Merge branch 'litellm_internal_staging' of https://github.com/BerriAI/litellm into litellm_decrease_anys_opus5_r2

# Conflicts:
#	tests/logging_callback_tests/gcs_pub_sub_body/spend_logs_payload.json

.github/workflows/test-unit.yml

@@ -103,6 +103,7 @@ jobs:
               tests/test_litellm/completion_extras
               tests/test_litellm/compression
               tests/test_litellm/containers
+              tests/test_litellm/endpoints
               tests/test_litellm/experimental_mcp_client
               tests/test_litellm/models
               tests/test_litellm/repositories

gateway/routes/allowlist.py

@@ -86,6 +86,7 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = (
     "/comprehendmedical",
     "/cohere/",
     "/gemini/",
+    "/gigachat/",
     "/google/",
     "/vertex_ai/",
     "/vertex-ai/",

litellm/__init__.py

@@ -7,6 +7,9 @@ warnings.filterwarnings("ignore", message=".*conflict with protected namespace.*
 # Suppress Pydantic 2.11+ deprecation warning about accessing model_fields on instances
 # This warning can accumulate during streaming and cause memory leaks
 warnings.filterwarnings("ignore", message=".*Accessing the.*attribute on the instance is deprecated.*")
+# ReadOnly on TypedDict fields is repo-wide static discipline (LIT012); pydantic warns it
+# cannot enforce it at runtime, which floods proxy boot once such a type is schema-walked
+warnings.filterwarnings("ignore", message=".*`ReadOnly` qualifier.*")
 ### INIT VARIABLES #########################
 import threading
 import os

litellm/constants.py

@@ -806,6 +806,7 @@ openai_compatible_endpoints: Final[list] = [
     "https://api.meta.ai/v1",
     "https://api.cognition.ai/v1",
     "https://api.scx.ai/v1",
+    "https://gigachat.devices.sberbank.ru/api/v1",
 ]
 
 

litellm/endpoints/speech/speech_to_completion_bridge/transformation.py

@@ -1,10 +1,14 @@
+from collections.abc import Mapping
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Final, cast
 
+from typing_extensions import NotRequired, ReadOnly, TypedDict
+
 from litellm.constants import OPENAI_CHAT_COMPLETION_PARAMS
 
 if TYPE_CHECKING:
     from litellm import Logging as LiteLLMLoggingObj
-    from litellm.types.llms.openai import HttpxBinaryResponseContent
+    from litellm.types.llms.openai import ChatCompletionUserMessage, HttpxBinaryResponseContent
     from litellm.types.utils import ModelResponse
 
 
@@ -16,7 +20,42 @@ def _completion_response_cost(model_response: "ModelResponse") -> float | None:
     return response_cost if isinstance(response_cost, float) else None
 
 
+GEMINI_TTS_CHAT_AUDIO_FORMAT: Final = "pcm16"
+
+
+class ChatAudioParam(TypedDict):
+    voice: ReadOnly[str]
+    format: ReadOnly[NotRequired[str]]
+
+
 class SpeechToCompletionBridgeTransformationHandler:
+    def _chat_completion_params(self, optional_params: Mapping[str, object]) -> Mapping[str, object]:
+        return MappingProxyType(
+            {
+                param: value
+                for param, value in optional_params.items()
+                if param in OPENAI_CHAT_COMPLETION_PARAMS and param != "response_format"
+            }
+        )
+
+    def _chat_audio_format(self, model: str, optional_params: Mapping[str, object]) -> str | None:
+        if self._is_gemini_tts_model(model):
+            return GEMINI_TTS_CHAT_AUDIO_FORMAT
+        response_format: Final = optional_params.get("response_format")
+        return response_format if isinstance(response_format, str) else None
+
+    def _chat_audio_param(
+        self, model: str, voice: str | Mapping[str, object] | None, optional_params: Mapping[str, object]
+    ) -> ChatAudioParam | None:
+        if not isinstance(voice, str):
+            return None
+        audio_format: Final = self._chat_audio_format(model, optional_params)
+        if audio_format is None:
+            voice_only: Final[ChatAudioParam] = {"voice": voice}
+            return voice_only
+        audio: Final[ChatAudioParam] = {"voice": voice, "format": audio_format}
+        return audio
+
     def transform_request(
         self,
         model: str,
@@ -28,36 +67,19 @@ class SpeechToCompletionBridgeTransformationHandler:
         litellm_logging_obj: "LiteLLMLoggingObj",
         custom_llm_provider: str,
     ) -> dict:
-        passed_optional_params: Final = {}
-        for op in optional_params:
-            if op in OPENAI_CHAT_COMPLETION_PARAMS:
-                passed_optional_params[op] = optional_params[op]
-
-        if voice is not None:
-            if isinstance(voice, str):
-                passed_optional_params["audio"] = {"voice": voice}
-                if "response_format" in optional_params:
-                    passed_optional_params["audio"]["format"] = optional_params["response_format"]
-
-        return_kwargs = {
+        user_message: Final[ChatCompletionUserMessage] = {"role": "user", "content": input}
+        return_kwargs: Final = {
             "model": model,
-            "messages": [
-                {
-                    "role": "user",
-                    "content": input,
-                }
-            ],
+            "messages": [user_message],
             "modalities": ["audio"],
-            **passed_optional_params,
+            **self._chat_completion_params(optional_params),
+            "audio": self._chat_audio_param(model, voice, optional_params),
             **litellm_params,
             "headers": headers,
             "litellm_logging_obj": litellm_logging_obj,
             "custom_llm_provider": custom_llm_provider,
         }
-
-        # filter out None values
-        return_kwargs = {k: v for k, v in return_kwargs.items() if v is not None}
-        return return_kwargs
+        return {k: v for k, v in return_kwargs.items() if v is not None}
 
     def _convert_pcm16_to_wav(self, pcm_data: bytes, sample_rate: int = 24000, channels: int = 1) -> bytes:
         """

litellm/google_genai/adapters/transformation.py

@@ -108,7 +108,6 @@ class GoogleGenAIStreamWrapper(AdapterCompletionStreamWrapper):
 
     def __init__(self, completion_stream: object):
         self.sent_first_chunk = False
-        # State tracking for accumulating partial tool calls
         self.accumulated_tool_calls = dict[int, _ToolCallAccumulator]()
         self._returned_response = False
         super().__init__(completion_stream)

litellm/litellm_core_utils/get_litellm_params.py

@@ -50,6 +50,9 @@ OPTIONAL_KWARGS_KEYS: Final = (
             "vertex_ai_project",
             "vertex_ai_location",
             "vertex_ai_credentials",
+            "gigachat_scope",
+            "gigachat_auth_url",
+            "gigachat_access_token",
             "tpm",
             "rpm",
             "itpm",

litellm/litellm_core_utils/get_llm_provider_logic.py

@@ -369,6 +369,9 @@ def get_llm_provider(
                     elif endpoint == "https://api.meta.ai/v1":
                         custom_llm_provider = "meta"
                         dynamic_api_key = get_secret_str("META_API_KEY")
+                    elif endpoint == "https://gigachat.devices.sberbank.ru/api/v1":
+                        custom_llm_provider = "gigachat"
+                        dynamic_api_key = get_secret_str("GIGACHAT_API_KEY")
                     elif (json_provider := JSONProviderRegistry.get_by_base_url(endpoint)) is not None:
                         custom_llm_provider = json_provider.slug
                         dynamic_api_key = api_key if api_key is not None else get_secret_str(json_provider.api_key_env)
@@ -867,6 +870,9 @@ def _get_openai_compatible_provider_info(
         # Manus is OpenAI compatible for responses API
         api_base = api_base or get_secret_str("MANUS_API_BASE") or "https://api.manus.im"
         dynamic_api_key = api_key or get_secret_str("MANUS_API_KEY")
+    elif custom_llm_provider == "gigachat":
+        api_base = api_base or get_secret_str("GIGACHAT_API_BASE") or "https://gigachat.devices.sberbank.ru/api/v1"
+        dynamic_api_key = api_key or get_secret_str("GIGACHAT_API_KEY")
 
     if api_base is not None and not isinstance(api_base, str):
         raise Exception(f"api base needs to be a string. api_base={api_base}")

litellm/litellm_core_utils/litellm_logging.py

@@ -2141,6 +2141,9 @@ class Logging(LiteLLMLoggingBaseClass):
 
             logging_result: Final = self.normalize_logging_result(result=result)
 
+            if isinstance(result, Response) and isinstance(logging_result, (ModelResponse, EmbeddingResponse)):
+                result = logging_result
+
             if standard_logging_object is None and result is not None and self.stream is not True:
                 if self._is_recognized_call_type_for_logging(logging_result=logging_result) or isinstance(
                     logging_result, (dict, list)
@@ -6152,7 +6155,10 @@ def get_standard_logging_object_payload(
 
 def emit_standard_logging_payload(payload: StandardLoggingPayload):
     if os.getenv("LITELLM_PRINT_STANDARD_LOGGING_PAYLOAD"):
-        print(json.dumps(payload, indent=4), flush=True)  # noqa: T201
+        try:
+            print(json.dumps(payload, indent=4, default=str), flush=True)  # noqa: T201
+        except Exception as e:  # noqa: BLE001 # Safe catch-all for verbose logging
+            verbose_logger.exception("Error serializing standard logging payload for debug output: %s", e)
 
 
 def get_standard_logging_metadata(

litellm/litellm_core_utils/model_response_utils.py

@@ -144,7 +144,6 @@ def _is_choice_non_empty(choice: StreamingChoices) -> bool:
     # Check model_extra for dynamically added fields on the choice
     choice_extra_fields: Final[Mapping[str, object]] = choice.model_extra or {}
     for extra_field_name, extra_field_value in choice_extra_fields.items():
-        # Skip certain structural fields that are just default/None placeholders
         if extra_field_name == "index" and extra_field_value == 0:
             continue
         if extra_field_name in {"finish_reason", "logprobs"} and extra_field_value is None:
@@ -192,7 +191,6 @@ def _is_delta_non_empty(delta: Delta) -> bool:
     # Check model_extra for dynamically added fields (this is where Pydantic stores them)
     delta_extra_fields: Final[Mapping[str, object]] = delta.model_extra or {}
     for extra_field_value in delta_extra_fields.values():
-        # Even structural fields are meaningful if they have actual content
         if _has_meaningful_content(extra_field_value):
             return True
 

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

@@ -859,12 +859,28 @@ class AnthropicMessagesHandler(BaseTranslation):
 
     @staticmethod
     def _image_sources(block: Mapping[str, object]) -> tuple[str, ...]:
+        """Normalize an Anthropic image block into strings a guardrail can read.
+
+        base64 becomes a data URI so the format travels with the payload, which is what
+        the OpenAI path already puts in this field. A file source yields nothing: those
+        bytes live behind the Files API and this extractor has no client to fetch them.
+        """
         source: Final = block.get("source")
         if not isinstance(source, Mapping):
             return ()
-        # Could be base64 or url
+
+        source_type: Final = source.get("type")
+        if source_type == "url":
+            url: Final = source.get("url")
+            return (url,) if isinstance(url, str) and url else ()
+
         data: Final = source.get("data")
-        return (data,) if data else ()
+        if not isinstance(data, str) or not data:
+            return ()
+        media_type: Final = source.get("media_type")
+        if isinstance(media_type, str) and media_type:
+            return (f"data:{media_type};base64,{data}",)
+        return (data,)
 
     async def _apply_guardrail_responses_to_input(
         self,

litellm/llms/bedrock/realtime/handler.py

@@ -7,13 +7,18 @@ This uses aws_sdk_bedrock_runtime for bidirectional streaming with Nova Sonic.
 import asyncio
 import contextlib
 import json
+from collections.abc import AsyncIterator, Mapping
 from typing import Final, Protocol
 
 from pydantic import JsonValue, TypeAdapter
 
+import litellm
 from litellm._logging import _redact_string, verbose_proxy_logger
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
+from litellm.litellm_core_utils.realtime_streaming import DefaultLoggedRealTimeEventTypes
+from litellm.types.llms.openai import OpenAIRealtimeEvents
 from litellm.types.realtime import RealtimeResponseTransformInput
 
 from ..base_aws_llm import BaseAWSLLM
@@ -32,6 +37,17 @@ def _json_str(value: JsonValue) -> str | None:
     return value if isinstance(value, str) else None
 
 
+def _should_log_event(openai_message: Mapping[str, object]) -> bool:
+    logged_types: Final = (
+        litellm.logged_real_time_event_types
+        if litellm.logged_real_time_event_types is not None
+        else DefaultLoggedRealTimeEventTypes
+    )
+    if logged_types == "*":
+        return True
+    return openai_message.get("type") in logged_types
+
+
 class RealtimeClientWebSocket(Protocol):
     """The client-facing websocket surface the realtime bridge talks to."""
 
@@ -205,16 +221,22 @@ class BedrockRealtime(BaseAWSLLM):
                 )
             )
 
-            bedrock_to_client_task: Final = asyncio.create_task(
-                self._forward_bedrock_to_client(
-                    bedrock_stream,
-                    websocket,
-                    transformation_config,
-                    model,
-                    logging_obj,
-                    session_state,
+            async def forward_bedrock_and_collect_logged_events() -> tuple[OpenAIRealtimeEvents, ...]:
+                return tuple(
+                    [
+                        event
+                        async for event in self._forward_bedrock_to_client(
+                            bedrock_stream,
+                            websocket,
+                            transformation_config,
+                            model,
+                            logging_obj,
+                            session_state,
+                        )
+                    ]
                 )
-            )
+
+            bedrock_to_client_task: Final = asyncio.create_task(forward_bedrock_and_collect_logged_events())
 
             # Wait for both tasks to complete
             await asyncio.gather(
@@ -223,6 +245,27 @@ class BedrockRealtime(BaseAWSLLM):
                 return_exceptions=True,
             )
 
+            forwarded_logged_events: Final = (
+                bedrock_to_client_task.result()
+                if not bedrock_to_client_task.cancelled() and bedrock_to_client_task.exception() is None
+                else ()
+            )
+            logged_events: Final = (
+                *forwarded_logged_events,
+                *(
+                    leftover_event
+                    for leftover_event in transformation_config.leftover_usage_done_events()
+                    if _should_log_event(leftover_event)
+                ),
+            )
+            if logged_events:
+                GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(
+                    logging_obj.dispatch_success_handlers(
+                        list(logged_events),  # mutable-ok: realtime spend logging requires a list result
+                        prefer_async_handlers=True,
+                    )
+                )
+
         except Exception as e:
             verbose_proxy_logger.exception("Error in BedrockRealtime.async_realtime: %s", e)
             try:
@@ -304,8 +347,8 @@ class BedrockRealtime(BaseAWSLLM):
         model: str,
         logging_obj: LiteLLMLogging,
         session_state: RealtimeResponseTransformInput,
-    ):
-        """Forward messages from Bedrock stream to client WebSocket."""
+    ) -> AsyncIterator[OpenAIRealtimeEvents]:
+        """Forward messages from Bedrock to the client, yielding the ones to record for spend logging."""
         try:
             while True:
                 # Receive from Bedrock
@@ -353,11 +396,14 @@ class BedrockRealtime(BaseAWSLLM):
                     )
 
                     # Send transformed messages to client
-                    openai_messages = transformed_response.get("response", [])
+                    response_value = transformed_response["response"]
+                    openai_messages = response_value if isinstance(response_value, list) else (response_value,)
                     for openai_message in openai_messages:
                         message_json = json.dumps(openai_message)
                         await client_ws.send_text(message_json)
                         verbose_proxy_logger.debug("Bedrock Realtime: Sent to client: %s", message_json[:200])
+                        if _should_log_event(openai_message):
+                            yield openai_message
 
         except Exception as e:
             verbose_proxy_logger.debug("Bedrock to client forwarding ended: %s", e, exc_info=True)

litellm/llms/bedrock/realtime/transformation.py

@@ -7,7 +7,7 @@ Transforms between OpenAI Realtime API format and Bedrock Nova Sonic format.
 import base64
 import json
 import uuid as uuid_lib
-from typing import Any, Final
+from typing import Any, Final, cast
 
 from pydantic import BaseModel
 
@@ -20,29 +20,54 @@ from litellm.types.llms.openai import (
     OpenAIRealtimeContentPartDone,
     OpenAIRealtimeDoneEvent,
     OpenAIRealtimeEvents,
+    OpenAIRealtimeInputAudioBufferSpeechEvent,
+    OpenAIRealtimeInputAudioTranscriptionCompleted,
+    OpenAIRealtimeInputAudioTranscriptionDelta,
     OpenAIRealtimeOutputItemDone,
     OpenAIRealtimeResponseAudioDone,
     OpenAIRealtimeResponseContentPartAdded,
     OpenAIRealtimeResponseDelta,
     OpenAIRealtimeResponseDoneObject,
     OpenAIRealtimeResponseTextDone,
+    OpenAIRealtimeResponseUsage,
     OpenAIRealtimeStreamResponseBaseObject,
     OpenAIRealtimeStreamResponseOutputItemAdded,
     OpenAIRealtimeStreamSession,
     OpenAIRealtimeStreamSessionEvents,
+    OpenAIRealtimeUsageTokenDetails,
 )
 from litellm.types.realtime import (
     ALL_DELTA_TYPES,
     RealtimeResponseTransformInput,
     RealtimeResponseTypedDict,
 )
-from litellm.utils import get_empty_usage
 
 
 class BedrockContentEnd(BaseModel):
     stopReason: str | None = None
 
 
+class BedrockUsageTokenDetails(BaseModel):
+    speechTokens: int = 0
+    textTokens: int = 0
+
+
+class BedrockUsageDetailsTotal(BaseModel):
+    input: BedrockUsageTokenDetails = BedrockUsageTokenDetails()
+    output: BedrockUsageTokenDetails = BedrockUsageTokenDetails()
+
+
+class BedrockUsageDetails(BaseModel):
+    total: BedrockUsageDetailsTotal = BedrockUsageDetailsTotal()
+
+
+class BedrockUsageEvent(BaseModel):
+    totalInputTokens: int = 0
+    totalOutputTokens: int = 0
+    totalTokens: int = 0
+    details: BedrockUsageDetails = BedrockUsageDetails()
+
+
 TRIGGER_AUDIO_SAMPLE_RATE_HERTZ: Final = 16000
 TRIGGER_AUDIO_BYTES_PER_SECOND: Final = TRIGGER_AUDIO_SAMPLE_RATE_HERTZ * 2
 TRIGGER_LEADING_SILENCE: Final = bytes(TRIGGER_AUDIO_BYTES_PER_SECOND // 2)
@@ -87,6 +112,15 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
         # Text configuration
         self.text_media_type = "text/plain"
 
+        # Response-stream state (Bedrock events carry no role on textOutput,
+        # so the USER/ASSISTANT split from contentStart is tracked here)
+        self._user_transcript_active = False
+        self._user_transcript_generation_stage: str | None = None
+        self._user_item_id: str | None = None
+        self._user_transcript_buffer = ""
+        self._cumulative_usage = BedrockUsageEvent()
+        self._reported_usage = BedrockUsageEvent()
+
     def validate_environment(self, headers: dict, model: str, api_key: str | None = None) -> dict:
         """Validate environment - no special validation needed for Bedrock."""
         return headers
@@ -691,6 +725,11 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
         role: Final = content_start.get("role")
 
         if role != "ASSISTANT":
+            if role == "USER" and content_start.get("type") == "TEXT":
+                self._user_transcript_active = True
+                self._user_transcript_generation_stage = self._parse_generation_stage(
+                    content_start.get("additionalModelFields")
+                )
             return (
                 [],
                 current_response_id,
@@ -700,6 +739,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
             )
 
         verbose_logger.debug("Handling ASSISTANT contentStart")
+        is_new_response: Final = current_response_id is None
 
         # Initialize IDs if needed
         if not current_response_id:
@@ -715,7 +755,8 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
 
         returned_messages: Final[list[OpenAIRealtimeEvents]] = []
 
-        # Send response.created
+        # Send response.created only once per response (a response can contain
+        # multiple content blocks, e.g. TEXT then AUDIO)
         response_created: Final = OpenAIRealtimeStreamResponseBaseObject(
             type="response.created",
             event_id=f"event_{uuid.uuid4()}",
@@ -727,7 +768,8 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
                 "conversation_id": current_conversation_id,
             },
         )
-        returned_messages.append(response_created)
+        if is_new_response:
+            returned_messages.append(response_created)
 
         # Send response.output_item.added
         output_item_added: Final = OpenAIRealtimeStreamResponseOutputItemAdded(
@@ -767,6 +809,108 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
             current_delta_type,
         )
 
+    @staticmethod
+    def _parse_generation_stage(additional_model_fields: object) -> str | None:
+        if not isinstance(additional_model_fields, str):
+            return None
+        try:
+            parsed: Final = json.loads(additional_model_fields)
+        except json.JSONDecodeError:
+            return None
+        stage: Final = parsed.get("generationStage") if isinstance(parsed, dict) else None
+        return stage if isinstance(stage, str) else None
+
+    def _current_user_item_id(self, new_utterance: bool = False) -> str:
+        """Item id shared by all events of one user utterance (speech boundaries and transcript)."""
+        if new_utterance or self._user_item_id is None:
+            self._user_item_id = f"item_{uuid.uuid4()}"
+        return self._user_item_id
+
+    def transform_user_speech_event(self, is_speech_start: bool) -> tuple[OpenAIRealtimeEvents, ...]:
+        """Transform Bedrock userSpeechStart/userSpeechEnd to OpenAI speech boundary events."""
+        verbose_logger.debug("Handling userSpeech%s", "Start" if is_speech_start else "End")
+        speech_event: Final[OpenAIRealtimeInputAudioBufferSpeechEvent] = {
+            "type": "input_audio_buffer.speech_started" if is_speech_start else "input_audio_buffer.speech_stopped",
+            "event_id": f"event_{uuid.uuid4()}",
+            "item_id": self._current_user_item_id(new_utterance=is_speech_start),
+        }
+        return (speech_event,)
+
+    def transform_usage_event(self, usage_event: BedrockUsageEvent) -> None:
+        """Record Bedrock's session-cumulative usage totals for the next response.done."""
+        verbose_logger.debug("Handling usageEvent")
+        self._cumulative_usage = usage_event
+
+    def _take_usage_delta(self) -> OpenAIRealtimeResponseUsage:
+        """Usage for the response now completing: cumulative totals minus what prior response.done events reported."""
+        prior: Final = self._reported_usage
+        latest: Final = self._cumulative_usage
+        self._reported_usage = latest
+        input_details: Final[OpenAIRealtimeUsageTokenDetails] = {
+            "audio_tokens": latest.details.total.input.speechTokens - prior.details.total.input.speechTokens,
+            "text_tokens": latest.details.total.input.textTokens - prior.details.total.input.textTokens,
+            "cached_tokens": 0,
+        }
+        output_details: Final[OpenAIRealtimeUsageTokenDetails] = {
+            "audio_tokens": latest.details.total.output.speechTokens - prior.details.total.output.speechTokens,
+            "text_tokens": latest.details.total.output.textTokens - prior.details.total.output.textTokens,
+        }
+        usage_delta: Final[OpenAIRealtimeResponseUsage] = {
+            "input_tokens": latest.totalInputTokens - prior.totalInputTokens,
+            "output_tokens": latest.totalOutputTokens - prior.totalOutputTokens,
+            "total_tokens": latest.totalTokens - prior.totalTokens,
+            "input_token_details": input_details,
+            "output_token_details": output_details,
+        }
+        return usage_delta
+
+    def leftover_usage_done_events(self) -> tuple[OpenAIRealtimeEvents, ...]:
+        """Logged-only response.done for usage Bedrock reports after the final turn's contentEnd."""
+        if self._cumulative_usage == self._reported_usage:
+            return ()
+        usage: Final = self._take_usage_delta()
+        leftover_done: Final = OpenAIRealtimeDoneEvent(
+            type="response.done",
+            event_id=f"event_{uuid.uuid4()}",
+            response=OpenAIRealtimeResponseDoneObject(
+                object="realtime.response",
+                id=f"resp_{uuid.uuid4()}",
+                status="completed",
+                conversation_id=f"conv_{uuid.uuid4()}",
+                usage=dict(usage),  # mutable-ok: OpenAIRealtimeResponseDoneObject types usage as plain dict
+            ),
+        )
+        return (leftover_done,)
+
+    def transform_user_transcript_event(self, transcript: str) -> tuple[OpenAIRealtimeEvents, ...]:
+        """Transform a USER-role Bedrock textOutput (ASR transcript) to an OpenAI transcription delta."""
+        verbose_logger.debug("Handling USER textOutput (ASR transcript)")
+        delta_event: Final[OpenAIRealtimeInputAudioTranscriptionDelta] = {
+            "type": "conversation.item.input_audio_transcription.delta",
+            "event_id": f"event_{uuid.uuid4()}",
+            "item_id": self._current_user_item_id(),
+            "content_index": 0,
+            "delta": transcript,
+        }
+        if self._user_transcript_generation_stage != "SPECULATIVE":
+            self._user_transcript_buffer += transcript
+        return (delta_event,)
+
+    def user_transcript_completed_events(self) -> tuple[OpenAIRealtimeEvents, ...]:
+        """One completed event with the full transcript once the FINAL user content block ends."""
+        transcript: Final = self._user_transcript_buffer
+        if not transcript:
+            return ()
+        self._user_transcript_buffer = ""
+        completed_event: Final[OpenAIRealtimeInputAudioTranscriptionCompleted] = {
+            "type": "conversation.item.input_audio_transcription.completed",
+            "event_id": f"event_{uuid.uuid4()}",
+            "item_id": self._current_user_item_id(),
+            "content_index": 0,
+            "transcript": transcript,
+        }
+        return (completed_event,)
+
     def transform_text_output_event(
         self,
         event: dict,
@@ -985,7 +1129,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
         if not current_response_id or not current_conversation_id:
             return [], None, None, None
 
