumulator),
+            args=(prisma_client, gateway_request_accumulator, _gateway_request_redis_buffer()),
             id="update_gateway_requests_job",
             replace_existing=True,
             misfire_grace_time=APSCHEDULER_MISFIRE_GRACE_TIME,
@@ -12714,7 +12817,9 @@ async def token_counter(request: TokenCountRequest, call_endpoint: bool = False)
             CustomHuggingfaceTokenizer | None,
             model_info.get("custom_tokenizer", None),
         )
-    _tokenizer_used: Final = litellm.utils._select_tokenizer(model=model_to_use, custom_tokenizer=custom_tokenizer)
+    _tokenizer_used: Final = await asyncify(litellm.utils._select_tokenizer)(
+        model=model_to_use, custom_tokenizer=custom_tokenizer
+    )
 
     tokenizer_used: Final = str(_tokenizer_used["type"])
     system_message: Final = _system_message(system)
@@ -12727,7 +12832,7 @@ async def token_counter(request: TokenCountRequest, call_endpoint: bool = False)
     counted_tools: Final = cast(  # cast-ok: raw OpenAI or Anthropic tool dicts, both of which token_counter formats
         list[ChatCompletionToolParam] | None, tools if counted_messages is not None else None
     )
-    total_tokens: Final = await asyncify(litellm.token_counter)(
+    total_tokens: Final = await offload_token_count(litellm.token_counter)(
         model=model_to_use,
         text=prompt,
         messages=counted_messages,

litellm/proxy/response_polling/background_streaming.py

@@ -15,6 +15,7 @@ from typing import TYPE_CHECKING, Final, TypeAlias
 
 from fastapi import Request, Response
 from fastapi.responses import StreamingResponse
+from starlette.types import Message
 from typing_extensions import ReadOnly, TypedDict
 
 from litellm._logging import verbose_proxy_logger
@@ -74,6 +75,20 @@ class _StreamEventParser:
     parse: Callable[[str], _StreamEvent] = staticmethod(json.loads)
 
 
+async def _never_receive() -> Message:
+    await asyncio.Event().wait()
+    raise AssertionError("unreachable")
+
+
+def detach_request_from_client(request: Request) -> Request:
+    """Same scope (headers, parsed body, auth) but a receive() that never yields http.disconnect.
+
+    The polling client closes its connection right after getting the polling id, so the
+    upstream call must not be cancelled by the client-disconnect guards.
+    """
+    return Request(request.scope, _never_receive)
+
+
 async def background_streaming_task(
     polling_id: str,
     data: dict[str, object],
@@ -123,7 +138,7 @@ async def background_streaming_task(
         # Pre-call checks (rate limits, guardrails, budget) were already run
         # before polling ID creation, so skip them here to avoid double-counting.
         response: Final[StreamingResponse] = await processor.base_process_llm_request(
-            request=request,
+            request=detach_request_from_client(request),
             fastapi_response=fastapi_response,
             user_api_key_dict=user_api_key_dict,
             route_type="aresponses",