-        usage_obj: Final = get_empty_usage()
+        usage: Final = self._take_usage_delta()
         response_done: Final = OpenAIRealtimeDoneEvent(
             type="response.done",
             event_id=f"event_{uuid.uuid4()}",
@@ -995,11 +1139,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
                 status="completed",
                 output=[],
                 conversation_id=current_conversation_id,
-                usage={
-                    "prompt_tokens": usage_obj.prompt_tokens,
-                    "completion_tokens": usage_obj.completion_tokens,
-                    "total_tokens": usage_obj.total_tokens,
-                },
+                usage=dict(usage),
             ),
         )
 
@@ -1042,8 +1182,6 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
 
         # Create a function call arguments done event
         # This is a custom event format that matches what clients expect
-        from typing import cast
-
         function_call_event: Final[dict[str, Any]] = {
             "type": "response.function_call_arguments.done",
             "event_id": f"event_{uuid.uuid4()}",
@@ -1194,18 +1332,26 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
             returned_messages.extend(events)
 
         elif "textOutput" in event:
-            events, current_delta_chunks = self.transform_text_output_event(
-                event,
-                current_output_item_id,
-                current_response_id,
-                current_delta_chunks,
-            )
-            returned_messages.extend(events)
+            if self._user_transcript_active:
+                returned_messages.extend(self.transform_user_transcript_event(event["textOutput"].get("content", "")))
+            else:
+                events, current_delta_chunks = self.transform_text_output_event(
+                    event,
+                    current_output_item_id,
+                    current_response_id,
+                    current_delta_chunks,
+                )
+                returned_messages.extend(events)
 
         elif "audioOutput" in event:
             events = self.transform_audio_output_event(event, current_output_item_id, current_response_id)
             returned_messages.extend(events)
 
+        elif "contentEnd" in event and self._user_transcript_active:
+            self._user_transcript_active = False
+            self._user_transcript_generation_stage = None
+            returned_messages.extend(self.user_transcript_completed_events())
+
         elif "contentEnd" in event:
             events, current_delta_chunks = self.transform_content_end_event(
                 event,
@@ -1224,6 +1370,12 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
                 ) = self._response_done_events(current_response_id, current_conversation_id)
                 returned_messages.extend(done_events)
 
+        elif "userSpeechStart" in event or "userSpeechEnd" in event:
+            returned_messages.extend(self.transform_user_speech_event("userSpeechStart" in event))
+
+        elif "usageEvent" in event:
+            self.transform_usage_event(BedrockUsageEvent.model_validate(event["usageEvent"]))
+
         elif "toolUse" in event:
             events, tool_call_id, tool_name = self.transform_tool_use_event(
                 event, current_output_item_id, current_response_id

litellm/llms/gigachat/__init__.py

@@ -15,9 +15,11 @@ API Documentation: https://developers.sber.ru/docs/ru/gigachat/api/overview
 
 from .chat.transformation import GigaChatConfig, GigaChatError
 from .embedding.transformation import GigaChatEmbeddingConfig
+from .passthrough.transformation import GigaChatPassthroughConfig
 
-__all__ = [
+__all__ = (
     "GigaChatConfig",
     "GigaChatEmbeddingConfig",
     "GigaChatError",
-]
+    "GigaChatPassthroughConfig",
+)

litellm/llms/gigachat/authenticator.py

@@ -7,6 +7,7 @@ Based on official GigaChat SDK authentication flow.
 
 import time
 import uuid
+from collections.abc import Mapping
 from typing import Final
 
 import httpx
@@ -16,7 +17,7 @@ from litellm.caching.caching import InMemoryCache
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.custom_httpx.http_handler import (
     HTTPHandler,
-    _get_httpx_client,
+    _get_httpx_client,  # pyright: ignore[reportPrivateUsage]  # house cached-client factory has no public alias
     get_async_httpx_client,
 )
 from litellm.secret_managers.main import get_secret_str
@@ -63,6 +64,7 @@ def get_access_token(
     credentials: str | None = None,
     scope: str | None = None,
     auth_url: str | None = None,
+    litellm_params: Mapping[str, object] | None = None,
 ) -> str:
     """
     Get valid access token, using cache if available.
@@ -78,71 +80,88 @@ def get_access_token(
     Raises:
         GigaChatAuthError: If authentication fails
     """
-    credentials = credentials or _get_credentials()
-    if not credentials:
+    if not litellm_params:
+        litellm_params = {}  # mutable-ok: empty dict default; rebind-ok: provide default
+
+    access_token: Final = litellm_params.get("gigachat_access_token") or get_secret_str("GIGACHAT_ACCESS_TOKEN")
+    if access_token:
+        return access_token
+
+    effective_credentials: Final = credentials or _get_credentials()
+    if not effective_credentials:
         raise GigaChatAuthError(
             status_code=401,
             message="GigaChat credentials not provided. Set GIGACHAT_CREDENTIALS or GIGACHAT_API_KEY environment variable.",
         )
 
-    scope = scope or _get_scope()
-    auth_url = auth_url or _get_auth_url()
+    effective_scope: Final = scope or litellm_params.get("gigachat_scope") or _get_scope()
+    effective_auth_url: Final = auth_url or litellm_params.get("gigachat_auth_url") or _get_auth_url()
 
     # Check cache
-    cache_key: Final = f"gigachat_token:{credentials[:16]}"
+    cache_key: Final = f"gigachat_token:{effective_credentials[:16]}"
     cached: Final = _token_cache.get_cache(cache_key)
     if cached:
-        token, expires_at = cached
+        _token, _expires_at = cached
         # Check if token is still valid (with buffer)
-        if time.time() * 1000 < expires_at - TOKEN_EXPIRY_BUFFER_MS:
+        if time.time() * 1000 < _expires_at - TOKEN_EXPIRY_BUFFER_MS:
             verbose_logger.debug("Using cached GigaChat access token")
-            return token
+            return _token
 
     # Request new token
-    token, expires_at = _request_token_sync(credentials, scope, auth_url)
+    new_token, new_expires_at = _request_token_sync(effective_credentials, effective_scope, effective_auth_url)  # pyright: ignore[reportArgumentType]  # credential keys may be broader than str
 
-    # Cache token
-    ttl_seconds: Final = max(0, (expires_at - TOKEN_EXPIRY_BUFFER_MS - time.time() * 1000) / 1000)
-    if ttl_seconds > 0:
-        _token_cache.set_cache(cache_key, (token, expires_at), ttl=ttl_seconds)
+    if new_expires_at:
+        # Cache token
+        ttl_seconds: Final = max(0, (new_expires_at - TOKEN_EXPIRY_BUFFER_MS - time.time() * 1000) / 1000)
+        if ttl_seconds > 0:
+            _token_cache.set_cache(cache_key, (new_token, new_expires_at), ttl=ttl_seconds)
 
-    return token
+    return new_token
 
 
 async def get_access_token_async(
     credentials: str | None = None,
     scope: str | None = None,
     auth_url: str | None = None,
+    litellm_params: Mapping[str, object] | None = None,
 ) -> str:
     """Async version of get_access_token."""
-    credentials = credentials or _get_credentials()
-    if not credentials:
+    if not litellm_params:
+        litellm_params = {}  # mutable-ok: empty dict default; rebind-ok: provide default
+
+    access_token: Final = litellm_params.get("gigachat_access_token") or get_secret_str("GIGACHAT_ACCESS_TOKEN")
+    if access_token:
+        return access_token
+
+    effective_credentials: Final = credentials or _get_credentials()
+    if not effective_credentials:
         raise GigaChatAuthError(
             status_code=401,
             message="GigaChat credentials not provided. Set GIGACHAT_CREDENTIALS or GIGACHAT_API_KEY environment variable.",
         )
 
-    scope = scope or _get_scope()
-    auth_url = auth_url or _get_auth_url()
+    effective_scope: Final = scope or litellm_params.get("gigachat_scope") or _get_scope()
+    effective_auth_url: Final = auth_url or litellm_params.get("gigachat_auth_url") or _get_auth_url()
 
     # Check cache
-    cache_key: Final = f"gigachat_token:{credentials[:16]}"
+    cache_key: Final = f"gigachat_token:{effective_credentials[:16]}"
     cached: Final = _token_cache.get_cache(cache_key)
     if cached:
-        token, expires_at = cached
-        if time.time() * 1000 < expires_at - TOKEN_EXPIRY_BUFFER_MS:
+        _token, _expires_at = cached
+        if time.time() * 1000 < _expires_at - TOKEN_EXPIRY_BUFFER_MS:
             verbose_logger.debug("Using cached GigaChat access token")
-            return token
+            return _token
 
     # Request new token
-    token, expires_at = await _request_token_async(credentials, scope, auth_url)
+    new_token, new_expires_at = await _request_token_async(effective_credentials, effective_scope, effective_auth_url)  # pyright: ignore[reportArgumentType]  # credential keys may be broader than str
 
-    # Cache token
-    ttl_seconds: Final = max(0, (expires_at - TOKEN_EXPIRY_BUFFER_MS - time.time() * 1000) / 1000)
-    if ttl_seconds > 0:
-        _token_cache.set_cache(cache_key, (token, expires_at), ttl=ttl_seconds)
+    if new_expires_at:
+        # Cache token
+        ttl_seconds: Final = max(0, (new_expires_at - TOKEN_EXPIRY_BUFFER_MS - time.time() * 1000) / 1000)
+        if ttl_seconds > 0:
+            _token_cache.set_cache(cache_key, (new_token, new_expires_at), ttl=ttl_seconds)
 
-    return token
+    return new_token
 
 
 def _request_token_sync(
@@ -154,7 +173,7 @@ def _request_token_sync(
     Request new access token from GigaChat OAuth endpoint (sync).
 
     Returns:
-        Tuple of (access_token, expires_at_ms)
+        tuple of (access_token, expires_at_ms)
     """
     headers: Final = {
         "Authorization": f"Basic {credentials}",
@@ -169,7 +188,7 @@ def _request_token_sync(
         client: Final = _get_http_client()
         response: Final = client.post(auth_url, headers=headers, data=data, timeout=30)
         response.raise_for_status()
-        return _parse_token_response(response)
+        return _parse_token_response(response)  # pyright: ignore[reportArgumentType]  # httpx Response may be None at type level
     except httpx.HTTPStatusError as e:
         raise GigaChatAuthError(
             status_code=e.response.status_code,
@@ -204,7 +223,7 @@ async def _request_token_async(
         )
         response: Final = await client.post(auth_url, headers=headers, data=data, timeout=30)
         response.raise_for_status()
-        return _parse_token_response(response)
+        return _parse_token_response(response)  # pyright: ignore[reportArgumentType]  # httpx Response may be None at type level
     except httpx.HTTPStatusError as e:
         raise GigaChatAuthError(
             status_code=e.response.status_code,
@@ -223,7 +242,7 @@ def _parse_token_response(response: httpx.Response) -> tuple[str, int]:
 
     # GigaChat returns either 'tok'/'exp' or 'access_token'/'expires_at'
     access_token: Final = data.get("tok") or data.get("access_token")
-    expires_at = data.get("exp") or data.get("expires_at")
+    expires_at_raw: Final = data.get("exp") or data.get("expires_at")
 
     if not access_token:
         raise GigaChatAuthError(
@@ -232,8 +251,11 @@ def _parse_token_response(response: httpx.Response) -> tuple[str, int]:
         )
 
     # expires_at is in milliseconds
-    if isinstance(expires_at, str):
-        expires_at = int(expires_at)
+    expires_at: int  # rebind-ok: conditionally assigned from str or int
+    if isinstance(expires_at_raw, str):
+        expires_at = int(expires_at_raw)  # rebind-ok: conditionally assigned from str or int
+    else:
+        expires_at = expires_at_raw  # pyright: ignore[reportAssignmentType]  # raw value is int or str; converted above; rebind-ok: conditionally assigned from str or int
 
     verbose_logger.debug("GigaChat access token obtained successfully")
     return access_token, expires_at

litellm/llms/gigachat/chat/__init__.py

@@ -5,8 +5,8 @@ GigaChat Chat Module
 from .streaming import GigaChatModelResponseIterator
 from .transformation import GigaChatConfig, GigaChatError
 
-__all__ = [
+__all__ = (
     "GigaChatConfig",
     "GigaChatError",
     "GigaChatModelResponseIterator",
-]
+)

litellm/llms/gigachat/chat/streaming.py

@@ -4,13 +4,15 @@ GigaChat Streaming Response Handler
 
 import json
 import uuid
+from collections.abc import Mapping, Sequence
 from typing import Any, Final
 
+from litellm.llms.gigachat.utils import convert_usage
 from litellm.types.llms.openai import (
     ChatCompletionToolCallChunk,
     ChatCompletionToolCallFunctionChunk,
 )
-from litellm.types.utils import GenericStreamingChunk
+from litellm.types.utils import ChatCompletionUsageBlock, GenericStreamingChunk
 
 
 class GigaChatModelResponseIterator:
@@ -26,14 +28,9 @@ class GigaChatModelResponseIterator:
         self.response_iterator = self.streaming_response
         self.json_mode = json_mode
 
-    def chunk_parser(self, chunk: dict) -> GenericStreamingChunk:
+    def chunk_parser(self, chunk: Mapping[str, object]) -> GenericStreamingChunk:
         """Parse a single streaming chunk from GigaChat."""
-        text = ""
-        tool_use: ChatCompletionToolCallChunk | None = None
-        is_finished = False
-        finish_reason: str | None = None
-
-        choices: Final = chunk.get("choices", [])
+        choices: Sequence = chunk.get("choices") or ()  # mutable-ok: tuple literal as default
         if not choices:
             return GenericStreamingChunk(
                 text="",
@@ -45,40 +42,63 @@ class GigaChatModelResponseIterator:
             )
 
         choice: Final = choices[0]
-        delta: Final = choice.get("delta", {})
-        finish_reason = choice.get("finish_reason")
+        delta: Mapping[str, object] = choice.get("delta") or {}  # mutable-ok: empty dict default for get
+        chunk_finish_reason: Final = choice.get("finish_reason")
 
         # Extract text content
-        text = delta.get("content", "") or ""
+        text: Final = delta.get("content", "") or ""
+
+        usage_block: ChatCompletionUsageBlock | None = None  # rebind-ok: conditionally assigned after stop detection
+        tool_use: ChatCompletionToolCallChunk | None = None  # rebind-ok: conditionally assigned on function_call
+        finish_reason: str | None = chunk_finish_reason
 
         # Handle function_call in stream
-        if finish_reason == "function_call" and delta.get("function_call"):
-            func_call: Final = delta["function_call"]
-            args = func_call.get("arguments", {})
-
-            if isinstance(args, dict):
-                args = json.dumps(args, ensure_ascii=False)
+        raw_function_call: Final = delta.get("function_call")
+        if chunk_finish_reason == "function_call" and isinstance(raw_function_call, Mapping) and raw_function_call:
+            func_call: Final[Mapping[str, object]] = raw_function_call
+            args_raw: Final[object] = func_call.get("arguments") or {}
+            args_str: str  # rebind-ok: conditionally assigned from dict or str
+            if isinstance(args_raw, dict):
+                args_str = json.dumps(args_raw, ensure_ascii=False)  # rebind-ok: build from dict
+            else:
+                args_str = str(args_raw)
 
+            name_raw: Final = func_call.get("name")
             tool_use = ChatCompletionToolCallChunk(
                 id=f"call_{uuid.uuid4().hex[:24]}",
                 type="function",
                 function=ChatCompletionToolCallFunctionChunk(
-                    name=func_call.get("name", ""),
-                    arguments=args,
+                    name=name_raw if isinstance(name_raw, str) else "",
+                    arguments=args_str,
                 ),
                 index=0,
             )
             finish_reason = "tool_calls"
 
-        if finish_reason is not None:
-            is_finished = True
+        usage_data: Final = chunk.get("usage") or {}  # mutable-ok: empty dict default
+        if usage_data and isinstance(usage_data, dict):
+            validated_usage: Final = {k: int(v) for k, v in usage_data.items()}
+            usage = convert_usage(validated_usage)
+            _prompt_details: dict | None = (
+                usage.prompt_tokens_details.model_dump() if usage.prompt_tokens_details else None
+            )  # rebind-ok: conditional
+            _completion_details: dict | None = (
+                usage.completion_tokens_details.model_dump() if usage.completion_tokens_details else None
+            )  # rebind-ok: conditional
+            usage_block = ChatCompletionUsageBlock(  # pyright: ignore[reportCallIssue]  # TypedDict kwarg constructor
+                prompt_tokens=usage.prompt_tokens,
+                completion_tokens=usage.completion_tokens,
+                total_tokens=usage.total_tokens,
+                prompt_tokens_details=_prompt_details,
+                completion_tokens_details=_completion_details,
+            )
 
         return GenericStreamingChunk(
-            text=text,
+            text=str(text),
             tool_use=tool_use,
-            is_finished=is_finished,
+            is_finished=chunk_finish_reason is not None,
             finish_reason=finish_reason or "",
-            usage=None,
+            usage=usage_block,
             index=choice.get("index", 0),
         )
 

litellm/llms/gigachat/chat/transformation.py

@@ -4,19 +4,22 @@ GigaChat Chat Transformation
 Transforms OpenAI-format requests to GigaChat format and back.
 """
 
+from __future__ import annotations
+
 import json
 import time
 import uuid
-from collections.abc import AsyncIterator, Iterator
+from collections.abc import AsyncIterator, Iterator, Mapping, Sequence
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
 from litellm._logging import verbose_logger
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
+from litellm.llms.gigachat.utils import convert_usage, get_api_base
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
-from litellm.types.utils import Choices, Message, ModelResponse, Usage
+from litellm.types.utils import Choices, Message, ModelResponse
 
 from ..authenticator import get_access_token
 from ..file_handler import upload_file_sync
@@ -30,9 +33,6 @@ if TYPE_CHECKING:
 else:
     LiteLLMLoggingObj = Any
 
-# GigaChat API endpoint
-GIGACHAT_BASE_URL: Final = "https://gigachat.devices.sberbank.ru/api/v1"
-
 
 def is_valid_json(value: str) -> bool:
     """Checks whether the value passed is a valid serialized JSON string"""
@@ -90,30 +90,30 @@ class GigaChatConfig(BaseConfig):
         api_base: str | None,
         api_key: str | None,
         model: str,
-        optional_params: dict,
-        litellm_params: dict,
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
         stream: bool | None = None,
     ) -> str:
         """Get complete API URL for chat completions."""
-        base: Final = api_base or get_secret_str("GIGACHAT_API_BASE") or GIGACHAT_BASE_URL
+        base: Final = get_api_base(api_base)
         return f"{base}/chat/completions"
 
     def validate_environment(
         self,
-        headers: dict,
+        headers: dict,  # mutable-ok: mutates in place per GigaChat OAuth setup
         model: str,
-        messages: list[AllMessageValues],
-        optional_params: dict,
-        litellm_params: dict,
+        messages: Sequence[AllMessageValues],
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
         api_key: str | None = None,
         api_base: str | None = None,
-    ) -> dict:
+    ) -> dict:  # mutable-ok: base class contract returns dict for httpx
         """
         Set up headers with OAuth token.
         """
         # Get access token
         credentials: Final = api_key or get_secret_str("GIGACHAT_CREDENTIALS") or get_secret_str("GIGACHAT_API_KEY")
-        access_token: Final = get_access_token(credentials=credentials)
+        access_token: Final = get_access_token(credentials=credentials, litellm_params=litellm_params)
 
         # Store credentials for image uploads
         self._current_credentials = credentials
@@ -125,9 +125,9 @@ class GigaChatConfig(BaseConfig):
 
         return headers
 
-    def get_supported_openai_params(self, model: str) -> list[str]:
+    def get_supported_openai_params(self, model: str) -> list[str]:  # mutable-ok: base class contract returns list
         """Return list of supported OpenAI parameters."""
-        return [
+        return [  # mutable-ok: base class contract returns list
             "stream",
             "temperature",
             "top_p",
@@ -143,11 +143,11 @@ class GigaChatConfig(BaseConfig):
 
     def map_openai_params(
         self,
-        non_default_params: dict,
-        optional_params: dict,
+        non_default_params: Mapping[str, object],
+        optional_params: dict,  # mutable-ok: mutated in place per GigaChat mapping
         model: str,
         drop_params: bool,
-    ) -> dict:
+    ) -> dict:  # mutable-ok: base class contract returns dict
         """Map OpenAI parameters to GigaChat parameters."""
         for param, value in non_default_params.items():
             if param == "stream":
@@ -167,42 +167,50 @@ class GigaChatConfig(BaseConfig):
                 pass
             elif param == "tools":
                 # Convert tools to functions format
-                optional_params["functions"] = self._convert_tools_to_functions(value)
+                if isinstance(value, Sequence):
+                    optional_params["functions"] = self._convert_tools_to_functions(value)
             elif param == "tool_choice":
                 # Map OpenAI tool_choice to GigaChat function_call
-                mapped_choice = self._map_tool_choice(value)
-                if mapped_choice is not None:
-                    optional_params["function_call"] = mapped_choice
+                if isinstance(value, (str, Mapping)):
+                    mapped_choice = self._map_tool_choice(value)
+                    if mapped_choice is not None:
+                        optional_params["function_call"] = mapped_choice
             elif param == "functions":
                 optional_params["functions"] = value
             elif param == "function_call":
                 optional_params["function_call"] = value
             elif param == "response_format":
                 # Handle structured output via function calling
-                if value.get("type") == "json_schema":
+                if isinstance(value, Mapping) and value.get("type") == "json_schema":
                     json_schema = value.get("json_schema", {})
                     schema_name = json_schema.get("name", "structured_output")
                     schema = json_schema.get("schema", {})
 
-                    function_def = {
+                    function_def = {  # mutable-ok: request payload for httpx
                         "name": schema_name,
                         "description": f"Output structured response: {schema_name}",
                         "parameters": schema,
                     }
 
-                    if "functions" not in optional_params:
-                        optional_params["functions"] = []
-                    optional_params["functions"].append(function_def)
-                    optional_params["function_call"] = {"name": schema_name}
+                    existing_functions = optional_params.get("functions")
+                    optional_params["functions"] = [
+                        *(
+                            existing_functions
+                            if isinstance(existing_functions, Sequence) and not isinstance(existing_functions, str)
+                            else ()
+                        ),
+                        function_def,
+                    ]
+                    optional_params["function_call"] = {"name": schema_name}  # mutable-ok: request payload
                     optional_params["_structured_output"] = True
 
         return optional_params
 
-    def _convert_tools_to_functions(self, tools: list[dict]) -> list[dict]:
+    def _convert_tools_to_functions(self, tools: Sequence) -> Sequence[dict]:
         """Convert OpenAI tools format to GigaChat functions format."""
-        functions: Final = []
+        functions: Final[list[dict]] = []  # mutable-ok: accumulator for building functions list
         for tool in tools:
-            if tool.get("type") == "function":
+            if isinstance(tool, dict) and tool.get("type") == "function":
                 func = tool.get("function", {})
                 functions.append(
                     {
@@ -213,7 +221,7 @@ class GigaChatConfig(BaseConfig):
                 )
         return functions
 
-    def _map_tool_choice(self, tool_choice: str | dict) -> str | dict | None:
+    def _map_tool_choice(self, tool_choice: str | Mapping[str, object]) -> str | Mapping[str, object] | None:
         """
         Map OpenAI tool_choice to GigaChat function_call format.
 
@@ -246,8 +254,9 @@ class GigaChatConfig(BaseConfig):
             # OpenAI format: {"type": "function", "function": {"name": "func_name"}}
             # GigaChat format: {"name": "func_name"}
             if tool_choice.get("type") == "function":
-                func_name: Final = tool_choice.get("function", {}).get("name")
-                if func_name:
+                function_spec: Final = tool_choice.get("function")
+                func_name: Final = function_spec.get("name") if isinstance(function_spec, Mapping) else None
+                if isinstance(func_name, str) and func_name:
                     return {"name": func_name}
 
         # Default to None (don't set function_call)
@@ -273,20 +282,51 @@ class GigaChatConfig(BaseConfig):
             verbose_logger.error("Failed to upload image: %s", e)
             return None
 
+    def _transform_list_content(self, content: Sequence) -> tuple[str, Sequence[str]]:
+        """
+        Extract text and image attachments from a multimodal message content list.
+
+        Args:
+            content: List of content parts (OpenAI multimodal format)
+
+        Returns:
+            Tuple of (combined text, list of attachment file ids)
+        """
+        texts: Final[list[str]] = []  # mutable-ok: accumulator
+        attachments: Final[list[str]] = []  # mutable-ok: accumulator
+        for part in content:
+            if isinstance(part, dict):
+                if part.get("type") == "text":
+                    texts.append(part.get("text", ""))
+                elif part.get("type") == "image_url":
+                    # Extract image URL and upload to GigaChat
+                    image_url: object = part.get("image_url", {})
+                    upload_url: str
+                    if isinstance(image_url, str):
+                        upload_url = image_url
+                    else:
+                        upload_url = str(image_url.get("url", "")) if isinstance(image_url, dict) else ""
+                    if upload_url:
+                        file_id = self._upload_image(upload_url)
+                        if file_id:
+                            attachments.append(file_id)
+        text: Final = "\n".join(texts) if texts else ""
+        return text, attachments
+
     def transform_request(
         self,
         model: str,
-        messages: list[AllMessageValues],
-        optional_params: dict,
-        litellm_params: dict,
-        headers: dict,
-    ) -> dict:
+        messages: Sequence[AllMessageValues],
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        headers: Mapping[str, object],
+    ) -> dict:  # mutable-ok: request payload sent to httpx
         """Transform OpenAI request to GigaChat format."""
         # Transform messages
         giga_messages: Final = self._transform_messages(messages)
 
         # Build request
-        request_data: Final = {
+        request_data: Final[dict[str, object]] = {
             "model": model.replace("gigachat/", ""),
             "messages": giga_messages,
         }
@@ -311,9 +351,9 @@ class GigaChatConfig(BaseConfig):
 
         return request_data
 
-    def _transform_messages(self, messages: list[AllMessageValues]) -> list[dict]:
+    def _transform_messages(self, messages: Sequence[AllMessageValues]) -> Sequence[dict]:
         """Transform OpenAI messages to GigaChat format."""
-        transformed: Final = []
+        transformed: Final[list[dict]] = []  # mutable-ok: accumulator for building transformed messages
 
         for i, msg in enumerate(messages):
             message = dict(msg)
@@ -341,24 +381,7 @@ class GigaChatConfig(BaseConfig):
             # Handle list content (multimodal) - extract text and images
             content = message.get("content")
             if isinstance(content, list):
-                texts = []
-                attachments = []
-                for part in content:
-                    if isinstance(part, dict):
-                        if part.get("type") == "text":
-                            texts.append(part.get("text", ""))
-                        elif part.get("type") == "image_url":
-                            # Extract image URL and upload to GigaChat
-                            image_url = part.get("image_url", {})
-                            if isinstance(image_url, str):
-                                url = image_url
-                            else:
-                                url = image_url.get("url", "")
-                            if url:
-                                file_id = self._upload_image(url)
-                                if file_id:
-                                    attachments.append(file_id)
-                message["content"] = "\n".join(texts) if texts else ""
+                message["content"], attachments = self._transform_list_content(content)
                 if attachments:
                     message["attachments"] = attachments
 
@@ -393,7 +416,7 @@ class GigaChatConfig(BaseConfig):
         messages: list[AllMessageValues],
         optional_params: dict,
         litellm_params: dict,
-        encoding: "tiktoken.Encoding | None",
+        encoding: tiktoken.Encoding | None,
         api_key: str | None = None,
         json_mode: bool | None = None,
     ) -> ModelResponse:
@@ -408,7 +431,7 @@ class GigaChatConfig(BaseConfig):
 
         is_structured_output: Final = optional_params.get("_structured_output", False)
 
-        choices: Final = []
+        choices: Final[list[Choices]] = []  # mutable-ok: accumulator for building response choices
         for choice in response_json.get("choices", []):
             message_data = choice.get("message", {})
             finish_reason = choice.get("finish_reason", "stop")
@@ -462,11 +485,7 @@ class GigaChatConfig(BaseConfig):
 
         # Build usage
         usage_data: Final = response_json.get("usage", {})
-        usage: Final = Usage(
-            prompt_tokens=usage_data.get("prompt_tokens", 0),
-            completion_tokens=usage_data.get("completion_tokens", 0),
-            total_tokens=usage_data.get("total_tokens", 0),
-        )
+        usage: Final = convert_usage(usage_data)
 
         model_response.id = response_json.get("id", f"chatcmpl-{uuid.uuid4().hex[:12]}")
         model_response.created = response_json.get("created", int(time.time()))

litellm/llms/gigachat/embedding/transformation.py

@@ -5,6 +5,8 @@ Transforms OpenAI /v1/embeddings format to GigaChat format.
 API Documentation: https://developers.sber.ru/docs/ru/gigachat/api/reference/rest/post-embeddings
 """
 
+from __future__ import annotations
+
 import types
 from typing import Final
 
@@ -14,14 +16,12 @@ from litellm import LlmProviders
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.embedding.transformation import BaseEmbeddingConfig
+from litellm.llms.gigachat.utils import get_api_base
 from litellm.types.llms.openai import AllEmbeddingInputValues, AllMessageValues
 from litellm.types.utils import EmbeddingResponse
 
 from ..authenticator import get_access_token
 
-# GigaChat API endpoint
-GIGACHAT_BASE_URL: Final = "https://gigachat.devices.sberbank.ru/api/v1"
-
 
 class GigaChatEmbeddingError(BaseLLMException):
     """GigaChat Embedding API error."""
@@ -78,9 +78,9 @@ class GigaChatEmbeddingConfig(BaseEmbeddingConfig):
         Returns provider info for GigaChat.
 
         Returns:
-            Tuple of (custom_llm_provider, api_base, dynamic_api_key)
+            tuple of (custom_llm_provider, api_base, dynamic_api_key)
         """
-        api_base = api_base or GIGACHAT_BASE_URL
+        api_base = get_api_base(api_base)
         return LlmProviders.GIGACHAT.value, api_base, api_key
 
     def get_complete_url(
@@ -93,7 +93,7 @@ class GigaChatEmbeddingConfig(BaseEmbeddingConfig):
         stream: bool | None = None,
     ) -> str:
         """Get the complete URL for embeddings endpoint."""
-        base: Final = api_base or GIGACHAT_BASE_URL
+        base: Final = get_api_base(api_base)
         return f"{base}/embeddings"
 
     def transform_embedding_request(
@@ -114,14 +114,12 @@ class GigaChatEmbeddingConfig(BaseEmbeddingConfig):
         """
         # Normalize input to list
         if isinstance(input, str):
-            input_list: list = [input]
-        elif isinstance(input, list):
-            input_list = input
+            input_list: list = [input]  # rebind-ok: locally scoped conversion
         else:
-            input_list = [input]
+            input_list = input
 
         # Remove gigachat/ prefix from model if present
-        model = model.removeprefix("gigachat/")
+        model = model.removeprefix("gigachat/")  # rebind-ok: parameter reassignment for normalization
 
         return {
             "model": model,
@@ -191,7 +189,7 @@ class GigaChatEmbeddingConfig(BaseEmbeddingConfig):
         Set up headers with OAuth token for GigaChat.
         """
         # Get access token via OAuth
-        access_token: Final = get_access_token(api_key)
+        access_token: Final = get_access_token(credentials=api_key, litellm_params=litellm_params)
 
         default_headers: Final = {
             "Content-Type": "application/json",

litellm/llms/gigachat/file_handler.py

@@ -9,6 +9,7 @@ import base64
 import hashlib
 import re
 import uuid
+from collections.abc import Mapping
 from typing import Final
 
 from litellm._logging import verbose_logger
@@ -16,13 +17,11 @@ from litellm.llms.custom_httpx.http_handler import (
     _get_httpx_client,
     get_async_httpx_client,
 )
+from litellm.llms.gigachat.utils import get_api_base
 from litellm.types.utils import LlmProviders
 
 from .authenticator import get_access_token, get_access_token_async
 
-# GigaChat API endpoint
-GIGACHAT_BASE_URL: Final = "https://gigachat.devices.sberbank.ru/api/v1"
-
 # Simple in-memory cache for file IDs
 _file_cache: Final[dict[str, str]] = {}
 
@@ -82,6 +81,7 @@ def upload_file_sync(
     image_url: str,
     credentials: str | None = None,
     api_base: str | None = None,
+    litellm_params: Mapping[str, object] | None = None,
 ) -> str | None:
     """
     Upload file to GigaChat and return file_id (sync).
@@ -114,10 +114,10 @@ def upload_file_sync(
         filename: Final = f"{uuid.uuid4()}.{ext}"
 
         # Get access token
-        access_token: Final = get_access_token(credentials)
+        access_token: Final = get_access_token(credentials=credentials, litellm_params=litellm_params)
 
         # Upload to GigaChat
-        base_url: Final = api_base or GIGACHAT_BASE_URL
+        base_url: Final = get_api_base(api_base)
         upload_url: Final = f"{base_url}/files"
 
         client: Final = _get_httpx_client(params={"ssl_verify": False})
@@ -147,6 +147,7 @@ async def upload_file_async(
     image_url: str,
     credentials: str | None = None,
     api_base: str | None = None,
+    litellm_params: Mapping[str, object] | None = None,
 ) -> str | None:
     """
     Upload file to GigaChat and return file_id (async).
@@ -179,10 +180,10 @@ async def upload_file_async(
         filename: Final = f"{uuid.uuid4()}.{ext}"
 
         # Get access token
-        access_token: Final = await get_access_token_async(credentials)
+        access_token: Final = await get_access_token_async(credentials=credentials, litellm_params=litellm_params)
 
         # Upload to GigaChat
-        base_url: Final = api_base or GIGACHAT_BASE_URL
+        base_url: Final = get_api_base(api_base)
         upload_url: Final = f"{base_url}/files"
 
         client: Final = get_async_httpx_client(

litellm/llms/gigachat/passthrough/__init__.py

@@ -0,0 +1,7 @@
+"""
+GigaChat passthrough Module
+"""
+
+from .transformation import GigaChatPassthroughConfig
+
+__all__ = ("GigaChatPassthroughConfig",)

litellm/llms/gigachat/passthrough/transformation.py

@@ -0,0 +1,213 @@
+from __future__ import annotations
+
+import json
+from collections.abc import Mapping, Sequence
+from typing import TYPE_CHECKING, Final
+
+import httpx
+
+from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
+from litellm.llms.gigachat.authenticator import get_access_token
+from litellm.llms.gigachat.chat.streaming import GigaChatModelResponseIterator
+from litellm.llms.gigachat.utils import GIGACHAT_BASE_URL
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.llms.openai import AllMessageValues
+from litellm.types.utils import EmbeddingResponse
+
+if TYPE_CHECKING:
+    from httpx import URL, Response
+
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.types.utils import CostResponseTypes
+
+
+class GigaChatPassthroughConfig(BasePassthroughConfig):
+    def is_streaming_request(self, endpoint: str, request_data: Mapping[str, object]) -> bool:
+        return request_data.get("stream", False)
+
+    def get_complete_url(
+        self,
+        api_base: str | None,
+        api_key: str | None,
+        model: str,
+        endpoint: str,
+        request_query_params: Mapping[str, object] | None,
+        litellm_params: Mapping[str, object],
+    ) -> tuple[URL, str]:
+        """Get complete API URL for chat completions."""
+        base_target_url: Final = self.get_api_base(api_base)
+
+        if base_target_url is None:
+            raise Exception("GigaChat api base not found")
+
+        complete_url: Final = f"{base_target_url}/{endpoint.lstrip('/')}"
+
+        return (
+            httpx.URL(complete_url),
+            base_target_url,
+        )
+
+    def validate_environment(
+        self,
+        headers: dict,  # mutable-ok: mutates in place to set OAuth headers
+        model: str,
+        messages: Sequence[AllMessageValues],
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        api_key: str | None = None,
+        api_base: str | None = None,
+    ) -> dict:  # mutable-ok: base class contract returns dict for httpx
+        """
+        Set up headers with OAuth token.
+        """
+        # Get access token
+        access_token: Final = get_access_token(credentials=api_key, litellm_params=litellm_params)
+
+        headers["Authorization"] = f"Bearer {access_token}"  # rebind-ok: mutating for OAuth setup
+        headers["Content-Type"] = "application/json"  # rebind-ok: mutating for OAuth setup
+        headers["Accept"] = "application/json"  # rebind-ok: mutating for OAuth setup
+
+        return headers
+
+    def logging_non_streaming_response(
+        self,
+        model: str,
+        custom_llm_provider: str,
+        httpx_response: Response,
+        request_data: Mapping[str, object],
+        logging_obj: LiteLLMLoggingObj,
+        endpoint: str,
+    ) -> CostResponseTypes | None:
+        from litellm import encoding
+        from litellm.types.utils import LlmProviders, ModelResponse
+        from litellm.utils import ProviderConfigManager
+
+        # cost tracking only for completions and embeddings
+        if "completions" in endpoint:
+            provider_chat_config: Final = ProviderConfigManager.get_provider_chat_config(
+                provider=LlmProviders(custom_llm_provider),
+                model=model,
+            )
+
+            if provider_chat_config is None:
+                raise ValueError(f"No provider config found for model: {model}")
+
+            raw_messages: Final = request_data.get("messages")
+            litellm_model_response: Final = provider_chat_config.transform_response(
+                model=model,
+                messages=list(raw_messages)
+                if isinstance(raw_messages, list)
+                else [],  # mutable-ok: transform_response wants a list
+                raw_response=httpx_response,
+                model_response=ModelResponse(),
+                logging_obj=logging_obj,
+                optional_params={},  # mutable-ok: empty dict kwarg for transform_response
+                litellm_params={},  # mutable-ok: empty dict kwarg for transform_response
+                api_key="",
+                request_data=dict(request_data),  # mutable-ok: transform_response wants a dict
+                encoding=encoding,
+            )
+
+            return litellm_model_response
+
+        if "embeddings" in endpoint:
+            provider_embedding_config: Final = ProviderConfigManager.get_provider_embedding_config(
+                provider=LlmProviders(custom_llm_provider),
+                model=model,
+            )
+
+            if provider_embedding_config is None:
+                raise ValueError(f"No provider config found for model: {model}")
+
+            litellm_embedding_response: Final[EmbeddingResponse] = (
+                provider_embedding_config.transform_embedding_response(
+                    model=model,
+                    raw_response=httpx_response,
+                    model_response=EmbeddingResponse(),
+                    logging_obj=logging_obj,
+                    optional_params={},  # mutable-ok: empty dict kwarg for transform_embedding_response
+                    api_key="",
+                    request_data=dict(request_data),  # mutable-ok: transform_embedding_response wants a dict
+                    litellm_params={},  # mutable-ok: empty dict kwarg for transform_embedding_response
+                )
+            )
+
+            return litellm_embedding_response
+
+        return None
+
+    def handle_logging_collected_chunks(
+        self,
+        all_chunks: Sequence[str],
+        litellm_logging_obj: LiteLLMLoggingObj,
+        model: str,
+        custom_llm_provider: str,
+        endpoint: str,
+    ) -> CostResponseTypes | None:
+        """
+        1. Convert all_chunks to a ModelResponseStream
+        2. combine model_response_stream to model_response
+        3. Return the model_response
+        """
+
+        from litellm.litellm_core_utils.streaming_handler import (
+            convert_generic_chunk_to_model_response_stream,
+            generic_chunk_has_all_required_fields,
+        )
+        from litellm.main import stream_chunk_builder
+        from litellm.types.utils import ModelResponseStream
+
+        all_translated_chunks: Final[list[object]] = []  # mutable-ok: accumulator
+
+        for chunk in all_chunks:
+            chunk = chunk.strip()
+            if not chunk or chunk == "[DONE]":
+                continue
+            chunk = chunk.removeprefix("data: ")
+            try:
+                message = json.loads(chunk)
+            except json.JSONDecodeError:
+                continue
+
+            gigachat_iterator = GigaChatModelResponseIterator(
+                streaming_response=None,
+                sync_stream=False,
+            )
+            translated_chunk = gigachat_iterator.chunk_parser(chunk=message)
+
+            if isinstance(translated_chunk, dict) and generic_chunk_has_all_required_fields(  # pyright: ignore[reportUnnecessaryIsInstance]  # runtime guard for patched chunk_parser
+                dict(translated_chunk)
+            ):
+                chunk_obj = convert_generic_chunk_to_model_response_stream(
+                    translated_chunk  # pyright: ignore[reportArgumentType]  # validated TypedDict
+                )
+            elif isinstance(translated_chunk, ModelResponseStream):
+                chunk_obj = translated_chunk
+            else:
+                continue
+
+            all_translated_chunks.append(chunk_obj)
+
+        if len(all_translated_chunks) > 0:
+            return stream_chunk_builder(
+                chunks=all_translated_chunks,
+                logging_obj=litellm_logging_obj,
+            )
+        return None
+
+    @staticmethod
+    def get_api_base(api_base: str | None = None) -> str | None:
+        return api_base or get_secret_str("GIGACHAT_API_BASE") or GIGACHAT_BASE_URL
+
+    @staticmethod
+    def get_api_key(
+        api_key: str | None = None,
+    ) -> str | None:
+        return api_key or get_secret_str("GIGACHAT_API_KEY")
+
+    @staticmethod
+    def get_base_model(model: str) -> str | None:
+        return model
+
+    def get_models(self, api_key: str | None = None, api_base: str | None = None) -> list[str]:
+        return list(super().get_models(api_key, api_base))

litellm/llms/gigachat/utils.py

@@ -0,0 +1,26 @@
+from collections.abc import Mapping
+from typing import Final
+
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.utils import PromptTokensDetailsWrapper, Usage
+
+# GigaChat API endpoint
+GIGACHAT_BASE_URL: Final = "https://gigachat.devices.sberbank.ru/api/v1"
+
+
+def convert_usage(usage_data: Mapping[str, int]) -> Usage:
+    precached_prompt_tokens: Final = usage_data.get("precached_prompt_tokens", 0)
+    prompt_tokens_details: Final = (
+        PromptTokensDetailsWrapper(cached_tokens=precached_prompt_tokens) if precached_prompt_tokens > 0 else None
+    )
+
+    return Usage(
+        prompt_tokens=usage_data.get("prompt_tokens", 0) + precached_prompt_tokens,
+        completion_tokens=usage_data.get("completion_tokens", 0),
+        prompt_tokens_details=prompt_tokens_details,
+        total_tokens=usage_data.get("total_tokens", 0) + precached_prompt_tokens,
+    )
+
+
+def get_api_base(api_base: str | None = None) -> str | None:
+    return api_base or get_secret_str("GIGACHAT_API_BASE") or GIGACHAT_BASE_URL

litellm/llms/oci/chat/cohere.py

@@ -9,7 +9,7 @@ response parsing, and streaming chunk parsing for models served with
 import datetime
 import json
 from collections.abc import Iterable, Mapping, Sequence
-from typing import Any, Final
+from typing import Final
 
 import httpx
 from pydantic import JsonValue, TypeAdapter, ValidationError
@@ -76,7 +76,7 @@ def _content_text(content: str | Iterable[Mapping[str, object]] | None) -> str:
     return str(content)
 
 
-def _extract_text_content(content: Any) -> str:
+def _extract_text_content(content: str | Iterable[Mapping[str, object]] | None) -> str:
     """Return the plain-text representation of a message content value."""
     return _content_text(content)
 

litellm/llms/runwayml/videos/transformation.py

@@ -160,14 +160,12 @@ class RunwayMLVideoConfig(BaseVideoConfig):
             **self._prompt_image_param(video_create_optional_params),
             **self._ratio_param(video_create_optional_params),
             **self._duration_param(video_create_optional_params),
-            # Pass through other parameters that aren't OpenAI-specific
             **{key: value for key, value in video_create_optional_params.items() if key not in supported_openai_params},
         }
 
     @staticmethod
     def _prompt_image_param(video_create_optional_params: VideoCreateOptionalRequestParams) -> Mapping[str, object]:
         # Handle input_reference parameter - map to promptImage
-        # RunwayML supports URLs and data URIs directly
         if "input_reference" in video_create_optional_params:
             return {"promptImage": video_create_optional_params["input_reference"]}
         return {}

litellm/llms/vertex_ai/image_edit/vertex_gemini_transformation.py

@@ -182,7 +182,6 @@ class VertexAIGeminiImageEditConfig(BaseImageEditConfig, VertexLLM):
             else None
         )
 
-        # Generation config with proper structure for image editing
         generation_config: Final[dict[str, object]] = {
             key: value for key, value in (("response_modalities", ["IMAGE"]), ("image_config", image_config)) if value
         }

litellm/llms/vertex_ai/vector_stores/rag_api/transformation.py

@@ -203,7 +203,6 @@ class VertexVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
             if value is not None
         }
 
-        # Build the request body for Vertex AI RAG API
         query_body: Final[Mapping[str, object]] = {
             key: value
             for key, value in (("text", query), ("rag_retrieval_config", rag_retrieval_config or None))
@@ -294,7 +293,6 @@ class VertexVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
         # Add metadata if provided
         metadata: Final = vector_store_create_optional_params.get("metadata")
 
-        # Build the request body for Vertex AI RAG Corpus creation
         request_body: Final[dict[str, object]] = {
             key: value
             for key, value in (

litellm/main.py

@@ -5507,6 +5507,9 @@ def completion(
             tpm=kwargs.get("tpm"),
             rpm=kwargs.get("rpm"),
             use_xai_oauth=kwargs.get("use_xai_oauth", False),
+            gigachat_scope=kwargs.get("gigachat_scope"),
+            gigachat_auth_url=kwargs.get("gigachat_auth_url"),
+            gigachat_access_token=kwargs.get("gigachat_access_token"),
             **{key: kwargs[key] for key in FORWARDED_KWARGS_KEYS if key in kwargs},
         )
         cast(LiteLLMLoggingObj, logging).update_environment_variables(

litellm/model_prices_and_context_window_backup.json

@@ -553,6 +553,26 @@
         "supports_response_schema": true,
         "supports_vision": true
     },
+    "amazon.nova-sonic-v1:0": {
+        "input_cost_per_audio_token": 3.4e-06,
+        "input_cost_per_token": 6e-08,
+        "litellm_provider": "bedrock",
+        "mode": "realtime",
+        "output_cost_per_audio_token": 1.36e-05,
+        "output_cost_per_token": 2.4e-07,
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
+    "amazon.nova-2-sonic-v1:0": {
+        "input_cost_per_audio_token": 3e-06,
+        "input_cost_per_token": 3.3e-07,
+        "litellm_provider": "bedrock",
+        "mode": "realtime",
+        "output_cost_per_audio_token": 1.2e-05,
+        "output_cost_per_token": 2.75e-06,
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
     "amazon.rerank-v1:0": {
         "input_cost_per_query": 0.001,
         "input_cost_per_token": 0.0,
@@ -19566,6 +19586,61 @@
         "supports_system_messages": true,
         "supports_tool_choice": true
     },
+    "friendliai/zai-org/GLM-5.3-Flash": {
+        "litellm_provider": "friendliai",
+        "supports_reasoning": true,
+        "supports_function_calling": true,
+        "max_input_tokens": 1048576,
+        "max_tokens": 1048576,
+        "max_output_tokens": 1048576,
+        "input_cost_per_token": 1.5e-07,
+        "output_cost_per_token": 5e-07,
+        "cache_read_input_token_cost": 3e-08,
+        "supports_prompt_caching": true,
+        "reasoning_effort_levels": [
+            "low",
+            "high",
+            "max"
+        ],
+        "supports_parallel_function_calling": true,
+        "supports_response_schema": true,
+        "supports_native_structured_output": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "mode": "chat",
+        "comment": "Native multimodal GLM model for efficient coding and long-horizon agent tasks",
+        "source": "https://api.friendli.ai/serverless/v1/models",
+        "supports_vision": true,
+        "supports_image_input": true,
+        "supports_video_input": true
+    },
+    "friendliai/zai-org/GLM-5.3": {
+        "litellm_provider": "friendliai",
+        "supports_reasoning": true,
+        "supports_function_calling": true,
+        "max_input_tokens": 1048576,
+        "max_tokens": 1048576,
+        "max_output_tokens": 1048576,
+        "input_cost_per_token": 1.26e-06,
+        "output_cost_per_token": 3.96e-06,
+        "cache_read_input_token_cost": 2.34e-07,
+        "supports_prompt_caching": true,
+        "reasoning_effort_levels": [
+            "low",
+            "high",
+            "max"
+        ],
+        "supports_parallel_function_calling": true,
+        "supports_response_schema": true,
+        "supports_native_structured_output": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "mode": "chat",
+        "comment": "Flagship GLM model for long-horizon coding, agents, and complex project delivery",
+        "source": "https://api.friendli.ai/serverless/v1/models",
+        "supports_vision": false,
+        "supports_image_input": false
+    },
     "ft:babbage-002": {
         "deprecation_date": "2026-10-23",
         "input_cost_per_token": 1.6e-06,
@@ -24344,7 +24419,7 @@
         "supports_response_schema": true,
         "supports_vision": true
     },
-    "gigachat/GigaChat-2-Lite": {
+    "gigachat/GigaChat-2": {
         "input_cost_per_token": 0.0,
         "litellm_provider": "gigachat",
         "max_input_tokens": 128000,
@@ -24406,6 +24481,15 @@
         "output_cost_per_token": 0.0,
         "output_vector_size": 2560
     },
+    "gigachat/GigaEmbeddings-3B-2025-09": {
+        "input_cost_per_token": 0.0,
+        "litellm_provider": "gigachat",
+        "max_input_tokens": 4096,
+        "max_tokens": 4096,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 2048
+    },
     "gmi/anthropic/claude-opus-4.5": {
         "input_cost_per_token": 5e-06,
         "litellm_provider": "gmi",

litellm/models/model.py

@@ -29,6 +29,8 @@ class LiteLLM_ProxyModelTable(LiteLLMPydanticObjectBase):
     @model_validator(mode="before")
     @classmethod
     def check_potential_json_str(cls, values):
+        if not isinstance(values, dict):
+            return values
         if isinstance(values.get("litellm_params"), str):
             try:
                 values["litellm_params"] = json.loads(values["litellm_params"])

litellm/passthrough/main.py

@@ -2,17 +2,22 @@
 This module is used to pass through requests to the LLM APIs.
 """
 
+from __future__ import annotations
+
 import asyncio
 import contextvars
-from collections.abc import AsyncGenerator, Coroutine, Generator
+from collections.abc import AsyncGenerator, AsyncIterator, Coroutine, Generator, Iterator
 from functools import partial
-from typing import TYPE_CHECKING, Any, Final, Optional, cast
+from types import TracebackType
+from typing import Any, Final, cast
 
 import httpx
-from httpx._types import CookieTypes, QueryParamTypes, RequestFiles
+from httpx._types import CookieTypes, QueryParamTypes, RequestContent, RequestFiles
 
 from litellm._logging import verbose_logger
 from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
+from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
 from litellm.passthrough.utils import CommonUtils
@@ -21,9 +26,222 @@ from litellm.utils import client
 base_llm_http_handler = BaseLLMHTTPHandler()
 from .utils import BasePassthroughUtils
 
-if TYPE_CHECKING:
-    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
-    from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
+
+async def _as_async_generator(iterable: AsyncIterator[bytes]) -> AsyncGenerator[bytes, Any]:
+    async for chunk in iterable:
+        yield chunk
+
+
+def _as_generator(iterable: Iterator[bytes]) -> Generator[bytes, Any, Any]:
+    yield from iterable
+
+
+class AsyncPassthroughStreamingResponse(AsyncGenerator[Any, Any]):
+    def __init__(
+        self,
+        response: Coroutine[Any, Any, httpx.Response],
+        litellm_logging_obj: LiteLLMLoggingObj,
+        provider_config: BasePassthroughConfig,
+    ) -> None:
+        self._initialized = False
+        self._status_code: int = 0
+        self._headers = httpx.Headers()
+        self._response_coro = response
+        self._response: httpx.Response
+        self._iterator: AsyncGenerator[bytes, Any]
+        self._litellm_logging_obj = litellm_logging_obj
+        self._provider_config = provider_config
+        self._raw_bytes: list[bytes] = []  # mutable-ok: instance buffer for streaming chunks
+        self._flush_scheduled = False
+        self._background_tasks: set[asyncio.Task] = set()  # mutable-ok: instance set for background task tracking
+        self._hidden_params: dict[str, object] = {}  # mutable-ok: router attaches response headers here in place
+
+    @property
+    def status_code(self) -> int:
+        if not self._initialized:
+            raise RuntimeError("AsyncPassthroughStreamingResponse must be awaited before accessing status_code")
+        return self._status_code
+
+    @status_code.setter
+    def status_code(self, value: int) -> None:
+        self._status_code = value
+
+    @property
+    def headers(self) -> httpx.Headers:
+        if not self._initialized:
+            raise RuntimeError("AsyncPassthroughStreamingResponse must be awaited before accessing headers")
+        return self._headers
+
+    @headers.setter
+    def headers(self, value: httpx.Headers) -> None:
+        self._headers = value
+
+    def __await__(self) -> Iterator[Any]:
+        async def _init():
+            if not self._initialized:
+                self._response = await self._response_coro
+                self.headers = self._response.headers
+                self.status_code = self._response.status_code
+                self._initialized = True
+                try:
+                    self._response.raise_for_status()
+                    self._iterator = _as_async_generator(self._response.aiter_bytes())
+                except Exception:  # noqa: BLE001 # Safe catch-all for cleanup logic
+                    try:
+                        await self._response.aread()
+                    except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                        pass
+                    try:
+                        await self._response.aclose()
+                    except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                        pass
+                    raise
+            return self
+
+        return _init().__await__()
+
+    def _start_flush(self) -> None:
+        if self._flush_scheduled or not self._raw_bytes:
+            return
+        self._flush_scheduled = True
+
+        try:
+            task: Final = asyncio.create_task(
+                self._litellm_logging_obj.async_flush_passthrough_collected_chunks(
+                    raw_bytes=self._raw_bytes,
+                    provider_config=self._provider_config,
+                )
+            )
+
+            # Compliant: Save a strong reference to prevent GC
+            self._background_tasks.add(task)
+
+            # Remove the task from the set when it finishes to avoid memory leaks
+            task.add_done_callback(self._background_tasks.discard)
+        except Exception as e:  # noqa: BLE001 # Safe catch-all for verbose logging
+            verbose_logger.exception(
+                "Failed to schedule passthrough spend-tracking flush; %d buffered chunks dropped: %s",
+                len(self._raw_bytes),
+                e,
+            )
+
+    def __aiter__(self) -> AsyncPassthroughStreamingResponse:
+        return self
+
+    def aiter_bytes(self) -> AsyncPassthroughStreamingResponse:
+        return self
+
+    async def __anext__(self) -> bytes:
+        if not self._initialized:
+            await self  # pyright: ignore[reportGeneralTypeIssues]  # structural type check misses __await__
+        try:
+            chunk: Final = await anext(self._iterator)
+            self._raw_bytes.append(chunk)
+        except Exception:  # noqa: BLE001 # Safe catch-all for cleanup logic
+            self._start_flush()
+            try:
+                await self._response.aclose()
+            except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                pass
+            raise
+        else:
+            return chunk
+
+    async def asend(self, value: bytes) -> bytes:
+        if not self._initialized:
+            await self  # pyright: ignore[reportGeneralTypeIssues]  # structural type check misses __await__
+        return await self._iterator.asend(value)
+
+    async def athrow(
+        self,
+        typ: BaseException | type[BaseException],
+        val: BaseException | object = None,
+        tb: TracebackType | None = None,
+    ) -> bytes:
+        if not self._initialized:
+            await self  # pyright: ignore[reportGeneralTypeIssues]  # structural type check misses __await__
+        return await self._iterator.athrow(typ, val, tb)  # pyright: ignore[reportCallIssue, reportArgumentType]  # matches one of the athrow overloads
+
+    async def aclose(self) -> None:
+        self._start_flush()
+        try:
+            if self._initialized:
+                await self._iterator.aclose()
+                await self._response.aclose()
+        except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+            pass
+
+
+class PassthroughStreamingResponse(Generator[Any, Any, Any]):
+    def __init__(
+        self,
+        response: httpx.Response,
+        litellm_logging_obj: LiteLLMLoggingObj,
+        provider_config: BasePassthroughConfig,
+    ) -> None:
+        self._response = response
+        self.headers = response.headers
+        self.status_code = response.status_code
+        self._litellm_logging_obj = litellm_logging_obj
+        self._provider_config = provider_config
+        self._iterator: Generator[bytes, Any, Any] = _as_generator(response.iter_bytes())
+        self._raw_bytes: list[bytes] = []  # mutable-ok: instance buffer for streaming chunks
+        self._flush_scheduled = False
+
+    def _start_flush(self) -> None:
+        if self._flush_scheduled or not self._raw_bytes:
+            return
+        self._flush_scheduled = True
+
+        from litellm.utils import executor
+
+        try:
+            executor.submit(
+                self._litellm_logging_obj.flush_passthrough_collected_chunks,
+                raw_bytes=self._raw_bytes,
+                provider_config=self._provider_config,
+            )
+        except Exception as e:  # noqa: BLE001 # Safe catch-all for verbose logging
+            verbose_logger.exception(
+                "Failed to schedule passthrough spend-tracking flush; %d buffered chunks dropped: %s",
+                len(self._raw_bytes),
+                e,
+            )
+
+    def __iter__(self) -> PassthroughStreamingResponse:
+        return self
+
+    def __next__(self) -> bytes:
+        try:
+            chunk: Final = next(self._iterator)
+            self._raw_bytes.append(chunk)
+        except Exception:  # noqa: BLE001 # Safe catch-all for cleanup logic
+            self._start_flush()
+            try:
+                self._response.close()
+            except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                pass
+            raise
+        else:
+            return chunk
+
+    def send(self, value: bytes) -> bytes:
+        return self._iterator.send(value)
+
+    def throw(
+        self,
+        typ: BaseException | type[BaseException],
+        val: BaseException | object = None,
+        tb: TracebackType | None = None,
+    ) -> bytes:
+        return self._iterator.throw(typ, val, tb)  # pyright: ignore[reportCallIssue, reportArgumentType]  # matches one of the throw overloads
+
+    def close(self) -> None:
+        self._start_flush()
+        try:
+            self._response.close()
+        except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+            pass
 
 
 @client
@@ -37,10 +255,10 @@ async def allm_passthrough_route(
     api_key: str | None = None,
     request_query_params: dict | None = None,
     request_headers: dict | None = None,
-    content: Any | None = None,
+    content: RequestContent | None = None,
     data: dict | None = None,
     files: RequestFiles | None = None,
-    json: Any | None = None,
+    json: object | None = None,
     params: QueryParamTypes | None = None,
     cookies: CookieTypes | None = None,
     client: HTTPHandler | AsyncHTTPHandler | None = None,
@@ -64,7 +282,7 @@ async def allm_passthrough_route(
         from litellm.utils import ProviderConfigManager
 
         provider_config = cast(
-            Optional["BasePassthroughConfig"], kwargs.get("provider_config")
+            BasePassthroughConfig | None, kwargs.get("provider_config")
         ) or ProviderConfigManager.get_provider_passthrough_config(
             provider=LlmProviders(custom_llm_provider),
             model=model,
@@ -132,12 +350,12 @@ async def allm_passthrough_route(
         if resolved_custom_llm_provider:
             try:
                 provider_config = cast(
-                    Optional["BasePassthroughConfig"], kwargs.get("provider_config")
+                    BasePassthroughConfig | None, kwargs.get("provider_config")
                 ) or ProviderConfigManager.get_provider_passthrough_config(
                     provider=LlmProviders(resolved_custom_llm_provider),
                     model=model,
                 )
-            except Exception:
+            except Exception:  # noqa: BLE001 S110
                 # If we can't get provider config, pass None
                 pass
 
@@ -162,10 +380,10 @@ def llm_passthrough_route(
     api_key: str | None = None,
     request_query_params: dict | None = None,
     request_headers: dict | None = None,
-    content: Any | None = None,
+    content: RequestContent | None = None,
     data: dict | None = None,
     files: RequestFiles | None = None,
-    json: Any | None = None,
+    json: object | None = None,
     params: QueryParamTypes | None = None,
     cookies: CookieTypes | None = None,
     client: HTTPHandler | AsyncHTTPHandler | None = None,
@@ -190,7 +408,9 @@ def llm_passthrough_route(
 
     _is_async: Final = bool(kwargs.get("allm_passthrough_route", False))
 
-    litellm_logging_obj: Final = cast("LiteLLMLoggingObj", kwargs.get("litellm_logging_obj"))
+    litellm_logging_obj: Final = cast(
+        LiteLLMLoggingObj, kwargs.get("litellm_logging_obj")
+    )  # cast-ok: logging obj is constructed upstream; tests inject mocks
 
     model, custom_llm_provider, api_key, api_base = get_llm_provider(
         model=model,
@@ -235,7 +455,7 @@ def llm_passthrough_route(
     )
 
     provider_config: Final = cast(
-        Optional["BasePassthroughConfig"], kwargs.get("provider_config")
+        BasePassthroughConfig | None, kwargs.get("provider_config")
     ) or ProviderConfigManager.get_provider_passthrough_config(
         provider=LlmProviders(custom_llm_provider),
         model=model,
@@ -276,10 +496,13 @@ def llm_passthrough_route(
         forward_headers=False,
     )
 
+    _request_data: dict | None = (
+        data if isinstance(data, dict) else (json if isinstance(json, dict) else None)
+    )  # rebind-ok: conditional
     headers, signed_json_body = provider_config.sign_request(
         headers=headers,
         litellm_params=litellm_params_dict,
-        request_data=data if data else json,
+        request_data=_request_data,
         api_base=str(updated_url),
         model=model,
     )
@@ -301,9 +524,12 @@ def llm_passthrough_route(
     )
 
     ## IS STREAMING REQUEST
+    _streaming_request_data: dict = (
+        data if isinstance(data, dict) else (json if isinstance(json, dict) else {})
+    )  # rebind-ok: conditional
     is_streaming_request: Final = provider_config.is_streaming_request(
         endpoint=endpoint,
-        request_data=data or json or {},
+        request_data=_streaming_request_data,
     )
 
     # Update logging object with streaming status
@@ -334,18 +560,26 @@ def llm_passthrough_route(
         else:
             # Sync path - client.client.send returns Response directly
             response: httpx.Response = client.client.send(request=request, stream=is_streaming_request)
-            response.raise_for_status()
+            try:
+                response.raise_for_status()
+            except Exception:  # noqa: BLE001 # Safe catch-all for cleanup logic
+                try:
+                    response.read()
+                except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                    pass
+                try:
+                    response.close()
+                except Exception:  # noqa: BLE001 S110 # Safe catch-all for cleanup logic
+                    pass
+                raise
 
-            if (
-                hasattr(response, "iter_bytes") and is_streaming_request
-            ):  # yield the chunk, so we can store it in the logging object
-                return _sync_streaming(response, litellm_logging_obj, provider_config)
+            if hasattr(response, "iter_bytes") and is_streaming_request:
+                return PassthroughStreamingResponse(response, litellm_logging_obj, provider_config)
             else:
-                # For non-streaming responses, yield the entire response
                 return response
     except Exception as e:
-        if provider_config is None:
-            raise e
+        # provider_config is guaranteed non-None here due to the earlier guard
+        assert provider_config is not None
         raise base_llm_http_handler._handle_error(
             e=e,
             provider_config=provider_config,
@@ -356,8 +590,8 @@ async def _async_passthrough_request(
     client: HTTPHandler | AsyncHTTPHandler,
     request: httpx.Request,
     is_streaming_request: bool,
-    litellm_logging_obj: "LiteLLMLoggingObj",
-    provider_config: "BasePassthroughConfig",
+    litellm_logging_obj: LiteLLMLoggingObj,
+    provider_config: BasePassthroughConfig,
 ) -> httpx.Response | AsyncGenerator[Any, Any]:
     """
     Handle async passthrough requests.
@@ -369,8 +603,7 @@ async def _async_passthrough_request(
     # Check if it's a coroutine and await it
     if asyncio.iscoroutine(response_result):
         if is_streaming_request:
-            # Pass the coroutine to _async_streaming which will await it
-            return _async_streaming(
+            return await AsyncPassthroughStreamingResponse(  # pyright: ignore[reportGeneralTypeIssues]  # structural type check misses __await__
                 response=response_result,
                 litellm_logging_obj=litellm_logging_obj,
                 provider_config=provider_config,
@@ -383,84 +616,3 @@ async def _async_passthrough_request(
     else:
         # Fallback for sync-like behavior (shouldn't happen in async path)
         raise Exception("Expected coroutine from async client")
-
-
-def _sync_streaming(
-    response: httpx.Response,
-    litellm_logging_obj: "LiteLLMLoggingObj",
-    provider_config: "BasePassthroughConfig",
-):
-    from litellm.utils import executor
-
-    raw_bytes: Final[list[bytes]] = []
-    flush_scheduled = False
-    try:
-        for chunk in response.iter_bytes():
-            raw_bytes.append(chunk)
-            yield chunk
-    finally:
-        if not flush_scheduled and raw_bytes:
-            flush_scheduled = True
-            try:
-                executor.submit(
-                    litellm_logging_obj.flush_passthrough_collected_chunks,
-                    raw_bytes=raw_bytes,
-                    provider_config=provider_config,
-                )
-            except Exception as e:
-                verbose_logger.exception(
-                    "Failed to schedule passthrough spend-tracking flush "
-                    "in _sync_streaming; %d buffered chunks dropped: %s",
-                    len(raw_bytes),
-                    e,
-                )
-
-
-async def _async_streaming(
-    response: Coroutine[Any, Any, httpx.Response],
-    litellm_logging_obj: "LiteLLMLoggingObj",
-    provider_config: "BasePassthroughConfig",
-):
-    iter_response: Final = await response
-
-    try:
-        iter_response.raise_for_status()
-    except Exception:
-        try:
-            await iter_response.aclose()
-        except Exception:
-            pass
-        raise
-
-    raw_bytes: Final[list[bytes]] = []
-    flush_scheduled = False
-    try:
-        async for chunk in iter_response.aiter_bytes():
-            raw_bytes.append(chunk)
-            yield chunk
-    except Exception:
-        try:
-            await iter_response.aclose()
-        except Exception:
-            pass
-        raise
-    finally:
-        # GeneratorExit (raised on client disconnect) is not caught by
-        # `except Exception`; the finally block ensures partial usage
-        # still gets flushed for spend tracking. See LIT-2642.
-        if not flush_scheduled and raw_bytes:
-            flush_scheduled = True
-            try:
-                asyncio.create_task(
-                    litellm_logging_obj.async_flush_passthrough_collected_chunks(
-                        raw_bytes=raw_bytes,
-                        provider_config=provider_config,
-                    )
-                )
-            except Exception as e:
-                verbose_logger.exception(
-                    "Failed to schedule passthrough spend-tracking flush "
-                    "in _async_streaming; %d buffered chunks dropped: %s",
-                    len(raw_bytes),
-                    e,
-                )

litellm/proxy/_types.py

@@ -474,6 +474,7 @@ class LiteLLMRoutes(enum.Enum):
         "/vllm",
         "/mistral",
         "/milvus",
+        "/gigachat",
         "/watsonx",
     ]
 

litellm/proxy/common_request_processing.py

@@ -1495,6 +1495,35 @@ class ProxyBaseLLMRequestProcessing:
     def __init__(self, data: dict):
         self.data = data
 
+    @staticmethod
+    def _merge_passthrough_streaming_headers(
+        response_headers: httpx.Headers | dict | None,
+        custom_headers: dict,
+    ) -> dict:
+        """
+        Merge upstream passthrough headers with proxy/custom headers.
+
+        Proxy/custom headers win on key collisions.
+        """
+        excluded_headers: Final = {  # mutable-ok: set of header names to exclude from forwarding
+            "transfer-encoding",
+            "content-encoding",
+            "set-cookie",
+            "connection",
+            "keep-alive",
+            "proxy-authenticate",
+            "proxy-authorization",
+            "te",
+            "trailer",
+            "upgrade",
+        }
+
+        merged_headers: Final = {  # mutable-ok: dict comprehension for merged headers forwarded to httpx
+            key: value for key, value in dict(response_headers or {}).items() if key.lower() not in excluded_headers
+        }
+        merged_headers.update(custom_headers)
+        return merged_headers
+
     @staticmethod
     def get_custom_headers(
         *,
@@ -2389,6 +2418,16 @@ class ProxyBaseLLMRequestProcessing:
                     )
 
                 if route_type == "allm_passthrough_route":
+                    upstream_response_headers: Final = getattr(response, "headers", None)
+                    streaming_headers: Final = (
+                        ProxyBaseLLMRequestProcessing._merge_passthrough_streaming_headers(
+                            response_headers=upstream_response_headers,
+                            custom_headers=custom_headers,
+                        )
+                        if upstream_response_headers is not None
+                        else custom_headers
+                    )
+
                     # Check if response is an async generator
                     if self._is_streaming_response(response):
                         if asyncio.iscoroutine(response):
@@ -2418,11 +2457,11 @@ class ProxyBaseLLMRequestProcessing:
 
                         # For passthrough routes, stream directly without error parsing
                         # since we're dealing with raw binary data (e.g., AWS event streams)
-                        return StreamingResponse(
-                            content=generator,
-                            status_code=status.HTTP_200_OK,
+                        return _UpstreamClosingStreamingResponse(
+                            content=generator,  # pyright: ignore[reportArgumentType]  # generator-configured StreamingResponse
+                            status_code=getattr(response, "status_code", status.HTTP_200_OK),
                             media_type=self._passthrough_event_stream_media_type(),
-                            headers=custom_headers,
+                            headers=streaming_headers,
                         )
                     else:
                         _early = await self._handle_non_streaming_allm_passthrough_route(
@@ -2437,7 +2476,7 @@ class ProxyBaseLLMRequestProcessing:
                         return StreamingResponse(
                             content=response.aiter_bytes(),
                             status_code=response.status_code,
-                            headers=custom_headers,
+                            headers=streaming_headers,
                         )
                 elif route_type == "anthropic_messages":
                     # Check if response is actually a streaming response (async generator)

litellm/proxy/guardrails/guardrail_hooks/headroom/headroom.py

@@ -433,7 +433,7 @@ class HeadroomGuardrail(CustomGuardrail):
             payload["model"] = model
 
         try:
-            raw_response: HttpxResponse = await self.async_handler.post(  # pyright: ignore[reportUnknownMemberType]
+            raw_response: HttpxResponse = await self.async_handler.post(  # pyright: ignore[reportUnknownMemberType]  # AsyncHTTPHandler.post is untyped
                 url=f"{self.headroom_api_base}/v1/compress",
                 json=payload,
                 headers=self._request_headers(),
@@ -570,7 +570,7 @@ class HeadroomGuardrail(CustomGuardrail):
             params["query"] = query
 
         try:
-            raw_response: HttpxResponse = await self.async_handler.get(  # pyright: ignore[reportUnknownMemberType]
+            raw_response: HttpxResponse = await self.async_handler.get(  # pyright: ignore[reportUnknownMemberType]  # AsyncHTTPHandler.get is untyped
                 url=f"{self.headroom_api_base}/v1/retrieve/{hash_value}",
                 params=params,
                 headers=self._request_headers(),

litellm/proxy/guardrails/guardrail_hooks/repelloai/repelloai.py

@@ -197,7 +197,7 @@ class RepelloAIGuardrail(CustomGuardrail):
         repelloai_response: RepelloAIAnalyzeResponse | None = None
         try:
             verbose_proxy_logger.debug("RepelloAI Argus request: %s", request)
-            response: Final[HttpxResponse] = await self.async_handler.post(  # pyright: ignore[reportUnknownMemberType]
+            response: Final[HttpxResponse] = await self.async_handler.post(  # pyright: ignore[reportUnknownMemberType]  # AsyncHTTPHandler.post is untyped
                 url=endpoint,
                 headers={"X-API-Key": self.repelloai_api_key},
                 json=request,

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -19,13 +19,13 @@ import re
 import secrets
 import traceback
 from collections.abc import Awaitable, Callable, Iterator, Mapping, Sequence
-from contextlib import AbstractAsyncContextManager
 from datetime import datetime, timedelta, timezone
 from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Protocol, TypeVar, cast
 
 import fastapi
 import yaml
 from fastapi import APIRouter, Depends, Header, HTTPException, Query, Request, status
+from typing_extensions import ReadOnly, TypedDict
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -155,6 +155,7 @@ from litellm.types.utils import (
 )
 
 if TYPE_CHECKING:
+    import prisma
     from prisma import Prisma
     from prisma import models as prisma_models
 
@@ -182,6 +183,14 @@ class _TxTables(Protocol):
     litellm_proxymodeltable: TableActions[object]
 
 
+class _ModelParamsUpdate(TypedDict):
+    litellm_params: ReadOnly["prisma.Json"]
+
+
+class _ModelRowWhere(TypedDict):
+    model_id: ReadOnly[str]
+
+
 class _ConfigTableActions(Protocol):
     """Config table surface this module needs; the shared repository seam exposes no ``update``."""
 
@@ -273,12 +282,6 @@ def _env_vars_param_value(param: _EnvVarsParam) -> Mapping[str, str] | None:
     return param.param_value
 
 
-def _tx_tables_context(
-    open_tx: Callable[[], AbstractAsyncContextManager[_TxTables]],
-) -> AbstractAsyncContextManager[_TxTables]:
-    return open_tx()
-
-
 async def _check_custom_key_allowed(custom_key_value: str | None) -> None:
     """Raise 403 if custom API keys are disabled and a custom key was provided."""
     if custom_key_value is None:
@@ -4492,27 +4495,29 @@ async def _rotate_master_key(
     if models:
         decrypted_models: Final = proxy_config.decrypt_model_list_from_db(new_models=models)
         verbose_proxy_logger.debug("ABLE TO DECRYPT MODELS - len(decrypted_models): %s", len(decrypted_models))
-        new_models: Final[list[dict[str, object]]] = []
-        for model in decrypted_models:
-            new_model = await _add_model_to_db(
-                model_params=Deployment(**model),
-                user_api_key_dict=user_api_key_dict,
-                prisma_client=prisma_client,
-                new_encryption_key=new_master_key,
-                should_create_model_in_db=False,
-            )
-            if new_model:
-                _dumped = dict[str, object](_as_object_dict(new_model.model_dump(exclude_none=True)))
-                _dumped["litellm_params"] = prisma.Json(_dumped["litellm_params"])
-                _dumped["model_info"] = prisma.Json(_dumped["model_info"])
-                new_models.append(_dumped)
-        verbose_proxy_logger.debug("Resetting proxy model table")
-        async with _tx_tables_context(prisma_client.db.tx) as tx:
-            await tx.litellm_proxymodeltable.delete_many()
-            verbose_proxy_logger.debug("Creating %s models", len(new_models))
-            await tx.litellm_proxymodeltable.create_many(
-                data=new_models,
-            )
+        reencrypted_models: Final = tuple(
+            [
+                reencrypted
+                for model in decrypted_models
+                if (
+                    reencrypted := await _add_model_to_db(
+                        model_params=Deployment(**model),
+                        user_api_key_dict=user_api_key_dict,
+                        prisma_client=prisma_client,
+                        new_encryption_key=new_master_key,
+                        should_create_model_in_db=False,
+                    )
+                )
+            ]
+        )
+        verbose_proxy_logger.debug("Re-encrypting litellm_params on %s model rows", len(reencrypted_models))
+        async with prisma_client.db.tx(timeout=timedelta(minutes=2)) as tx_ctx:
+            tx: Final[_TxTables] = tx_ctx
+            for reencrypted_model in reencrypted_models:
+                await tx.litellm_proxymodeltable.update_many(
+                    data=_ModelParamsUpdate(litellm_params=prisma.Json(reencrypted_model.litellm_params)),
+                    where=_ModelRowWhere(model_id=reencrypted_model.model_id),
+                )
         await publish_config_change(redis_cache=coordination_redis_cache(), object_type="litellm_proxymodeltable")
     # 3. process config table
     try:

litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py

@@ -6,6 +6,8 @@ Provider-specific Pass-Through Endpoints
 Use litellm with Anthropic SDK, Vertex AI SDK, Cohere SDK, etc.
 """
 
+from __future__ import annotations
+
 import hmac
 import json
 import os
@@ -28,6 +30,7 @@ from litellm.constants import (
 )
 from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix
 from litellm.llms.anthropic.common_utils import AnthropicModelInfo
+from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
 from litellm.llms.vertex_ai.vertex_llm_base import VertexBase
 from litellm.proxy._types import *
 from litellm.proxy.auth.handle_jwt import JWTHandler
@@ -51,6 +54,7 @@ from litellm.proxy.pass_through_endpoints.pass_through_endpoints import (
     create_websocket_passthrough_route,
     websocket_passthrough_request,
 )
+from litellm.proxy.utils import ProxyLogging as ProxyLoggingType
 from litellm.proxy.utils import is_known_model
 from litellm.proxy.vector_store_endpoints.utils import (
     assert_proxy_admin_for_vector_store_index_management,
@@ -71,13 +75,17 @@ from litellm.utils import ProviderConfigManager
 from .passthrough_endpoint_router import PassthroughEndpointRouter
 
 if TYPE_CHECKING:
+    from litellm.proxy.proxy_server import ProxyConfig as _ProxyConfig
     from litellm.router import Router
 
+    ProxyConfig = _ProxyConfig  # rebind-ok: conditional type alias
+else:
+    ProxyConfig = Any  # rebind-ok: runtime fallback
+
 vertex_llm_base: Final = VertexBase()
 router: Final = APIRouter()
 openai_passthrough_router: Final = APIRouter()
 default_vertex_config: Final = None
-
 passthrough_endpoint_router: Final = PassthroughEndpointRouter()
 
 
@@ -510,8 +518,14 @@ async def milvus_proxy_route(
     request_body: Final = await get_request_body(request)
 
     # check collectionName
-    collection_name: Final = cast(str | None, request_body.get("collectionName"))
-    extra_headers = {}
+    _raw_collection_name: Final = request_body.get("collectionName")
+    if _raw_collection_name is not None and not isinstance(_raw_collection_name, str):
+        raise HTTPException(
+            status_code=400,
+            detail=f"collectionName must be a string. Got {type(_raw_collection_name).__name__}",
+        )
+    collection_name: str | None = _raw_collection_name  # rebind-ok: locally scoped conversion
+    extra_headers = {}  # mutable-ok: dict for extra headers; rebind-ok: reassigned later from credentials
     base_target_url: str | None = None
     if not collection_name:
         raise HTTPException(
@@ -1288,7 +1302,7 @@ def _resolve_vertex_model_from_router(
         vertex_location: Current vertex location (may be from URL)
 
     Returns:
-        Tuple of (encoded_endpoint, endpoint, vertex_project, vertex_location)
+        tuple of (encoded_endpoint, endpoint, vertex_project, vertex_location)
         with resolved values from router config
     """
     if not llm_router:
@@ -1717,7 +1731,7 @@ def get_vertex_ai_allowed_incoming_headers(request: Request) -> dict:
 
 
 def get_vertex_pass_through_handler(
-    call_type: Literal["discovery", "aiplatform"],
+    call_type: Literal["discovery", "aiplatform"],  # noqa: UP037
 ) -> BaseVertexAIPassThroughHandler:
     if call_type == "discovery":
         return VertexAIDiscoveryPassThroughHandler()
@@ -1741,7 +1755,7 @@ def _override_vertex_params_from_router_credentials(
         vertex_location: Current vertex location (from URL)
 
     Returns:
-        Tuple of (vertex_project, vertex_location) with overridden values if applicable
+        tuple of (vertex_project, vertex_location) with overridden values if applicable
     """
     if router_credentials is None:
         return vertex_project, vertex_location
@@ -1906,12 +1920,12 @@ async def _prepare_vertex_auth_headers(
             authenticated them is stripped on the credential-less branch
 
     Returns:
-        Tuple containing:
+        tuple containing:
             - headers: dict - Authentication headers to use
-            - base_target_url: Optional[str] - Updated base target URL
+            - base_target_url: str | None - Updated base target URL
             - headers_passed_through: bool - Whether headers were passed through from request
-            - vertex_project: Optional[str] - Updated vertex project ID
-            - vertex_location: Optional[str] - Updated vertex location
+            - vertex_project: str | None - Updated vertex project ID
+            - vertex_location: str | None - Updated vertex location
     """
     vertex_llm_base: Final = VertexBase()
     headers_passed_through = False
@@ -2151,8 +2165,6 @@ async def vertex_discovery_proxy_route(
     """
     import re
 
-    from litellm.types.vector_stores import LiteLLM_ManagedVectorStore
-
     # Extract vector store ID from endpoint if present (e.g., dataStores/test-litellm-app_1761094730750)
     vector_store_credentials: LiteLLM_ManagedVectorStore | None = None
     vector_store_id_match: Final = re.search(r"dataStores/([^/]+)", endpoint)
@@ -2559,7 +2571,7 @@ def _vertex_publisher_model_suffix(model: str) -> str:
     return f"{VERTEX_PUBLISHER_MODEL_PREFIX}{model.rsplit('/', 1)[-1]}"
 
 
-def _get_llm_router() -> "Router | None":
+def _get_llm_router() -> Router | None:
     from litellm.proxy.proxy_server import llm_router
 
     return llm_router
@@ -2599,7 +2611,7 @@ def _resolve_vertex_live_credentials(
 def _build_vertex_live_setup_model_rewriter(
     vertex_project: str | None,
     vertex_location: str | None,
-    llm_router: "Router | None",
+    llm_router: Router | None,
 ) -> Callable[[str], str] | None:
     """
     Rewrite the ``setup`` frame's model into the full Vertex resource path the Live API requires.
@@ -2619,7 +2631,7 @@ def _build_vertex_live_setup_model_rewriter(
     return rewrite
 
 
-def _resolve_alias_to_upstream_model(setup_model: str, llm_router: "Router | None") -> str:
+def _resolve_alias_to_upstream_model(setup_model: str, llm_router: Router | None) -> str:
     """
     The Live SDK wraps whatever the caller typed as ``models/<name>``, so a gateway alias arrives prefixed
     """
@@ -2809,6 +2821,238 @@ def create_generic_websocket_passthrough_endpoint(
     )
 
 
+@router.api_route(
+    "/gigachat/{endpoint:path}",
+    methods=["GET", "POST", "PUT", "DELETE", "PATCH"],  # mutable-ok: FastAPI route methods
+    tags=["Gigachat Pass-through", "pass-through"],  # mutable-ok: FastAPI route tags
+)
+async def gigachat_proxy_route(
+    endpoint: str,
+    request: Request,
+    fastapi_response: Response,
+    user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth)],
+) -> Response:
+    """
+    [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+    """
+    from litellm.proxy.proxy_server import (
+        general_settings,
+        llm_router,
+        proxy_config,
+        proxy_logging_obj,
+        select_data_generator,
+        user_api_base,
+        user_max_tokens,
+        user_model,
+        user_request_timeout,
+        user_temperature,
+        version,
+    )
+
+    ## check for streaming
+    request_body: Final[dict[str, object]] = await get_request_body(request)
+    is_router_model = False  # rebind-ok: conditionally set to True when model uses router
+
+    raw_model: Final = request_body.get("model")
+    model: Final = raw_model if isinstance(raw_model, str) else None
+    if model:
+        is_router_model = is_passthrough_request_using_router_model(
+            request_body, llm_router
+        )  # rebind-ok: conditionally set to True
+    elif any(word in endpoint for word in ("completions", "embeddings")):
+        raise HTTPException(
+            status_code=400, detail={"error": "Model is required in request body"}
+        )  # mutable-ok: HTTPException detail dict
+
+    # If router model, use dedicated router passthrough handler
+    # This uses the same common processing path as non-router models
+    if model and is_router_model and llm_router:
+        return await handle_gigachat_passthrough_router_model(
+            model=model,
+            endpoint=endpoint,
+            request=request,
+            request_body=request_body,
+            fastapi_response=fastapi_response,
+            llm_router=llm_router,
+            user_api_key_dict=user_api_key_dict,
+            proxy_logging_obj=proxy_logging_obj,
+            general_settings=general_settings,
+            proxy_config=proxy_config,
+            select_data_generator=select_data_generator,
+            user_model=user_model,
+            user_temperature=user_temperature,
+            user_request_timeout=user_request_timeout,
+            user_max_tokens=user_max_tokens,
+            user_api_base=user_api_base,
+            version=version,
+        )
+
+    verbose_proxy_logger.debug(
+        "Gigachat passthrough: Using direct Gigachat model '%s' for endpoint '%s'", model, endpoint
+    )
+
+    from litellm.llms.gigachat.authenticator import get_access_token
+    from litellm.llms.gigachat.utils import GIGACHAT_BASE_URL
+
+    base_target_url: Final = get_secret_str("GIGACHAT_API_BASE") or GIGACHAT_BASE_URL
+    request_path: Final = httpx.URL(endpoint).path
+    encoded_endpoint: Final = request_path if request_path.startswith("/") else f"/{request_path}"
+
+    base_url: Final = httpx.URL(base_target_url)
+    updated_url: Final = base_url.copy_with(
+        path=HttpPassThroughEndpointHelpers.join_base_and_endpoint_path(base_url, encoded_endpoint)
+    )
+
+    is_streaming_request: Final = await is_streaming_request_fn(request)
+
+    endpoint_func: Final = create_pass_through_route(
+        endpoint=endpoint,
+        target=str(updated_url),
+        custom_headers={"Authorization": f"Bearer {get_access_token()}"},
+        is_streaming_request=is_streaming_request,
+    )
+    return await endpoint_func(
+        request,
+        fastapi_response,
+        user_api_key_dict,
+    )
+
+
+async def handle_gigachat_passthrough_router_model(
+    model: str,
+    endpoint: str,
+    request: Request,
+    request_body: dict,
+    fastapi_response: Response,
+    llm_router: litellm.Router,
+    user_api_key_dict: UserAPIKeyAuth,
+    proxy_logging_obj: ProxyLoggingType,
+    general_settings: dict,
+    proxy_config: ProxyConfig,
+    select_data_generator: Callable,
+    user_model: str | None,
+    user_temperature: float | None,
+    user_request_timeout: float | None,
+    user_max_tokens: int | None,
+    user_api_base: str | None,
+    version: str | None,
+) -> Response | StreamingResponse:
+    """
+    Handle Gigachat passthrough for router models (models defined in config.yaml).
+
+    Uses the same common processing path as non-router models to ensure
+    metadata and hooks are properly initialized.
+
+    Args:
+        model: The router model name (e.g., "gigachat/gigachat-2")
+        endpoint: The Gigachat endpoint path (e.g., "/chat/completions")
+        request: The FastAPI request object
+        request_body: The parsed request body
+        llm_router: The LiteLLM router instance
+        user_api_key_dict: The user API key authentication dictionary
+        proxy_logging_obj: Proxy logging
+        general_settings: Proxy general settings
+        proxy_config: Proxy config
+        select_data_generator: Select data generator function
+        (additional args for common processing)
+
+    Returns:
+        Response or StreamingResponse depending on endpoint type
+    """
+    from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
+
+    # Detect streaming based on request body
+    is_streaming: Final = request_body.get("stream", False)  # pyright: ignore[reportUnknownVariableType]  # request_body is dict[Unknown, Unknown]
+
+    data: dict[str, Any] = await _read_request_body(
+        request=request
+    )  # mutable-ok: mutated in place by proxy pipeline; pyright: ignore[reportExplicitAny]  # Any needed for proxy pipeline
+    if user_api_key_dict is not None:
+        auth_metadata: Final = {
+            metadata_key: value
+            for metadata_key, value in (
+                ("user_api_key_user_id", getattr(user_api_key_dict, "user_id", None)),
+                ("user_api_key_team_id", getattr(user_api_key_dict, "team_id", None)),
+                ("user_api_key_org_id", getattr(user_api_key_dict, "org_id", None)),
+                ("agent_id", getattr(user_api_key_dict, "agent_id", None)),
+            )
+            if value is not None
+        }
+        existing_metadata: Final = data.get("metadata")
+        data["metadata"] = {
+            **(existing_metadata if isinstance(existing_metadata, dict) else {}),
+            **auth_metadata,
+        }
+
+    verbose_proxy_logger.debug(
+        "Gigachat router passthrough: model='%s', endpoint='%s', streaming=%s", model, endpoint, is_streaming
+    )
+
+    # Use the common processing path (same as non-router models)
+    # This ensures all metadata, hooks, and logging are properly initialized
+
+    data["model"] = model
+    data["method"] = request.method
+    data["endpoint"] = endpoint
+    data["json"] = request_body
+    data["custom_llm_provider"] = "gigachat"
+
+    # Remove sensitive keys from data
+    keys: Final = [  # mutable-ok: list of keys to remove from data
+        "gigachat_auth_url",
+        "gigachat_access_token",
+        "gigachat_scope",
+        "api_base",
+        "api_key",
+    ]
+    for key in keys:
+        data.pop(key, None)
+
+    client: Final = get_async_httpx_client(
+        llm_provider=LlmProviders.GIGACHAT,
+        params={  # mutable-ok: httpx client params
+            "timeout": httpx.Timeout(timeout=600.0, connect=5.0),
+        },
+    )
+
+    data["client"] = client
+    base_llm_response_processor: Final = ProxyBaseLLMRequestProcessing(data=data)
+
+    # Use the common passthrough processing to handle metadata and hooks
+    # This also handles all response formatting (streaming/non-streaming) and exceptions
+    try:
+        result = await base_llm_response_processor.base_passthrough_process_llm_request(  # rebind-ok: assigned once in try block
+            request=request,
+            fastapi_response=fastapi_response,
+            user_api_key_dict=user_api_key_dict,
+            proxy_logging_obj=proxy_logging_obj,
+            llm_router=llm_router,
+            general_settings=general_settings,
+            proxy_config=proxy_config,
+            select_data_generator=select_data_generator,
+            model=model,
+            user_model=user_model,
+            user_temperature=user_temperature,
+            user_request_timeout=user_request_timeout,
+            user_max_tokens=user_max_tokens,
+            user_api_base=user_api_base,
+            version=version,
+        )
+    except Exception as e:  # noqa: BLE001 # Safe catch-all for handle exception
+        # Use common exception handling
+        raise await base_llm_response_processor._handle_llm_api_exception(
+            e=e,
+            user_api_key_dict=user_api_key_dict,
+            proxy_logging_obj=proxy_logging_obj,
+        )
+    else:
+        if isinstance(result, StreamingResponse):
+            if result.headers.get("Content-Type") is None:
+                result.headers["Content-Type"] = "text/event-stream; charset=utf-8"
+
+        return result
+
+
 @router.api_route(
     "/watsonx/{endpoint:path}",
     methods=["GET", "POST", "PUT", "DELETE", "PATCH"],

litellm/proxy/public_endpoints/provider_create_fields.json

@@ -1318,6 +1318,68 @@
     ],
     "default_model_placeholder": "gpt-3.5-turbo"
   },
+  {
+    "provider": "GIGACHAT",
+    "provider_display_name": "GigaChat",
+    "litellm_provider": "gigachat",
+    "credential_fields": [
+      {
+        "key": "api_base",
+        "label": "API Base",
+        "placeholder": null,
+        "tooltip": null,
+        "required": false,
+        "field_type": "text",
+        "options": null,
+        "default_value": null
+      },
+      {
+        "key": "api_key",
+        "label": "API Key",
+        "placeholder": null,
+        "tooltip": null,
+        "required": false,
+        "field_type": "password",
+        "options": null,
+        "default_value": null
+      },
+      {
+        "key": "gigachat_scope",
+        "label": "Scope",
+        "placeholder": null,
+        "tooltip": null,
+        "required": false,
+        "field_type": "select",
+        "options": [
+          "GIGACHAT_API_PERS",
+          "GIGACHAT_API_B2B",
+          "GIGACHAT_API_CORP"
+        ],
+        "default_value": "GIGACHAT_API_PERS"
+      },
+      {
+        "key": "gigachat_auth_url",
+        "label": "Auth URL",
+        "placeholder": null,
+        "tooltip": null,
+        "required": false,
+        "field_type": "text",
+        "options": null,
+        "default_value": null
+      },
+      {
+        "key": "gigachat_access_token",
+        "label": "Access token",
+        "placeholder": null,
+        "tooltip": "Disable OAuth, provide value to authorization.",
+        "required": false,
+        "field_type": "password",
+        "options": null,
+        "default_value": null
+      }
+    ],
+    "default_model_placeholder": "GigaChat-2"
+  },
   {
     "provider": "GITHUB",
     "provider_display_name": "Github",

litellm/proxy/response_polling/background_streaming.py

@@ -144,10 +144,8 @@ async def background_streaming_task(
 
         # Process streaming response following OpenAI events format
         # https://platform.openai.com/docs/api-reference/responses-streaming
-        output_items: Final = dict[str, _OutputItem]()  # Track output items by ID
-        accumulated_text: Final = dict[
-            tuple[str, int], str
-        ]()  # Track accumulated text deltas by (item_id, content_index)
+        output_items: Final = dict[str, _OutputItem]()
+        accumulated_text: Final = dict[tuple[str, int], str]()
 
         # ResponsesAPIResponse fields to extract from response.completed
         usage_data = None
@@ -262,7 +260,6 @@ async def background_streaming_task(
                                 if "content" in delta_item:
                                     content_list = delta_item["content"]
                                     if content_index < len(content_list):
-                                        # Update existing content part with accumulated text
                                         content_entry = content_list[content_index]
                                         if isinstance(content_entry, dict):
                                             content_entry["text"] = accumulated_text[key]

litellm/rag/ingestion/vertex_ai_ingestion.py

@@ -186,7 +186,6 @@ class VertexAIRAGIngestion(BaseRAGIngestion, VertexBase):
         base_url: Final = get_vertex_base_url(self.location)
         url: Final = f"{base_url}/v1beta1/projects/{self.project_id}/locations/{self.location}/ragCorpora"
 
-        # Build request body with camelCase keys (Vertex AI API format)
         vector_db_config: Final = self.vector_store_config.get("vector_db_config")
         embedding_model: Final = self.vector_store_config.get("embedding_model")
         embedding_model_config: Final = (
@@ -447,7 +446,6 @@ class VertexAIRAGIngestion(BaseRAGIngestion, VertexBase):
         # Add max embedding requests per minute if specified
         max_embedding_qpm: Final = self.vector_store_config.get("max_embedding_requests_per_min")
 
-        # Build request body with camelCase keys (Vertex AI API format)
         chunking_config: Final = (
             {"chunkSize": chunk_size or 1024, "chunkOverlap": chunk_overlap or 200}
             if chunk_size or chunk_overlap

litellm/router.py

@@ -6486,6 +6486,8 @@ class Router:
                     **kwargs,
                 )
             elif call_type == "allm_passthrough_route":
+                if client:
+                    kwargs["client"] = client
                 return await self._ageneric_api_call_with_fallbacks(
                     original_function=original_function,
                     passthrough_on_no_deployment=True,
@@ -9152,7 +9154,8 @@ class Router:
             if _deployment_on_router is not None:
                 # deployment with this model_id exists on the router
                 if (
-                    deployment.litellm_params == _deployment_on_router.litellm_params
+                    deployment.model_name == _deployment_on_router.model_name
+                    and deployment.litellm_params == _deployment_on_router.litellm_params
                     and deployment.model_info == _deployment_on_router.model_info
                 ):
                     # No need to update

litellm/router_strategy/complexity_router/complexity_router.py

@@ -479,6 +479,33 @@ def _last_human_ask_index(
     )
 
 
+def _newest_turn_is_human_ask(
+    messages: Sequence[Mapping[str, object]] | None,
+    marker_pairs: tuple[tuple[str, str], ...] = _DEFAULT_REMINDER_MARKERS,
+) -> bool:
+    """Whether the request's newest turn carries a real human ask, i.e. this is a new ask rather
+    than an agent loop's continuation traffic.
+
+    Anchored on `_last_human_ask_index` so every surface's plumbing reads as a continuation:
+    chat-completions tool turns are role=tool, Messages-surface tool_result turns flatten to empty
+    human text, and a hybrid turn carrying an ask alongside a tool_result still counts as an ask.
+    Compared against the newest non-system message rather than the raw tail, because Claude Code
+    appends a system-role reminder after the human turn; that trailing plumbing is neither an ask
+    nor loop traffic and must not turn a fresh ask into a continuation. An unreadable request (no
+    messages) is treated as a continuation: there is no ask to classify, which is the same reading
+    `_extract_current_ask_and_system_prompt` gives it downstream.
+    """
+    if not messages:
+        return False
+    newest_non_system: Final = next(
+        (index for index in range(len(messages) - 1, -1, -1) if messages[index].get("role") != "system"),
+        None,
+    )
+    if newest_non_system is None:
+        return False
+    return _last_human_ask_index(messages, marker_pairs) == newest_non_system
+
+
 def _iter_system_scope_texts(
     body_system: object,
     messages: Sequence[Mapping[str, object]],
@@ -2247,14 +2274,18 @@ class ComplexityRouter(CustomLogger):
 
     @property
     def _uses_tier_pin(self) -> bool:
-        return bool(self.config.session_affinity and not self.config.plugins)
+        """classification_mode 'user_turn' implies the tier pin machinery: the pin write after each
+        pinnable classification is what gives a continuation a held decision to replay."""
+        return bool(
+            (self.config.session_affinity or self.config.classification_mode == "user_turn") and not self.config.plugins
+        )
 
     @property
     def _uses_deployment_pin(self) -> bool:
-        """session_affinity implies the deployment pin: a session frozen onto one model
+        """The tier pin implies the deployment pin: a session frozen onto one model
         group but load-balanced across its deployments would still go cache-cold, which
         is the exact failure both flags exist to prevent."""
-        return bool((self.config.deployment_affinity or self.config.session_affinity) and not self.config.plugins)
+        return bool(self.config.deployment_affinity and not self.config.plugins) or self._uses_tier_pin
 
     def _with_session_deployment_affinity(
         self, response: PreRoutingHookResponse | None
@@ -2282,6 +2313,11 @@ class ComplexityRouter(CustomLogger):
         pins the model chosen on the session's first turn and reuses it for every later
         turn, skipping classification entirely. Otherwise delegates to `_classify_and_route`.
 
+        When `classification_mode` is 'user_turn', the same pin is replayed only on
+        continuation turns (an agent loop's tool traffic); a new human ask always falls
+        through to classification, so the session can still move tiers between asks.
+        With both knobs on, session_affinity's pin-first behavior wins.
+
         Skipped entirely when `plugins` are configured: reusing a stale pin would bypass
         the plugin pipeline on every turn after the first, since a pinned model was never
         re-checked against a policy plugin whose decision can change between turns (e.g. a
@@ -2305,7 +2341,13 @@ class ComplexityRouter(CustomLogger):
         session_id: Final = self._get_session_id_from_request_kwargs(request_kwargs) if use_session_affinity else None
         cache_key = self._get_session_affinity_cache_key(session_id, request_kwargs) if session_id is not None else None
 
-        if cache_key is not None:
+        # In 'user_turn' mode a held pin is replayed only on continuation turns; a new human
+        # ask falls through and re-classifies. session_affinity restores pin-first for asks too.
+        pin_replay_allowed: Final = bool(self.config.session_affinity) or not _newest_turn_is_human_ask(
+            resolved_messages, self._reminder_markers
+        )
+
+        if cache_key is not None and pin_replay_allowed:
             pinned_value: Final = await self.litellm_router_instance.cache.async_get_cache(key=cache_key)
             pinned_pin: Final = _parse_session_affinity_pin(pinned_value)
             if pinned_pin is not None:
@@ -2354,10 +2396,11 @@ class ComplexityRouter(CustomLogger):
                         kwargs_metadata: Final = request_kwargs.setdefault("metadata", {})
                         if isinstance(kwargs_metadata, dict):
                             kwargs_metadata[ADAPTIVE_ROUTER_CHOSEN_MODEL_KEY] = routed_model
+                    replay_cause: Final[RoutingDecisionCause] = (
+                        "session_affinity_pin" if self.config.session_affinity else "user_turn_continuation"
+                    )
                     cause: RoutingDecisionCause = (
-                        "plan_mode"
-                        if plan_floored
-                        else ("session_affinity_escalation" if escalated else "session_affinity_pin")
+                        "plan_mode" if plan_floored else ("session_affinity_escalation" if escalated else replay_cause)
                     )
                     verbose_router_logger.info(
                         "ComplexityRouter: routing decision cause=%s, routed_model=%s", cause, routed_model

litellm/router_strategy/complexity_router/config.py

@@ -839,6 +839,21 @@ class ComplexityRouterConfig(BaseModel):
         description="Minimum cosine similarity for a semantic keyword match",
     )
 
+    classification_mode: Literal["every_request", "user_turn"] = Field(
+        default="every_request",
+        description=(
+            "When to run the complexity classifier. 'every_request' (the default) classifies every "
+            "inference request, including the tool-result continuation turns of an agentic loop. "
+            "'user_turn' classifies only requests whose newest turn is a new human ask and replays "
+            "the session's held routing decision on continuation turns, which cuts classifier "
+            "spend and eliminates mid-loop model switches. Continuations with no held decision to "
+            "replay (no resolvable session_id, expired pin, fresh restart) still classify. Unlike "
+            "session_affinity, a new human ask always re-classifies, so a session can still move "
+            "tiers between asks. Suppressed when plugins are configured, for the same reason "
+            "session_affinity is: a replayed decision would bypass the plugin pipeline."
+        ),
+    )
+
     # Session affinity: pin the first turn's routed model for the rest of the session
     session_affinity: bool = Field(
         default=False,

litellm/types/llms/openai.py

@@ -2162,6 +2162,42 @@ class OpenAIRealtimeDoneEvent(TypedDict):
     type: Literal["response.done"]
 
 
+class OpenAIRealtimeInputAudioBufferSpeechEvent(TypedDict):
+    type: ReadOnly[Literal["input_audio_buffer.speech_started", "input_audio_buffer.speech_stopped"]]
+    event_id: ReadOnly[str]
+    item_id: ReadOnly[str]
+
+
+class OpenAIRealtimeInputAudioTranscriptionDelta(TypedDict):
+    type: ReadOnly[Literal["conversation.item.input_audio_transcription.delta"]]
+    event_id: ReadOnly[str]
+    item_id: ReadOnly[str]
+    content_index: ReadOnly[int]
+    delta: ReadOnly[str]
+
+
+class OpenAIRealtimeInputAudioTranscriptionCompleted(TypedDict):
+    type: ReadOnly[Literal["conversation.item.input_audio_transcription.completed"]]
+    event_id: ReadOnly[str]
+    item_id: ReadOnly[str]
+    content_index: ReadOnly[int]
+    transcript: ReadOnly[str]
+
+
+class OpenAIRealtimeUsageTokenDetails(TypedDict):
+    audio_tokens: ReadOnly[int]
+    text_tokens: ReadOnly[int]
+    cached_tokens: NotRequired[ReadOnly[int]]
+
+
+class OpenAIRealtimeResponseUsage(TypedDict):
+    input_tokens: ReadOnly[int]
+    output_tokens: ReadOnly[int]
+    total_tokens: ReadOnly[int]
+    input_token_details: NotRequired[ReadOnly[OpenAIRealtimeUsageTokenDetails]]
+    output_token_details: NotRequired[ReadOnly[OpenAIRealtimeUsageTokenDetails]]
+
+
 class OpenAIRealtimeEventTypes(Enum):
     SESSION_CREATED = "session.created"
     # Beta delta event names
@@ -2199,6 +2235,9 @@ OpenAIRealtimeEvents = (
     | OpenAIRealtimeOutputItemDone
     | OpenAIRealtimeFunctionCallArgumentsDone
     | OpenAIRealtimeDoneEvent
+    | OpenAIRealtimeInputAudioBufferSpeechEvent
+    | OpenAIRealtimeInputAudioTranscriptionDelta
+    | OpenAIRealtimeInputAudioTranscriptionCompleted
 )
 
 OpenAIRealtimeStreamList = list[OpenAIRealtimeEvents]

litellm/types/utils.py

@@ -2842,6 +2842,11 @@ RoutingDecisionCause = Literal[
     "housekeeping",
     "session_affinity_pin",
     "session_affinity_escalation",
+    # classification_mode 'user_turn': the request is an agent loop's continuation turn (no new
+    # human ask), so the session's held routing decision was replayed and the classifier was never
+    # called. Distinct from "session_affinity_pin", which reports the session_affinity flag pinning
+    # every turn including new asks; this cause only appears when session_affinity is off.
+    "user_turn_continuation",
     "default_fallback",
     "keyword",
     "quality_tier",

litellm/utils.py

@@ -8830,6 +8830,12 @@ class ProviderConfigManager:
             )
 
             return AzurePassthroughConfig()
+        elif LlmProviders.GIGACHAT == provider:
+            from litellm.llms.gigachat.passthrough.transformation import (
+                GigaChatPassthroughConfig,
+            )
+
+            return GigaChatPassthroughConfig()
         elif LlmProviders.WATSONX == provider:
             from litellm.llms.watsonx.passthrough.transformation import (
                 WatsonxPassthroughConfig,

model_prices_and_context_window.json

@@ -553,6 +553,26 @@
         "supports_response_schema": true,
         "supports_vision": true
     },
+    "amazon.nova-sonic-v1:0": {
+        "input_cost_per_audio_token": 3.4e-06,
+        "input_cost_per_token": 6e-08,
+        "litellm_provider": "bedrock",
+        "mode": "realtime",
+        "output_cost_per_audio_token": 1.36e-05,
+        "output_cost_per_token": 2.4e-07,
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
+    "amazon.nova-2-sonic-v1:0": {
+        "input_cost_per_audio_token": 3e-06,
+        "input_cost_per_token": 3.3e-07,
+        "litellm_provider": "bedrock",
+        "mode": "realtime",
+        "output_cost_per_audio_token": 1.2e-05,
+        "output_cost_per_token": 2.75e-06,
+        "supports_audio_input": true,
+        "supports_audio_output": true
+    },
     "amazon.rerank-v1:0": {
         "input_cost_per_query": 0.001,
         "input_cost_per_token": 0.0,
@@ -19566,6 +19586,61 @@
         "supports_system_messages": true,
         "supports_tool_choice": true
     },
+    "friendliai/zai-org/GLM-5.3-Flash": {
+        "litellm_provider": "friendliai",
+        "supports_reasoning": true,
+        "supports_function_calling": true,
+        "max_input_tokens": 1048576,
+        "max_tokens": 1048576,
+        "max_output_tokens": 1048576,
+        "input_cost_per_token": 1.5e-07,
+        "output_cost_per_token": 5e-07,
+        "cache_read_input_token_cost": 3e-08,
+        "supports_prompt_caching": true,
+        "reasoning_effort_levels": [
+            "low",
+            "high",
+            "max"
+        ],
+        "supports_parallel_function_calling": true,
+        "supports_response_schema": true,
+        "supports_native_structured_output": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "mode": "chat",
+        "comment": "Native multimodal GLM model for efficient coding and long-horizon agent tasks",
+        "source": "https://api.friendli.ai/serverless/v1/models",
+        "supports_vision": true,
+        "supports_image_input": true,
+        "supports_video_input": true
+    },
+    "friendliai/zai-org/GLM-5.3": {
+        "litellm_provider": "friendliai",
+        "supports_reasoning": true,
+        "supports_function_calling": true,
+        "max_input_tokens": 1048576,
+        "max_tokens": 1048576,
+        "max_output_tokens": 1048576,
+        "input_cost_per_token": 1.26e-06,
+        "output_cost_per_token": 3.96e-06,
+        "cache_read_input_token_cost": 2.34e-07,
+        "supports_prompt_caching": true,
+        "reasoning_effort_levels": [
+            "low",
+            "high",
+            "max"
+        ],
+        "supports_parallel_function_calling": true,
+        "supports_response_schema": true,
+        "supports_native_structured_output": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "mode": "chat",
+        "comment": "Flagship GLM model for long-horizon coding, agents, and complex project delivery",
+        "source": "https://api.friendli.ai/serverless/v1/models",
+        "supports_vision": false,
+        "supports_image_input": false
+    },
     "ft:babbage-002": {
         "deprecation_date": "2026-10-23",
         "input_cost_per_token": 1.6e-06,
@@ -24344,7 +24419,7 @@
         "supports_response_schema": true,
         "supports_vision": true
     },
-    "gigachat/GigaChat-2-Lite": {
+    "gigachat/GigaChat-2": {
         "input_cost_per_token": 0.0,
         "litellm_provider": "gigachat",
         "max_input_tokens": 128000,
@@ -24406,6 +24481,15 @@
         "output_cost_per_token": 0.0,
         "output_vector_size": 2560
     },
+    "gigachat/GigaEmbeddings-3B-2025-09": {
+        "input_cost_per_token": 0.0,
+        "litellm_provider": "gigachat",
+        "max_input_tokens": 4096,
+        "max_tokens": 4096,
+        "mode": "embedding",
+        "output_cost_per_token": 0.0,
+        "output_vector_size": 2048
+    },
     "gmi/anthropic/claude-opus-4.5": {
         "input_cost_per_token": 5e-06,
         "litellm_provider": "gmi",

osv-scanner.toml

@@ -2,3 +2,8 @@
 id = "GHSA-w8v5-vhqr-4h9v"
 ignoreUntil = 2026-09-09
 reason = "diskcache has no fixed release published; remove this entry once one exists"
+
+[[IgnoredVulns]]
+id = "GHSA-h7x2-h6g9-p789"
+ignoreUntil = 2026-09-14
+reason = "mlflow has no fixed release published; remove this entry once one exists"

tests/e2e/management/test_model_tag_accessgroup_e2e.py

@@ -180,6 +180,12 @@ class ModelBlockBody(BaseModel):
     model_id: str
 
 
+class ModelBlockResponse(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+    model_id: str
+    blocked: bool
+
+
 class ModelInfoBlockDetail(BaseModel):
     id: str | None = None
     blocked: bool | None = None
@@ -245,11 +251,6 @@ class TestModelRoutes:
     def test_block_then_unblock_persists_to_model_info(
         self, client: ManagementClient, resources: ResourceManager
     ) -> None:
-        """The blocked flag's persistence is read back from /model/info, not from the
-        /model/block response: that route currently returns a non-2xx serialization
-        envelope even though the DB write lands, so the /model/info read-back is the
-        authoritative persistence contract and keeps this test valid once the
-        response shape is fixed."""
         model_name = f"e2e-mgmt-model-block-{unique_marker()}"
         model_id = _create_db_model(client, resources, model_name)
 
@@ -257,27 +258,25 @@ class TestModelRoutes:
             f"{model_name!r} already reports blocked in /model/info before /model/block ran"
         )
 
-        _ = client.proxy.transport.send(
-            "/model/block",
-            headers=client.proxy.transport.master,
-            json=ModelBlockBody(model_id=model_id),
-        )
-        _ = _poll(
-            client.proxy,
-            lambda: True if _model_blocked_flag(client, model_id) is True else None,
-            f"/model/info never reported {model_name!r} blocked after /model/block",
-        )
-
-        _ = client.proxy.transport.send(
-            "/model/unblock",
-            headers=client.proxy.transport.master,
-            json=ModelBlockBody(model_id=model_id),
-        )
-        _ = _poll(
-            client.proxy,
-            lambda: True if _model_blocked_flag(client, model_id) is not True else None,
-            f"/model/info never cleared blocked for {model_name!r} after /model/unblock",
-        )
+        for action, expected in (("block", True), ("unblock", False)):
+            response = unwrap(
+                client.proxy.transport.post(
+                    f"/model/{action}",
+                    headers=client.proxy.transport.master,
+                    json=ModelBlockBody(model_id=model_id),
+                    response_type=ModelBlockResponse,
+                )
+            )
+            assert response.model_id == model_id
+            assert response.blocked is expected
+            _ = _poll(
+                client.proxy,
+                lambda want=expected: True
+                if _model_blocked_flag(client, model_id) is want
+                else None,
+                f"/model/info never reported blocked={expected} for {model_name!r} "
+                f"after /model/{action}",
+            )
 
 
 class TestTagRoutes:

tests/e2e/ui/tests/settings/scim.spec.ts

@@ -0,0 +1,53 @@
+import { test, expect, Page as PlaywrightPage } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { Page } from "../../fixtures/pages";
+import { navigateToPage } from "../../helpers/navigation";
+
+const rootPath = (): string => process.env.SERVER_ROOT_PATH ?? "";
+
+async function createScimTokenViaUi(page: PlaywrightPage, alias: string): Promise<string> {
+  await navigateToPage(page, Page.AdminPanel);
+  await page.getByRole("tab", { name: "SCIM" }).click();
+
+  await expect(page.getByText("SCIM Tenant URL")).toBeVisible();
+  await expect(page.locator("input[disabled]").first()).toHaveValue(/\/scim\/v2$/);
+
+  await page.getByLabel("Token Name").fill(alias);
+  await page.getByRole("button", { name: "Create SCIM Token" }).click();
+
+  await expect(page.getByText(/copy this token now/i)).toBeVisible({ timeout: 15_000 });
+  const token = await page.locator('input[type="password"]').inputValue();
+  expect(token, "the one-time token panel shows a usable virtual key").toMatch(/^sk-/);
+  return token;
+}
+
+test.describe("Admin Settings - SCIM", () => {
+  test.use({ storageState: ADMIN_STORAGE_PATH });
+
+  test("Create SCIM Token shows the token once and offers to create another", async ({ page }) => {
+    await createScimTokenViaUi(page, `e2e-scim-ui-${Date.now()}`);
+
+    await page.getByRole("button", { name: "Create Another Token" }).click();
+    await expect(page.getByRole("button", { name: "Create SCIM Token" })).toBeVisible();
+    await expect(page.getByText(/copy this token now/i)).toBeHidden();
+  });
+
+  test("a UI-minted SCIM token authorizes the SCIM API", async ({ page, request }) => {
+    test.skip(!process.env.LITELLM_LICENSE, "LITELLM_LICENSE not set in test env — /scim/v2 is premium-gated");
+
+    const token = await createScimTokenViaUi(page, `e2e-scim-api-${Date.now()}`);
+
+    const denied = await request.get(`${rootPath()}/scim/v2/Groups`, {
+      headers: { Authorization: "Bearer sk-not-a-real-key" },
+    });
+    expect(denied.status(), "an unknown key must not reach SCIM").toBe(401);
+
+    const res = await request.get(`${rootPath()}/scim/v2/Groups`, {
+      headers: { Authorization: `Bearer ${token}` },
+    });
+    expect(res.status(), `SCIM Groups listing failed: ${await res.text()}`).toBe(200);
+    const body = await res.json();
+    expect(body.schemas, "SCIM answers with a ListResponse").toContain("urn:ietf:params:scim:api:messages:2.0:ListResponse");
+    expect(Array.isArray(body.Resources), "SCIM ListResponse carries a Resources array").toBe(true);
+  });
+});

tests/logging_callback_tests/gcs_pub_sub_body/spend_logs_payload.json

@@ -11,7 +11,7 @@
     "user": "",
     "team_id": "",
     "organization_id": "",
-    "metadata": "{\"applied_guardrails\": [], \"attempted_fallbacks\": null, \"original_model_group\": null, \"batch_models\": null, \"batch_successful_requests\": null, \"batch_failed_requests\": null, \"mcp_tool_call_metadata\": null, \"vector_store_request_metadata\": null, \"routing_decision\": null, \"internal_call_origin\": null, \"guardrail_information\": null, \"compression_savings\": null, \"litellm_gateway_injected_cache\": null, \"router_metadata\": null, \"usage_object\": {\"completion_tokens\": 20, \"prompt_tokens\": 10, \"total_tokens\": 30, \"completion_tokens_details\": null, \"prompt_tokens_details\": null}, \"model_map_information\": {\"model_map_key\": \"gpt-4o\", \"model_map_value\": {\"key\": \"gpt-4o\", \"max_tokens\": 16384, \"max_input_tokens\": 128000, \"max_output_tokens\": 16384, \"input_cost_per_token\": 2.5e-06, \"cache_creation_input_token_cost\": null, \"cache_read_input_token_cost\": 1.25e-06, \"input_cost_per_character\": null, \"input_cost_per_token_above_128k_tokens\": null, \"input_cost_per_token_above_200k_tokens\": null, \"input_cost_per_query\": null, \"input_cost_per_second\": null, \"input_cost_per_audio_token\": null, \"input_cost_per_token_batches\": 1.25e-06, \"output_cost_per_token_batches\": 5e-06, \"output_cost_per_token\": 1e-05, \"output_cost_per_audio_token\": null, \"output_cost_per_character\": null, \"output_cost_per_token_above_128k_tokens\": null, \"output_cost_per_character_above_128k_tokens\": null, \"output_cost_per_token_above_200k_tokens\": null, \"output_cost_per_second\": null, \"output_cost_per_image\": null, \"output_vector_size\": null, \"litellm_provider\": \"openai\", \"mode\": \"chat\", \"supports_system_messages\": true, \"supports_response_schema\": true, \"supports_vision\": true, \"supports_function_calling\": true, \"supports_tool_choice\": true, \"supports_assistant_prefill\": false, \"supports_prompt_caching\": true, \"supports_audio_input\": false, \"supports_audio_output\": false, \"supports_pdf_input\": false, \"supports_embedding_image_input\": false, \"supports_native_streaming\": null, \"supports_web_search\": true, \"supports_reasoning\": false, \"search_context_cost_per_query\": {\"search_context_size_low\": 0.03, \"search_context_size_medium\": 0.035, \"search_context_size_high\": 0.05}, \"tpm\": null, \"rpm\": null, \"supported_openai_params\": [\"frequency_penalty\", \"logit_bias\", \"logprobs\", \"top_logprobs\", \"max_tokens\", \"max_completion_tokens\", \"modalities\", \"prediction\", \"n\", \"presence_penalty\", \"seed\", \"stop\", \"stream\", \"stream_options\", \"temperature\", \"top_p\", \"tools\", \"tool_choice\", \"function_call\", \"functions\", \"max_retries\", \"extra_headers\", \"parallel_tool_calls\", \"audio\", \"response_format\", \"user\"]}}, \"additional_usage_values\": {\"completion_tokens_details\": null, \"prompt_tokens_details\": null}, \"user_api_key\": null, \"user_api_key_alias\": null, \"user_api_key_team_id\": null, \"user_api_key_project_id\": null, \"user_api_key_project_alias\": null, \"user_api_key_org_id\": null, \"user_api_key_user_id\": null, \"user_api_key_team_alias\": null, \"spend_logs_metadata\": null, \"requester_ip_address\": null, \"status\": null, \"proxy_server_request\": null, \"error_information\": null, \"attempted_retries\": null, \"max_retries\": null}",
+    "metadata": "{\"applied_guardrails\": [], \"attempted_fallbacks\": null, \"original_model_group\": null, \"batch_models\": null, \"batch_successful_requests\": null, \"batch_failed_requests\": null, \"mcp_tool_call_metadata\": null, \"vector_store_request_metadata\": null, \"routing_decision\": null, \"internal_call_origin\": null, \"router_metadata\": null, \"guardrail_information\": null, \"compression_savings\": null, \"litellm_gateway_injected_cache\": null, \"usage_object\": {\"completion_tokens\": 20, \"prompt_tokens\": 10, \"total_tokens\": 30, \"completion_tokens_details\": null, \"prompt_tokens_details\": null}, \"model_map_information\": {\"model_map_key\": \"gpt-4o\", \"model_map_value\": {\"key\": \"gpt-4o\", \"max_tokens\": 16384, \"max_input_tokens\": 128000, \"max_output_tokens\": 16384, \"input_cost_per_token\": 2.5e-06, \"cache_creation_input_token_cost\": null, \"cache_read_input_token_cost\": 1.25e-06, \"input_cost_per_character\": null, \"input_cost_per_token_above_128k_tokens\": null, \"input_cost_per_token_above_200k_tokens\": null, \"input_cost_per_query\": null, \"input_cost_per_second\": null, \"input_cost_per_audio_token\": null, \"input_cost_per_token_batches\": 1.25e-06, \"output_cost_per_token_batches\": 5e-06, \"output_cost_per_token\": 1e-05, \"output_cost_per_audio_token\": null, \"output_cost_per_character\": null, \"output_cost_per_token_above_128k_tokens\": null, \"output_cost_per_character_above_128k_tokens\": null, \"output_cost_per_token_above_200k_tokens\": null, \"output_cost_per_second\": null, \"output_cost_per_image\": null, \"output_vector_size\": null, \"litellm_provider\": \"openai\", \"mode\": \"chat\", \"supports_system_messages\": true, \"supports_response_schema\": true, \"supports_vision\": true, \"supports_function_calling\": true, \"supports_tool_choice\": true, \"supports_assistant_prefill\": false, \"supports_prompt_caching\": true, \"supports_audio_input\": false, \"supports_audio_output\": false, \"supports_pdf_input\": false, \"supports_embedding_image_input\": false, \"supports_native_streaming\": null, \"supports_web_search\": true, \"supports_reasoning\": false, \"search_context_cost_per_query\": {\"search_context_size_low\": 0.03, \"search_context_size_medium\": 0.035, \"search_context_size_high\": 0.05}, \"tpm\": null, \"rpm\": null, \"supported_openai_params\": [\"frequency_penalty\", \"logit_bias\", \"logprobs\", \"top_logprobs\", \"max_tokens\", \"max_completion_tokens\", \"modalities\", \"prediction\", \"n\", \"presence_penalty\", \"seed\", \"stop\", \"stream\", \"stream_options\", \"temperature\", \"top_p\", \"tools\", \"tool_choice\", \"function_call\", \"functions\", \"max_retries\", \"extra_headers\", \"parallel_tool_calls\", \"audio\", \"response_format\", \"user\"]}}, \"additional_usage_values\": {\"completion_tokens_details\": null, \"prompt_tokens_details\": null}, \"user_api_key\": null, \"user_api_key_alias\": null, \"user_api_key_team_id\": null, \"user_api_key_project_id\": null, \"user_api_key_project_alias\": null, \"user_api_key_org_id\": null, \"user_api_key_user_id\": null, \"user_api_key_team_alias\": null, \"spend_logs_metadata\": null, \"requester_ip_address\": null, \"status\": null, \"proxy_server_request\": null, \"error_information\": null, \"attempted_retries\": null, \"max_retries\": null}",
     "cache_key": "Cache OFF",
     "spend": 0.00022500000000000002,
     "total_tokens": 30,

tests/test_litellm/endpoints/speech/speech_to_completion_bridge/test_transformation.py

@@ -0,0 +1,62 @@
+from typing import Final
+from unittest.mock import MagicMock
+
+import pytest
+
+import litellm
+from litellm.constants import OPENAI_CHAT_COMPLETION_PARAMS
+from litellm.endpoints.speech.speech_to_completion_bridge.transformation import (
+    SpeechToCompletionBridgeTransformationHandler,
+)
+
+GEMINI_TTS_MODEL: Final = "gemini-3.1-flash-tts-preview"
+
+
+def _bridge_request(response_format: str | None) -> dict:
+    optional_params: Final = (
+        {"temperature": 0.4} if response_format is None else {"temperature": 0.4, "response_format": response_format}
+    )
+    return SpeechToCompletionBridgeTransformationHandler().transform_request(
+        model=GEMINI_TTS_MODEL,
+        input="Hello from LiteLLM",
+        voice="Kore",
+        optional_params=optional_params,
+        litellm_params={},
+        headers={},
+        litellm_logging_obj=MagicMock(),
+        custom_llm_provider="gemini",
+    )
+
+
+@pytest.mark.parametrize("response_format", ["wav", "mp3", "pcm", None])
+def test_gemini_tts_request_keeps_speech_response_format_out_of_chat_params(response_format: str | None) -> None:
+    request: Final = _bridge_request(response_format)
+
+    assert "response_format" not in request
+    assert request["audio"] == {"voice": "Kore", "format": "pcm16"}
+    assert request["temperature"] == 0.4
+    assert request["modalities"] == ["audio"]
+
+    gemini_params: Final = litellm.get_optional_params(
+        model=GEMINI_TTS_MODEL,
+        custom_llm_provider="gemini",
+        **{param: value for param, value in request.items() if param in OPENAI_CHAT_COMPLETION_PARAMS},
+    )
+    assert gemini_params["speechConfig"] == {"voiceConfig": {"prebuiltVoiceConfig": {"voiceName": "Kore"}}}
+    assert "responseMimeType" not in gemini_params
+
+
+def test_non_gemini_request_forwards_speech_response_format_as_audio_format() -> None:
+    request: Final = SpeechToCompletionBridgeTransformationHandler().transform_request(
+        model="gpt-4o-audio-preview",
+        input="Hello from LiteLLM",
+        voice="alloy",
+        optional_params={"response_format": "wav"},
+        litellm_params={},
+        headers={},
+        litellm_logging_obj=MagicMock(),
+        custom_llm_provider="openai",
+    )
+
+    assert "response_format" not in request
+    assert request["audio"] == {"voice": "alloy", "format": "wav"}

tests/test_litellm/litellm_core_utils/test_get_llm_provider_endpoint_match.py

@@ -184,3 +184,26 @@ class TestTogetherApiBaseResolvesProvider:
 
         assert provider == "together_ai"
         assert api_base == "https://api.together.ai/v1"
+
+
+class TestGigachatApiBaseResolvesProvider:
+    """
+    Regression for the GigaChat api_base branch: the provider-mapping chain
+    carried an ``endpoint == "https://gigachat.devices.sberbank.ru/api/v1"``
+    elif, but the URL was never added to ``openai_compatible_endpoints``, so
+    the endpoint loop never fired the branch and a caller-supplied GigaChat
+    api_base raised BadRequestError instead of resolving to ``gigachat``.
+    """
+
+    def test_gigachat_api_base_resolves_to_gigachat(self, monkeypatch):
+        monkeypatch.setenv("GIGACHAT_API_KEY", "gigachat-key-from-env")
+
+        model, provider, dynamic_api_key, returned_api_base = get_llm_provider(
+            model="GigaChat-2",
+            api_base="https://gigachat.devices.sberbank.ru/api/v1",
+        )
+
+        assert provider == "gigachat"
+        assert dynamic_api_key == "gigachat-key-from-env"
+        assert returned_api_base == "https://gigachat.devices.sberbank.ru/api/v1"
+        assert model == "GigaChat-2"

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -6101,3 +6101,64 @@ def test_response_timing_metrics_survive_deepcopy(logging_obj):
     logging_obj.set_response_timing_metrics({"_response_ms": 12.5})
 
     assert copy.deepcopy(logging_obj).response_timing_metrics == {"_response_ms": 12.5}
+
+
+def test_passthrough_embeddings_result_swapped_for_callbacks():
+    """
+    Regression: for gigachat passthrough /embeddings, normalize_logging_result
+    produces an EmbeddingResponse, but the result swap only accepted
+    ModelResponse, so callbacks kept receiving the raw httpx.Response (which
+    crashes attribute readers like OTEL). The swap must cover
+    EmbeddingResponse too.
+    """
+    import datetime as dt
+
+    from litellm.types.utils import EmbeddingResponse
+
+    logging_obj = LitellmLogging(
+        model="EmbeddingsGigaR",
+        messages=[],
+        stream=False,
+        call_type="allm_passthrough_route",
+        start_time=time.time(),
+        litellm_call_id="passthrough-embed-call-id",
+        function_id="passthrough-embed-fn-id",
+    )
+    logging_obj.update_environment_variables(
+        litellm_params={},
+        optional_params={},
+        model="EmbeddingsGigaR",
+        custom_llm_provider="gigachat",
+        endpoint="/embeddings",
+        request_data={"model": "EmbeddingsGigaR", "input": ["hello"]},
+        input=["hello"],
+    )
+
+    httpx_response = httpx.Response(
+        200,
+        json={
+            "object": "list",
+            "data": [
+                {
+                    "object": "embedding",
+                    "embedding": [0.1, 0.2, 0.3],
+                    "index": 0,
+                    "usage": {"prompt_tokens": 5},
+                }
+            ],
+            "model": "EmbeddingsGigaR",
+        },
+        request=httpx.Request(
+            "POST", "https://gigachat.devices.sberbank.ru/api/v1/embeddings"
+        ),
+    )
+
+    _, _, swapped_result = logging_obj._success_handler_helper_fn(
+        result=httpx_response,
+        start_time=dt.datetime.now(),
+        end_time=dt.datetime.now(),
+        cache_hit=False,
+    )
+
+    assert isinstance(swapped_result, EmbeddingResponse)
+    assert swapped_result.data[0]["embedding"] == [0.1, 0.2, 0.3]

tests/test_litellm/llms/anthropic/chat/guardrail_translation/test_anthropic_guardrail_handler.py

@@ -1490,6 +1490,92 @@ class MockCanaryMaskingGuardrail(CustomGuardrail):
         return inputs
 
 
+class TestAnthropicMessagesImageSources:
+    """An Anthropic image block has three source shapes (`AnthropicMessagesImageParam.source`).
+
+    Only the base64 one carries "data", so reading that key alone drops url images
+    entirely -- for every guardrail consuming GenericGuardrailAPIInputs["images"],
+    not just Bedrock.
+    """
+
+    def _data(self, messages):
+        return {"model": "claude-sonnet-4-5", "messages": messages}
+
+    async def _images_seen(self, content) -> list[str]:
+        handler = AnthropicMessagesHandler()
+
+        class ImageRecordingGuardrail(MockCanaryMaskingGuardrail):
+            def __init__(self):
+                super().__init__()
+                self.seen_images: list[str] = []  # mutable-ok: accumulator for the assertion
+
+            async def apply_guardrail(self, inputs, request_data, input_type, logging_obj=None):
+                self.seen_images.extend(inputs.get("images") or [])
+                return await super().apply_guardrail(inputs, request_data, input_type, logging_obj)
+
+        guardrail = ImageRecordingGuardrail()
+        # The text block is what gets the guardrail invoked at all: a message with
+        # no text gives the handler nothing to scan, so it never reaches the
+        # guardrail and every source shape would look equally "dropped".
+        await handler.process_input_messages(
+            data=self._data([{"role": "user", "content": [{"type": "text", "text": "describe it"}, *content]}]),
+            guardrail_to_apply=guardrail,
+        )
+        return guardrail.seen_images
+
+    @pytest.mark.asyncio
+    async def test_url_source_reaches_the_guardrail(self):
+        """A url source has no "data" key, so it used to yield nothing at all."""
+        seen = await self._images_seen(
+            [{"type": "image", "source": {"type": "url", "url": "https://example.com/a.png"}}]
+        )
+
+        assert seen == ["https://example.com/a.png"]
+
+    @pytest.mark.asyncio
+    async def test_base64_source_carries_its_media_type(self):
+        """Bare base64 leaves the consumer no way to recover the format.
+
+        An API like Bedrock's ApplyGuardrail needs it to build the request, so the
+        media_type travels with the payload as a data URI.
+        """
+        seen = await self._images_seen(
+            [{"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "AAAA"}}]
+        )
+
+        assert seen == ["data:image/png;base64,AAAA"]
+
+    @pytest.mark.asyncio
+    async def test_base64_source_without_a_media_type_is_passed_through(self):
+        """There is no format to attach, so the payload goes through unchanged."""
+        seen = await self._images_seen([{"type": "image", "source": {"type": "base64", "data": "AAAA"}}])
+
+        assert seen == ["AAAA"]
+
+    @pytest.mark.asyncio
+    async def test_file_source_yields_nothing(self):
+        """The bytes live behind the Files API and this extractor has no client.
+
+        Documented as a known gap rather than silently handed on as a file_id string,
+        which a consumer would try to decode as an image.
+        """
+        seen = await self._images_seen([{"type": "image", "source": {"type": "file", "file_id": "file_abc"}}])
+
+        assert seen == []
+
+    @pytest.mark.asyncio
+    async def test_a_malformed_source_is_dropped_rather_than_passed_on(self):
+        seen = await self._images_seen(
+            [
+                {"type": "image", "source": {"type": "base64"}},
+                {"type": "image", "source": {"type": "url"}},
+                {"type": "image", "source": {"type": "base64", "data": ""}},
+            ]
+        )
+
+        assert seen == []
+
+
 class TestAnthropicMessagesToolResultScanning:
     """LIT-5251: tool_result blocks carry whatever a client's local tool fetched, so
     they are the request-path payload an indirect prompt injection actually arrives in.

tests/test_litellm/llms/bedrock/realtime/test_bedrock_realtime_handler.py

@@ -6,7 +6,7 @@ from unittest.mock import MagicMock
 
 import pytest
 
-
+import litellm
 from litellm.llms.bedrock.common_utils import BedrockError
 from litellm.llms.bedrock.realtime.handler import BedrockRealtime
 from litellm.llms.bedrock.realtime.transformation import BedrockRealtimeConfig
@@ -104,12 +104,24 @@ class RealtimeClientWS:
         self.closed = True
 
 
-class ImmediatelyEndingBedrockStream:
-    def __init__(self):
+class ScriptedBedrockReceiver:
+    def __init__(self, payloads):
+        self._payloads = list(payloads)
+
+    async def receive(self):
+        if not self._payloads:
+            return None
+        payload = self._payloads.pop(0)
+        return SimpleNamespace(value=SimpleNamespace(bytes_=payload.encode("utf-8")))
+
+
+class ScriptedBedrockStream:
+    def __init__(self, payloads):
         self.input_stream = FakeInputStream()
+        self._receiver = ScriptedBedrockReceiver(payloads)
 
     async def await_output(self):
-        return (None, EndedBedrockReceiver())
+        return (None, self._receiver)
 
 
 class FakeStaticCredentialsResolver:
@@ -151,7 +163,7 @@ def stub_aws_sdk_client(monkeypatch):
 
         async def invoke_model_with_bidirectional_stream(self, operation_input):
             captured["operation_input"] = operation_input
-            return ImmediatelyEndingBedrockStream()
+            return ScriptedBedrockStream(captured.get("scripted_payloads", []))
 
     package = types.ModuleType("aws_sdk_bedrock_runtime")
     client_module = types.ModuleType("aws_sdk_bedrock_runtime.client")
@@ -271,19 +283,132 @@ class TestBedrockRealtimeHandler:
         assert "sessionEnd" in event_names
         assert stream.input_stream.closed
 
+    @pytest.mark.asyncio
+    async def test_forwarded_events_are_filtered_to_logged_types_for_spend_logging(self):
+        handler = BedrockRealtime()
+        stream = ScriptedBedrockStream(
+            [
+                json.dumps({"event": {"userSpeechStart": {}}}),
+                json.dumps({"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}}),
+                json.dumps({"event": {"textOutput": {"content": "Hi"}}}),
+                json.dumps({"event": {"contentEnd": {"stopReason": "END_TURN"}}}),
+            ]
+        )
+        client_ws = RealtimeClientWS()
+
+        logged_events = [
+            event
+            async for event in handler._forward_bedrock_to_client(
+                stream,
+                client_ws,
+                BedrockRealtimeConfig(),
+                "amazon.nova-sonic-v1:0",
+                FakeLogging(),
+                {},
+            )
+        ]
+
+        assert [event["type"] for event in logged_events] == ["response.done"]
+        sent_types = [json.loads(message)["type"] for message in client_ws.sent_to_client]
+        assert "input_audio_buffer.speech_started" in sent_types
+        assert "response.text.delta" in sent_types
+        assert "response.done" in sent_types
+        assert client_ws.closed
+
+    @pytest.mark.asyncio
+    async def test_logged_event_types_star_collects_every_forwarded_event(self, monkeypatch):
+        monkeypatch.setattr(litellm, "logged_real_time_event_types", "*")
+        handler = BedrockRealtime()
+        stream = ScriptedBedrockStream(
+            [
+                json.dumps({"event": {"userSpeechStart": {}}}),
+                json.dumps({"event": {"userSpeechEnd": {}}}),
+            ]
+        )
+        client_ws = RealtimeClientWS()
+
+        logged_events = [
+            event
+            async for event in handler._forward_bedrock_to_client(
+                stream,
+                client_ws,
+                BedrockRealtimeConfig(),
+                "amazon.nova-sonic-v1:0",
+                FakeLogging(),
+                {},
+            )
+        ]
+
+        assert [event["type"] for event in logged_events] == [
+            "input_audio_buffer.speech_started",
+            "input_audio_buffer.speech_stopped",
+        ]
+
+    @pytest.mark.asyncio
+    async def test_trailing_usage_after_last_done_is_dispatched_for_spend(self, stub_aws_sdk_client, monkeypatch):
+        import litellm.llms.bedrock.realtime.handler as handler_module
+
+        dispatched = {}
+
+        class RecordingLogging(FakeLogging):
+            async def dispatch_success_handlers(self, result=None, prefer_async_handlers=False, **kwargs):
+                dispatched["events"] = result
+
+        class RecordingLoggingWorker:
+            def ensure_initialized_and_enqueue(self, coro):
+                dispatched["coro"] = coro
+
+        monkeypatch.setattr(handler_module, "GLOBAL_LOGGING_WORKER", RecordingLoggingWorker())
+        stub_aws_sdk_client["scripted_payloads"] = [
+            json.dumps(
+                {
+                    "event": {
+                        "usageEvent": {
+                            "totalInputTokens": 3,
+                            "totalOutputTokens": 6,
+                            "totalTokens": 9,
+                            "details": {
+                                "total": {
+                                    "input": {"speechTokens": 3, "textTokens": 0},
+                                    "output": {"speechTokens": 0, "textTokens": 6},
+                                }
+                            },
+                        }
+                    }
+                }
+            )
+        ]
+
+        await BedrockRealtime().async_realtime(
+            model="amazon.nova-sonic-v1:0",
+            websocket=RealtimeClientWS(),
+            logging_obj=RecordingLogging(),
+            aws_region_name="us-east-1",
+            aws_access_key_id="k",
+            aws_secret_access_key="s",
+        )
+        await dispatched["coro"]
+
+        assert [event["type"] for event in dispatched["events"]] == ["response.done"]
+        usage = dispatched["events"][0]["response"]["usage"]
+        assert (usage["input_tokens"], usage["output_tokens"], usage["total_tokens"]) == (3, 6, 9)
+        assert usage["input_token_details"] == {"audio_tokens": 3, "text_tokens": 0, "cached_tokens": 0}
+        assert usage["output_token_details"] == {"audio_tokens": 0, "text_tokens": 6}
+
     @pytest.mark.asyncio
     async def test_bedrock_stream_end_closes_client_websocket(self):
         handler = BedrockRealtime()
         client_ws = ClosableClientWS()
 
-        await handler._forward_bedrock_to_client(
+        async for _ in handler._forward_bedrock_to_client(
             EndedBedrockStream(),
             client_ws,
             BedrockRealtimeConfig(),
             "amazon.nova-sonic-v1:0",
             MagicMock(),
             {},
-        )
+        ):
+            pass
 
         assert client_ws.closed
 
@@ -320,9 +445,7 @@ class TestBedrockRealtimeSessionLifecycle:
             [json.dumps({"type": "session.update", "session": {"instructions": "hi", "modalities": ["text"]}})]
         )
 
-        await handler._forward_client_to_bedrock(
-            client_ws, stream, config, "amazon.nova-sonic-v1:0", {}, FakeLogging()
-        )
+        await handler._forward_client_to_bedrock(client_ws, stream, config, "amazon.nova-sonic-v1:0", {}, FakeLogging())
 
         acked = [json.loads(message) for message in client_ws.sent_to_client]
         updated = [event for event in acked if event["type"] == "session.updated"]
@@ -334,9 +457,7 @@ class TestBedrockRealtimeSessionLifecycle:
         handler = BedrockRealtime()
         config = BedrockRealtimeConfig()
         stream = FakeBedrockStream()
-        client_ws = DisconnectingClientWS(
-            [json.dumps({"type": "session.update", "session": {"instructions": "hi"}})]
-        )
+        client_ws = DisconnectingClientWS([json.dumps({"type": "session.update", "session": {"instructions": "hi"}})])
 
         await handler._forward_client_to_bedrock(client_ws, stream, config, "amazon.nova-sonic-v1:0", {})
 

tests/test_litellm/llms/bedrock/realtime/test_bedrock_realtime_transformation.py

@@ -827,5 +827,310 @@ class TestBedrockRealtimeSessionEvents:
         assert event["session"]["modalities"] == ["text", "audio"]
 
 
+class TestBedrockRealtimeUserEventsAndUsage:
+    """Regression tests for #38346: USER ASR transcripts, speech boundary events,
+    usage propagation, and duplicate response.created"""
+
+    @staticmethod
+    def _run(config, messages):
+        logging_obj = MagicMock()
+        logging_obj.litellm_trace_id = "trace_123"
+        state = {
+            "session_configuration_request": json.dumps({"configured": True}),
+            "current_output_item_id": None,
+            "current_response_id": None,
+            "current_conversation_id": None,
+            "current_delta_chunks": [],
+            "current_item_chunks": [],
+            "current_delta_type": None,
+        }
+        all_events = []
+        for msg in messages:
+            result = config.transform_realtime_response(
+                json.dumps(msg),
+                "amazon.nova-2-sonic-v1:0",
+                logging_obj,
+                realtime_response_transform_input=dict(state),
+            )
+            all_events.extend(result["response"])
+            state.update(
+                {
+                    "current_output_item_id": result["current_output_item_id"],
+                    "current_response_id": result["current_response_id"],
+                    "current_conversation_id": result["current_conversation_id"],
+                    "current_delta_chunks": result["current_delta_chunks"],
+                    "current_delta_type": result["current_delta_type"],
+                }
+            )
+        return all_events
+
+    def test_user_speech_start_and_stop_events(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [{"event": {"userSpeechStart": {}}}, {"event": {"userSpeechEnd": {}}}],
+        )
+        assert [e["type"] for e in events] == [
+            "input_audio_buffer.speech_started",
+            "input_audio_buffer.speech_stopped",
+        ]
+        assert all(e["event_id"] and e["item_id"] for e in events)
+        assert events[0]["item_id"] == events[1]["item_id"]
+
+    def test_utterance_lifecycle_shares_one_item_id(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {"event": {"userSpeechStart": {}}},
+                {"event": {"userSpeechEnd": {}}},
+                {
+                    "event": {
+                        "contentStart": {
+                            "role": "USER",
+                            "type": "TEXT",
+                            "additionalModelFields": json.dumps({"generationStage": "FINAL"}),
+                        }
+                    }
+                },
+                {"event": {"textOutput": {"content": "ready"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+            ],
+        )
+        item_ids = {e["item_id"] for e in events if "item_id" in e}
+        assert len(item_ids) == 1
+
+    def test_new_utterance_gets_new_item_id(self):
+        config = BedrockRealtimeConfig()
+        first = self._run(config, [{"event": {"userSpeechStart": {}}}, {"event": {"userSpeechEnd": {}}}])
+        second = self._run(config, [{"event": {"userSpeechStart": {}}}, {"event": {"userSpeechEnd": {}}}])
+        assert first[0]["item_id"] == first[1]["item_id"]
+        assert second[0]["item_id"] == second[1]["item_id"]
+        assert first[0]["item_id"] != second[0]["item_id"]
+
+    def test_user_transcript_emits_input_audio_transcription_events(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {
+                    "event": {
+                        "contentStart": {
+                            "role": "USER",
+                            "type": "TEXT",
+                            "additionalModelFields": json.dumps({"generationStage": "FINAL"}),
+                        }
+                    }
+                },
+                {"event": {"textOutput": {"content": "ready"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+            ],
+        )
+        deltas = [e for e in events if e["type"] == "conversation.item.input_audio_transcription.delta"]
+        completed = [e for e in events if e["type"] == "conversation.item.input_audio_transcription.completed"]
+        assert len(deltas) == 1 and deltas[0]["delta"] == "ready"
+        assert len(completed) == 1 and completed[0]["transcript"] == "ready"
+        assert deltas[0]["item_id"] == completed[0]["item_id"]
+        assert not any(e["type"] == "response.text.delta" for e in events)
+
+    def test_speculative_user_transcript_emits_delta_only(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {
+                    "event": {
+                        "contentStart": {
+                            "role": "USER",
+                            "type": "TEXT",
+                            "additionalModelFields": json.dumps({"generationStage": "SPECULATIVE"}),
+                        }
+                    }
+                },
+                {"event": {"textOutput": {"content": "rea"}}},
+            ],
+        )
+        assert [e["type"] for e in events] == ["conversation.item.input_audio_transcription.delta"]
+
+    def test_user_transcript_state_resets_on_content_end(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {"event": {"contentStart": {"role": "USER", "type": "TEXT"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+                {"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}},
+                {"event": {"textOutput": {"content": "Hi there"}}},
+            ],
+        )
+        text_deltas = [e for e in events if e["type"] == "response.text.delta"]
+        assert len(text_deltas) == 1 and text_deltas[0]["delta"] == "Hi there"
+        assert not any(e["type"].startswith("conversation.item.input_audio_transcription") for e in events)
+
+    def test_response_created_emitted_once_per_response(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}},
+                {"event": {"textOutput": {"content": "Hi"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+                {"event": {"contentStart": {"role": "ASSISTANT", "type": "AUDIO"}}},
+            ],
+        )
+        assert sum(1 for e in events if e["type"] == "response.created") == 1
+
+    def test_usage_event_propagates_to_response_done(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {
+                    "event": {
+                        "usageEvent": {
+                            "totalInputTokens": 25,
+                            "totalOutputTokens": 40,
+                            "totalTokens": 65,
+                            "details": {
+                                "total": {
+                                    "input": {"speechTokens": 20, "textTokens": 5},
+                                    "output": {"speechTokens": 30, "textTokens": 10},
+                                }
+                            },
+                        }
+                    }
+                },
+                {"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}},
+                {"event": {"textOutput": {"content": "Hi"}}},
+                {"event": {"contentEnd": {"stopReason": "END_TURN"}}},
+            ],
+        )
+        done_events = [e for e in events if e["type"] == "response.done"]
+        assert len(done_events) == 1
+        usage = done_events[0]["response"]["usage"]
+        assert usage["input_tokens"] == 25
+        assert usage["output_tokens"] == 40
+        assert usage["total_tokens"] == 65
+        assert usage["input_token_details"]["audio_tokens"] == 20
+        assert usage["input_token_details"]["text_tokens"] == 5
+        assert usage["output_token_details"]["audio_tokens"] == 30
+        assert usage["output_token_details"]["text_tokens"] == 10
+
+    def test_response_done_without_usage_event_reports_zero_usage(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}},
+                {"event": {"textOutput": {"content": "Hi"}}},
+                {"event": {"contentEnd": {"stopReason": "END_TURN"}}},
+            ],
+        )
+        done_events = [e for e in events if e["type"] == "response.done"]
+        assert len(done_events) == 1
+        usage = done_events[0]["response"]["usage"]
+        assert usage["input_tokens"] == 0
+        assert usage["output_tokens"] == 0
+        assert usage["total_tokens"] == 0
+
+    @staticmethod
+    def _usage_event(total_input, total_output, in_speech, in_text, out_speech, out_text):
+        return {
+            "event": {
+                "usageEvent": {
+                    "totalInputTokens": total_input,
+                    "totalOutputTokens": total_output,
+                    "totalTokens": total_input + total_output,
+                    "details": {
+                        "total": {
+                            "input": {"speechTokens": in_speech, "textTokens": in_text},
+                            "output": {"speechTokens": out_speech, "textTokens": out_text},
+                        }
+                    },
+                }
+            }
+        }
+
+    _ASSISTANT_TURN = (
+        {"event": {"contentStart": {"role": "ASSISTANT", "type": "TEXT"}}},
+        {"event": {"textOutput": {"content": "Hi"}}},
+        {"event": {"contentEnd": {"stopReason": "END_TURN"}}},
+    )
+
+    def test_multi_turn_usage_reports_per_response_deltas_not_cumulative_totals(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                self._usage_event(25, 40, in_speech=20, in_text=5, out_speech=30, out_text=10),
+                *self._ASSISTANT_TURN,
+                self._usage_event(40, 100, in_speech=30, in_text=10, out_speech=75, out_text=25),
+                *self._ASSISTANT_TURN,
+            ],
+        )
+        usages = [e["response"]["usage"] for e in events if e["type"] == "response.done"]
+        assert len(usages) == 2
+        assert (usages[0]["input_tokens"], usages[0]["output_tokens"], usages[0]["total_tokens"]) == (25, 40, 65)
+        assert (usages[1]["input_tokens"], usages[1]["output_tokens"], usages[1]["total_tokens"]) == (15, 60, 75)
+        assert usages[1]["input_token_details"] == {"audio_tokens": 10, "text_tokens": 5, "cached_tokens": 0}
+        assert usages[1]["output_token_details"] == {"audio_tokens": 45, "text_tokens": 15}
+        assert sum(u["total_tokens"] for u in usages) == 140
+
+    def test_usage_reported_after_last_response_done_flushes_as_logged_only_done(self):
+        config = BedrockRealtimeConfig()
+        self._run(
+            config,
+            [
+                self._usage_event(25, 40, in_speech=20, in_text=5, out_speech=30, out_text=10),
+                *self._ASSISTANT_TURN,
+            ],
+        )
+        assert config.leftover_usage_done_events() == ()
+
+        self._run(config, [self._usage_event(25, 46, in_speech=20, in_text=5, out_speech=30, out_text=16)])
+        leftover = config.leftover_usage_done_events()
+        assert len(leftover) == 1
+        assert leftover[0]["type"] == "response.done"
+        usage = leftover[0]["response"]["usage"]
+        assert (usage["input_tokens"], usage["output_tokens"], usage["total_tokens"]) == (0, 6, 6)
+        assert usage["output_token_details"] == {"audio_tokens": 0, "text_tokens": 6}
+        assert config.leftover_usage_done_events() == ()
+
+    def test_final_transcript_fragments_emit_one_completed_with_full_transcript(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {
+                    "event": {
+                        "contentStart": {
+                            "role": "USER",
+                            "type": "TEXT",
+                            "additionalModelFields": json.dumps({"generationStage": "FINAL"}),
+                        }
+                    }
+                },
+                {"event": {"textOutput": {"content": "What is the "}}},
+                {"event": {"textOutput": {"content": "capital of France?"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+            ],
+        )
+        deltas = [e for e in events if e["type"] == "conversation.item.input_audio_transcription.delta"]
+        completed = [e for e in events if e["type"] == "conversation.item.input_audio_transcription.completed"]
+        assert [d["delta"] for d in deltas] == ["What is the ", "capital of France?"]
+        assert len(completed) == 1
+        assert completed[0]["transcript"] == "What is the capital of France?"
+        assert {e["item_id"] for e in deltas + completed} == {completed[0]["item_id"]}
+
+    def test_speculative_transcript_block_end_emits_no_completed(self):
+        events = self._run(
+            BedrockRealtimeConfig(),
+            [
+                {
+                    "event": {
+                        "contentStart": {
+                            "role": "USER",
+                            "type": "TEXT",
+                            "additionalModelFields": json.dumps({"generationStage": "SPECULATIVE"}),
+                        }
+                    }
+                },
+                {"event": {"textOutput": {"content": "rea"}}},
+                {"event": {"contentEnd": {"stopReason": "PARTIAL_TURN"}}},
+            ],
+        )
+        assert [e["type"] for e in events] == ["conversation.item.input_audio_transcription.delta"]
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])

tests/test_litellm/llms/gigachat/chat/test_gigachat_chat_streaming.py

@@ -0,0 +1,87 @@
+"""
+Tests for litellm.llms.gigachat.chat.streaming
+"""
+
+from litellm.llms.gigachat.chat.streaming import GigaChatModelResponseIterator
+
+
+def _parse(chunk: dict) -> dict:
+    iterator = GigaChatModelResponseIterator(streaming_response=None, sync_stream=True)
+    return dict(iterator.chunk_parser(chunk=chunk))
+
+
+class TestChunkParserUsage:
+    def test_usage_on_stop_chunk(self):
+        parsed = _parse(
+            {
+                "choices": [{"delta": {"content": ""}, "index": 0, "finish_reason": "stop"}],
+                "usage": {"prompt_tokens": 25, "completion_tokens": 7, "total_tokens": 32},
+            }
+        )
+
+        assert parsed["finish_reason"] == "stop"
+        assert parsed["usage"] is not None
+        assert parsed["usage"]["prompt_tokens"] == 25
+        assert parsed["usage"]["completion_tokens"] == 7
+        assert parsed["usage"]["total_tokens"] == 32
+
+    def test_usage_on_function_call_chunk(self):
+        """Regression: a final chunk ending in function_call still carries usage; it must not be dropped."""
+        parsed = _parse(
+            {
+                "choices": [
+                    {
+                        "delta": {"function_call": {"name": "get_weather", "arguments": {"city": "Moscow"}}},
+                        "index": 0,
+                        "finish_reason": "function_call",
+                    }
+                ],
+                "usage": {"prompt_tokens": 40, "completion_tokens": 12, "total_tokens": 52},
+            }
+        )
+
+        assert parsed["finish_reason"] == "tool_calls"
+        assert parsed["tool_use"] is not None
+        assert parsed["usage"] is not None
+        assert parsed["usage"]["prompt_tokens"] == 40
+        assert parsed["usage"]["completion_tokens"] == 12
+        assert parsed["usage"]["total_tokens"] == 52
+
+    def test_usage_on_length_chunk(self):
+        parsed = _parse(
+            {
+                "choices": [{"delta": {"content": "truncated"}, "index": 0, "finish_reason": "length"}],
+                "usage": {"prompt_tokens": 10, "completion_tokens": 128, "total_tokens": 138},
+            }
+        )
+
+        assert parsed["usage"] is not None
+        assert parsed["usage"]["total_tokens"] == 138
+
+    def test_no_usage_on_interim_chunk(self):
+        parsed = _parse({"choices": [{"delta": {"content": "hello"}, "index": 0, "finish_reason": None}]})
+
+        assert parsed["text"] == "hello"
+        assert parsed["is_finished"] is False
+        assert parsed["usage"] is None
+
+    def test_cache_hit_usage_folds_cached_tokens_back_in(self):
+        """GigaChat reports prompt_tokens and total_tokens after subtracting cached tokens
+        (docs example: prompt_tokens=1, precached_prompt_tokens=37, total_tokens=5), so the
+        OpenAI-convention usage must add them back and surface them as cached_tokens."""
+        parsed = _parse(
+            {
+                "choices": [{"delta": {"content": ""}, "index": 0, "finish_reason": "stop"}],
+                "usage": {
+                    "prompt_tokens": 25,
+                    "completion_tokens": 7,
+                    "total_tokens": 32,
+                    "precached_prompt_tokens": 20,
+                },
+            }
+        )
+
+        assert parsed["usage"] is not None
+        assert parsed["usage"]["prompt_tokens"] == 45
+        assert parsed["usage"]["total_tokens"] == 52
+        assert parsed["usage"]["prompt_tokens_details"]["cached_tokens"] == 20

tests/test_litellm/llms/gigachat/chat/test_gigachat_chat_transformation.py

@@ -0,0 +1,883 @@
+"""
+Unit tests for GigaChat chat transformation.
+
+Tests GigaChatConfig covering get_complete_url, validate_environment,
+get_supported_openai_params, map_openai_params, _convert_tools_to_functions,
+_map_tool_choice, _transform_messages, transform_request, transform_response,
+get_model_response_iterator, and get_error_class.
+"""
+
+import json
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+
+from litellm.llms.gigachat.chat.transformation import (
+    GigaChatConfig,
+    GigaChatError,
+    is_valid_json,
+)
+from litellm.types.utils import ModelResponse, Usage
+
+TRANSFORM_MODULE = "litellm.llms.gigachat.chat.transformation"
+
+
+def _make_httpx_response(
+    body: dict, status_code: int = 200
+) -> httpx.Response:
+    return httpx.Response(
+        status_code=status_code,
+        headers={"content-type": "application/json"},
+        content=json.dumps(body).encode("utf-8"),
+        request=httpx.Request(
+            "POST",
+            "https://gigachat.devices.sberbank.ru/api/v1/chat/completions",
+        ),
+    )
+
+
+# ---------------------------------------------------------------------------
+# is_valid_json
+# ---------------------------------------------------------------------------
+
+
+class TestIsValidJson:
+    def test_valid_json_object(self):
+        assert is_valid_json('{"key": "value"}') is True
+
+    def test_valid_json_array(self):
+        assert is_valid_json("[1, 2, 3]") is True
+
+    def test_valid_json_string(self):
+        assert is_valid_json('"hello"') is True
+
+    def test_invalid_json(self):
+        assert is_valid_json("{invalid}") is False
+
+    def test_empty_string(self):
+        assert is_valid_json("") is False
+
+
+# ---------------------------------------------------------------------------
+# GigaChatConfig
+# ---------------------------------------------------------------------------
+
+
+class TestGetCompleteUrl:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_uses_api_base_from_param(self):
+        url = self.config.get_complete_url(
+            api_base="https://custom.example.com",
+            api_key=None,
+            model="GigaChat",
+            optional_params={},
+            litellm_params={},
+            stream=False,
+        )
+        assert url == "https://custom.example.com/chat/completions"
+
+    def test_uses_api_base_with_trailing_slash(self):
+        url = self.config.get_complete_url(
+            api_base="https://custom.example.com/",
+            api_key=None,
+            model="GigaChat",
+            optional_params={},
+            litellm_params={},
+            stream=False,
+        )
+        # get_api_base passes the value through without stripping the slash
+        assert url == "https://custom.example.com//chat/completions"
+
+    def test_uses_api_base_from_get_api_base_when_none(self):
+        url = self.config.get_complete_url(
+            api_base=None,
+            api_key=None,
+            model="GigaChat",
+            optional_params={},
+            litellm_params={},
+            stream=False,
+        )
+        assert url.endswith("/chat/completions")
+
+
+class TestValidateEnvironment:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{TRANSFORM_MODULE}.get_secret_str", return_value=None)
+    def test_sets_auth_headers(self, mock_get_secret, mock_get_token):
+        headers: dict = {}
+        result = self.config.validate_environment(
+            headers=headers,
+            model="GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={},
+            litellm_params={},
+            api_key="creds",
+            api_base="https://api.example.com",
+        )
+        assert result["Authorization"] == "Bearer test-token"
+        assert result["Content-Type"] == "application/json"
+        assert result["Accept"] == "application/json"
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="token")
+    @patch(f"{TRANSFORM_MODULE}.get_secret_str", return_value=None)
+    def test_stores_credentials_and_api_base_for_image_uploads(
+        self, mock_get_secret, mock_get_token
+    ):
+        self.config.validate_environment(
+            headers={},
+            model="GigaChat",
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            api_key="my-creds",
+            api_base="https://my-api.example.com",
+        )
+        assert self.config._current_credentials == "my-creds"
+        assert self.config._current_api_base == "https://my-api.example.com"
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="token")
+    @patch(f"{TRANSFORM_MODULE}.get_secret_str")
+    def test_falls_back_to_env_for_credentials(  # test-quality-ok: mock-echo of internal wiring
+        self, mock_get_secret, mock_get_token
+    ):
+        mock_get_secret.return_value = "env-creds"
+        self.config.validate_environment(
+            headers={},
+            model="GigaChat",
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            api_key=None,
+            api_base=None,
+        )
+        mock_get_secret.assert_any_call("GIGACHAT_CREDENTIALS")  # test-quality-ok: mock-echo of internal wiring
+
+
+class TestGetSupportedOpenAiParams:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_returns_expected_params(self):
+        params = self.config.get_supported_openai_params("GigaChat")
+        expected = [
+            "stream",
+            "temperature",
+            "top_p",
+            "max_tokens",
+            "max_completion_tokens",
+            "stop",
+            "tools",
+            "tool_choice",
+            "functions",
+            "function_call",
+            "response_format",
+        ]
+        assert params == expected
+
+
+class TestMapOpenAiParams:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_stream(self):
+        result = self.config.map_openai_params(
+            non_default_params={"stream": True},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["stream"] is True
+
+    def test_temperature_zero_maps_to_top_p_zero(self):
+        result = self.config.map_openai_params(
+            non_default_params={"temperature": 0},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["top_p"] == 0
+        assert "temperature" not in result
+
+    def test_temperature_non_zero(self):
+        result = self.config.map_openai_params(
+            non_default_params={"temperature": 0.7},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["temperature"] == 0.7
+
+    def test_top_p(self):
+        result = self.config.map_openai_params(
+            non_default_params={"top_p": 0.5},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["top_p"] == 0.5
+
+    def test_max_tokens(self):
+        result = self.config.map_openai_params(
+            non_default_params={"max_tokens": 100},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["max_tokens"] == 100
+
+    def test_max_completion_tokens(self):
+        result = self.config.map_openai_params(
+            non_default_params={"max_completion_tokens": 200},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["max_tokens"] == 200
+
+    def test_stop_is_dropped(self):
+        result = self.config.map_openai_params(
+            non_default_params={"stop": ["\n\n"]},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert "stop" not in result
+
+    def test_tools_converted_to_functions(self):
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "get_weather",
+                    "description": "Get weather",
+                    "parameters": {"type": "object"},
+                },
+            }
+        ]
+        result = self.config.map_openai_params(
+            non_default_params={"tools": tools},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert "functions" in result
+        assert result["functions"] == [
+            {"name": "get_weather", "description": "Get weather", "parameters": {"type": "object"}}
+        ]
+
+    def test_tool_choice_auto(self):
+        result = self.config.map_openai_params(
+            non_default_params={"tool_choice": "auto"},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result.get("function_call") == "auto"
+
+    def test_tool_choice_none(self):
+        result = self.config.map_openai_params(
+            non_default_params={"tool_choice": "none"},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result.get("function_call") == "none"
+
+    def test_tool_choice_required(self):
+        result = self.config.map_openai_params(
+            non_default_params={"tool_choice": "required"},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result.get("function_call") == "auto"
+
+    def test_tool_choice_dict(self):
+        result = self.config.map_openai_params(
+            non_default_params={
+                "tool_choice": {
+                    "type": "function",
+                    "function": {"name": "get_weather"},
+                }
+            },
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result.get("function_call") == {"name": "get_weather"}
+
+    def test_functions(self):
+        funcs = [{"name": "my_func", "description": "desc", "parameters": {}}]
+        result = self.config.map_openai_params(
+            non_default_params={"functions": funcs},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["functions"] == funcs
+
+    def test_function_call(self):
+        result = self.config.map_openai_params(
+            non_default_params={"function_call": {"name": "my_func"}},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result["function_call"] == {"name": "my_func"}
+
+    def test_response_format_json_schema(self):
+        response_format = {
+            "type": "json_schema",
+            "json_schema": {
+                "name": "test_schema",
+                "schema": {"type": "object", "properties": {"name": {"type": "string"}}},
+            },
+        }
+        result = self.config.map_openai_params(
+            non_default_params={"response_format": response_format},
+            optional_params={"functions": []},
+            model="GigaChat",
+            drop_params=False,
+        )
+        # Should add a function for the schema
+        assert len(result["functions"]) == 1
+        assert result["functions"][0]["name"] == "test_schema"
+        assert result["function_call"] == {"name": "test_schema"}
+        assert result["_structured_output"] is True
+
+
+class TestConvertToolsToFunctions:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_converts_function_tools_only(self):
+        tools = [
+            {"type": "function", "function": {"name": "a", "description": "d", "parameters": {}}},
+            {"type": "code_interpreter"},  # should be ignored
+        ]
+        result = self.config._convert_tools_to_functions(tools)
+        assert len(result) == 1
+        assert result[0]["name"] == "a"
+
+    def test_empty_tools(self):
+        assert self.config._convert_tools_to_functions([]) == []
+
+
+class TestMapToolChoice:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_none(self):
+        assert self.config._map_tool_choice("none") == "none"
+
+    def test_auto(self):
+        assert self.config._map_tool_choice("auto") == "auto"
+
+    def test_required(self):
+        assert self.config._map_tool_choice("required") == "auto"
+
+    def test_dict_with_function(self):
+        result = self.config._map_tool_choice(
+            {"type": "function", "function": {"name": "get_weather"}}
+        )
+        assert result == {"name": "get_weather"}
+
+    def test_dict_without_name(self):
+        result = self.config._map_tool_choice(
+            {"type": "function", "function": {}}
+        )
+        assert result is None
+
+    def test_unknown_value(self):
+        assert self.config._map_tool_choice("unknown") is None
+
+
+class TestTransformMessages:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_developer_role_to_system(self):
+        result = self.config._transform_messages(
+            [{"role": "developer", "content": "be helpful"}]
+        )
+        assert result[0]["role"] == "system"
+        assert result[0]["content"] == "be helpful"
+
+    def test_system_message_not_first_becomes_user(self):
+        result = self.config._transform_messages([
+            {"role": "user", "content": "hi"},
+            {"role": "system", "content": "instruction"},
+        ])
+        assert result[0]["role"] == "user"
+        assert result[1]["role"] == "user"
+        assert result[1]["content"] == "instruction"
+
+    def test_tool_role_to_function(self):
+        result = self.config._transform_messages([
+            {"role": "tool", "content": '{"result": "ok"}'}
+        ])
+        assert result[0]["role"] == "function"
+
+    def test_tool_role_content_wraps_non_json(self):
+        result = self.config._transform_messages([
+            {"role": "tool", "content": "plain text"}
+        ])
+        assert result[0]["role"] == "function"
+        assert is_valid_json(result[0]["content"])
+
+    def test_none_content_becomes_empty_string(self):
+        result = self.config._transform_messages([
+            {"role": "user", "content": None}
+        ])
+        assert result[0]["content"] == ""
+
+    def test_name_field_removed(self):
+        result = self.config._transform_messages([
+            {"role": "user", "content": "hi", "name": "John"}
+        ])
+        assert "name" not in result[0]
+
+    def test_tool_calls_converted_to_function_call(self):
+        result = self.config._transform_messages([
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {
+                        "id": "call_abc",
+                        "type": "function",
+                        "function": {
+                            "name": "get_weather",
+                            "arguments": '{"city": "London"}',
+                        },
+                    }
+                ],
+            }
+        ])
+        assert "tool_calls" not in result[0]
+        assert result[0]["function_call"]["name"] == "get_weather"
+        assert result[0]["function_call"]["arguments"] == {"city": "London"}
+
+    def test_tool_calls_with_dict_arguments(self):
+        result = self.config._transform_messages([
+            {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {
+                        "id": "call_xyz",
+                        "type": "function",
+                        "function": {
+                            "name": "search",
+                            "arguments": {"query": "test"},
+                        },
+                    }
+                ],
+            }
+        ])
+        assert result[0]["function_call"]["arguments"] == {"query": "test"}
+
+    def test_list_content_multimodal(self):
+        content = [
+            {"type": "text", "text": "describe this"},
+            {
+                "type": "image_url",
+                "image_url": {"url": "https://example.com/img.jpg"},
+            },
+        ]
+        with patch.object(self.config, "_upload_image", return_value="file-123"):
+            result = self.config._transform_messages([
+                {"role": "user", "content": content}
+            ])
+        assert result[0]["content"] == "describe this"
+        assert result[0]["attachments"] == ["file-123"]
+
+    def test_list_content_with_image_url_string(self):
+        content = [
+            {"type": "text", "text": "look"},
+            {"type": "image_url", "image_url": "https://example.com/img.jpg"},
+        ]
+        with patch.object(self.config, "_upload_image", return_value="file-456"):
+            result = self.config._transform_messages([
+                {"role": "user", "content": content}
+            ])
+        assert result[0]["content"] == "look"
+        assert "file-456" in result[0]["attachments"]
+
+
+class TestTransformRequest:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_builds_basic_request(self):
+        body = self.config.transform_request(
+            model="gigachat/GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={},
+            litellm_params={},
+            headers={},
+        )
+        assert body["model"] == "GigaChat"
+        assert len(body["messages"]) == 1
+        assert body["messages"][0]["content"] == "hi"
+
+    def test_model_prefix_stripped(self):
+        body = self.config.transform_request(
+            model="gigachat/GigaChat-Pro",
+            messages=[{"role": "user", "content": "hello"}],
+            optional_params={},
+            litellm_params={},
+            headers={},
+        )
+        assert body["model"] == "GigaChat-Pro"
+
+    def test_includes_optional_params(self):
+        body = self.config.transform_request(
+            model="gigachat/GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={
+                "temperature": 0.5,
+                "max_tokens": 100,
+                "stream": True,
+            },
+            litellm_params={},
+            headers={},
+        )
+        assert body["temperature"] == 0.5
+        assert body["max_tokens"] == 100
+        assert body["stream"] is True
+
+    def test_includes_functions(self):
+        body = self.config.transform_request(
+            model="gigachat/GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={
+                "functions": [{"name": "my_func"}],
+                "function_call": {"name": "my_func"},
+            },
+            litellm_params={},
+            headers={},
+        )
+        assert body["functions"] == [{"name": "my_func"}]
+        assert body["function_call"] == {"name": "my_func"}
+
+    def test_skips_unsupported_params(self):
+        body = self.config.transform_request(
+            model="gigachat/GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={"n": 2, "user": "abc"},
+            litellm_params={},
+            headers={},
+        )
+        assert "n" not in body
+        assert "user" not in body
+
+
+class TestTransformResponse:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_basic_response(self):
+        raw = _make_httpx_response({
+            "id": "chatcmpl-123",
+            "created": 1700000000,
+            "model": "GigaChat",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": "Hello!"},
+                    "finish_reason": "stop",
+                }
+            ],
+            "usage": {"prompt_tokens": 5, "completion_tokens": 3, "total_tokens": 8},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        assert result.choices[0].message.content == "Hello!"
+        assert result.choices[0].finish_reason == "stop"
+        assert result.usage.prompt_tokens == 5
+        assert result.usage.total_tokens == 8
+
+    def test_function_call_into_tool_calls(self):
+        raw = _make_httpx_response({
+            "id": "chatcmpl-456",
+            "created": 1700000000,
+            "model": "GigaChat",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": None,
+                        "function_call": {
+                            "name": "get_weather",
+                            "arguments": {"city": "Moscow"},
+                        },
+                    },
+                    "finish_reason": "function_call",
+                }
+            ],
+            "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        assert result.choices[0].finish_reason == "tool_calls"
+        tool_calls = result.choices[0].message.tool_calls
+        assert tool_calls is not None
+        assert len(tool_calls) == 1
+        assert tool_calls[0].function.name == "get_weather"
+        assert '{"city": "Moscow"}' in tool_calls[0].function.arguments
+
+    def test_function_call_structured_output(self):
+        raw = _make_httpx_response({
+            "id": "chatcmpl-789",
+            "created": 1700000000,
+            "model": "GigaChat",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": None,
+                        "function_call": {
+                            "name": "test_schema",
+                            "arguments": {"name": "John"},
+                        },
+                    },
+                    "finish_reason": "function_call",
+                }
+            ],
+            "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={"_structured_output": True},
+            litellm_params={},
+            encoding=None,
+        )
+        # Structured output: function_call -> content
+        assert result.choices[0].finish_reason == "stop"
+        assert result.choices[0].message.content is not None
+        assert '"name": "John"' in result.choices[0].message.content
+
+    def test_function_call_string_arguments(self):
+        raw = _make_httpx_response({
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "function_call": {
+                            "name": "get_weather",
+                            "arguments": '{"city": "Moscow"}',
+                        },
+                    },
+                    "finish_reason": "function_call",
+                }
+            ],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        tc = result.choices[0].message.tool_calls[0]
+        assert '{"city": "Moscow"}' in tc.function.arguments
+
+    def test_cleans_up_gigachat_specific_fields(self):
+        raw = _make_httpx_response({
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": "done",
+                        "functions_state_id": "some-state",
+                    },
+                    "finish_reason": "stop",
+                }
+            ],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        # functions_state_id should have been removed from the message data
+        assert result.choices[0].message.content == "done"
+
+    def test_raises_on_invalid_json(self):
+        raw = httpx.Response(
+            status_code=500,
+            headers={"content-type": "text/plain"},
+            content=b"not json",
+            request=httpx.Request("POST", "https://example.com"),
+        )
+        model_response = ModelResponse()
+        with pytest.raises(GigaChatError) as exc_info:
+            self.config.transform_response(
+                model="gigachat/GigaChat",
+                raw_response=raw,
+                model_response=model_response,
+                logging_obj=MagicMock(),
+                request_data={},
+                messages=[],
+                optional_params={},
+                litellm_params={},
+                encoding=None,
+            )
+        assert "Invalid JSON response" in str(exc_info.value.message)
+
+    def test_empty_choices(self):
+        raw = _make_httpx_response({
+            "choices": [],
+            "usage": {},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        assert result.choices == []
+
+    def test_function_call_with_non_dict_arguments(self):
+        raw = _make_httpx_response({
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "function_call": {
+                            "name": "say_hello",
+                            "arguments": "hello",
+                        },
+                    },
+                    "finish_reason": "function_call",
+                }
+            ],
+            "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+        })
+        model_response = ModelResponse()
+        result = self.config.transform_response(
+            model="gigachat/GigaChat",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=MagicMock(),
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+        tc = result.choices[0].message.tool_calls[0]
+        assert tc.function.arguments == "hello"
+
+
+class TestGetModelResponseIterator:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_returns_gigachat_iterator_sync(self):
+        from litellm.llms.gigachat.chat.streaming import (
+            GigaChatModelResponseIterator,
+        )
+
+        result = self.config.get_model_response_iterator(
+            streaming_response=iter(["data"]),
+            sync_stream=True,
+            json_mode=False,
+        )
+        assert isinstance(result, GigaChatModelResponseIterator)
+
+
+class TestGetErrorClass:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    def test_returns_gigachat_error(self):
+        error = self.config.get_error_class(
+            error_message="something went wrong",
+            status_code=400,
+            headers={"x-request-id": "abc"},
+        )
+        assert isinstance(error, GigaChatError)
+        assert error.status_code == 400
+        assert error.message == "something went wrong"
+        assert error.headers == {"x-request-id": "abc"}
+
+
+class TestUploadImage:
+    def setup_method(self):
+        self.config = GigaChatConfig()
+
+    @patch(f"{TRANSFORM_MODULE}.upload_file_sync", return_value="file-uploaded")
+    def test_upload_image_success(self, mock_upload):
+        self.config._current_credentials = "creds"
+        self.config._current_api_base = "https://api.example.com"
+        result = self.config._upload_image("https://example.com/img.jpg")
+        assert result == "file-uploaded"
+        mock_upload.assert_called_once_with(
+            image_url="https://example.com/img.jpg",
+            credentials="creds",
+            api_base="https://api.example.com",
+        )
+
+    @patch(f"{TRANSFORM_MODULE}.upload_file_sync", side_effect=Exception("fail"))
+    def test_upload_image_failure_returns_none(self, mock_upload):
+        result = self.config._upload_image("https://example.com/img.jpg")
+        assert result is None

tests/test_litellm/llms/gigachat/embedding/test_gigachat_embedding_transformation.py

@@ -0,0 +1,372 @@
+"""
+Unit tests for GigaChat embedding transformation.
+
+Tests GigaChatEmbeddingConfig covering get_config, get_supported_openai_params,
+map_openai_params, _get_openai_compatible_provider_info, get_complete_url,
+transform_embedding_request, transform_embedding_response, validate_environment,
+and get_error_class.
+"""
+
+import json
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+
+from litellm import LlmProviders
+from litellm.llms.gigachat.embedding.transformation import (
+    GigaChatEmbeddingConfig,
+    GigaChatEmbeddingError,
+)
+from litellm.types.utils import EmbeddingResponse
+
+TRANSFORM_MODULE = "litellm.llms.gigachat.embedding.transformation"
+
+
+def _make_httpx_response(body: dict, status_code: int = 200) -> httpx.Response:
+    return httpx.Response(
+        status_code=status_code,
+        headers={"content-type": "application/json"},
+        content=json.dumps(body).encode("utf-8"),
+        request=httpx.Request("POST", "https://gigachat.devices.sberbank.ru/api/v1/embeddings"),
+    )
+
+
+# ---------------------------------------------------------------------------
+# GigaChatEmbeddingConfig
+# ---------------------------------------------------------------------------
+
+
+class TestGetConfig:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_contains_only_abc_impl(self):
+        """get_config returns ABC internal data due to inheritance."""
+        result = self.config.get_config()
+        # The only key should be _abc_impl from ABC base class
+        assert set(result.keys()) == {"_abc_impl"}
+
+
+class TestGetSupportedOpenAiParams:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_returns_empty_list(self):
+        params = self.config.get_supported_openai_params("GigaChat")
+        assert params == []
+
+
+class TestMapOpenAiParams:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_returns_optional_params_unchanged(self):
+        result = self.config.map_openai_params(
+            non_default_params={"model": "test"},
+            optional_params={"temperature": 0.5},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result == {"temperature": 0.5}
+
+    def test_returns_empty_dict_when_no_optional_params(self):
+        result = self.config.map_openai_params(
+            non_default_params={},
+            optional_params={},
+            model="GigaChat",
+            drop_params=False,
+        )
+        assert result == {}
+
+
+class TestGetOpenaiCompatibleProviderInfo:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_returns_gigachat_provider(self):
+        provider, api_base, api_key = self.config._get_openai_compatible_provider_info(
+            api_base="https://api.example.com", api_key="test-key"
+        )
+        assert provider == LlmProviders.GIGACHAT.value
+        assert api_base == "https://api.example.com"
+        assert api_key == "test-key"
+
+    def test_resolves_api_base_when_none(self, monkeypatch):
+        monkeypatch.delenv("GIGACHAT_API_BASE", raising=False)
+        provider, api_base, api_key = self.config._get_openai_compatible_provider_info(
+            api_base=None, api_key="key"
+        )
+        assert api_base is not None
+        assert api_base.endswith("/api/v1")
+
+    def test_returns_none_api_key(self):
+        _, _, api_key = self.config._get_openai_compatible_provider_info(
+            api_base="https://example.com", api_key=None
+        )
+        assert api_key is None
+
+
+class TestGetCompleteUrl:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_default_url(self):
+        url = self.config.get_complete_url(
+            api_base=None, api_key=None, model="GigaChat",
+            optional_params={}, litellm_params={},
+        )
+        assert url.endswith("/embeddings")
+
+    def test_custom_api_base(self):
+        url = self.config.get_complete_url(
+            api_base="https://custom.example.com", api_key=None, model="GigaChat",
+            optional_params={}, litellm_params={},
+        )
+        assert url == "https://custom.example.com/embeddings"
+
+    def test_trailing_slash_api_base(self):
+        url = self.config.get_complete_url(
+            api_base="https://custom.example.com/", api_key=None, model="GigaChat",
+            optional_params={}, litellm_params={},
+        )
+        # get_api_base doesn't strip slash, so we get double slash
+        assert url == "https://custom.example.com//embeddings"
+
+
+class TestTransformEmbeddingRequest:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_string_input(self):
+        result = self.config.transform_embedding_request(
+            model="gigachat/Embeddings",
+            input="hello world",
+            optional_params={},
+            headers={},
+        )
+        assert result == {"model": "Embeddings", "input": ["hello world"]}
+
+    def test_list_input(self):
+        result = self.config.transform_embedding_request(
+            model="gigachat/Embeddings",
+            input=["text1", "text2"],
+            optional_params={},
+            headers={},
+        )
+        assert result == {"model": "Embeddings", "input": ["text1", "text2"]}
+
+    def test_strips_gigachat_prefix(self):
+        result = self.config.transform_embedding_request(
+            model="gigachat/GigaChat-Pro",
+            input="test",
+            optional_params={},
+            headers={},
+        )
+        assert result["model"] == "GigaChat-Pro"
+
+    def test_model_without_prefix(self):
+        result = self.config.transform_embedding_request(
+            model="Embeddings",
+            input="test",
+            optional_params={},
+            headers={},
+        )
+        assert result["model"] == "Embeddings"
+
+
+class TestTransformEmbeddingResponse:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+        self.logging_obj = MagicMock()
+
+    def _make_gigachat_response(self, data: list[dict]) -> httpx.Response:
+        return _make_httpx_response({
+            "object": "list",
+            "data": data,
+            "model": "Embeddings",
+        })
+
+    def test_basic_response(self):
+        raw = self._make_gigachat_response([
+            {
+                "object": "embedding",
+                "embedding": [0.1, 0.2, 0.3],
+                "index": 0,
+            }
+        ])
+        model_response = EmbeddingResponse()
+        result = self.config.transform_embedding_response(
+            model="gigachat/Embeddings",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=self.logging_obj,
+            api_key="test-key",
+            request_data={"input": ["text"]},
+            optional_params={},
+            litellm_params={},
+        )
+        assert result.object == "list"
+        assert len(result.data) == 1
+        assert result.data[0]["embedding"] == [0.1, 0.2, 0.3]
+        assert result.data[0]["index"] == 0
+        assert result.usage.prompt_tokens == 0
+        assert result.usage.total_tokens == 0
+
+    def test_aggregates_per_embedding_usage(self):
+        raw = self._make_gigachat_response([
+            {
+                "object": "embedding",
+                "embedding": [0.1, 0.2],
+                "index": 0,
+                "usage": {"prompt_tokens": 5},
+            },
+            {
+                "object": "embedding",
+                "embedding": [0.3, 0.4],
+                "index": 1,
+                "usage": {"prompt_tokens": 7},
+            },
+        ])
+        model_response = EmbeddingResponse()
+        result = self.config.transform_embedding_response(
+            model="gigachat/Embeddings",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=self.logging_obj,
+            api_key="test-key",
+            request_data={"input": ["a", "b"]},
+            optional_params={},
+            litellm_params={},
+        )
+        # Total should be sum of per-embedding prompt_tokens
+        assert result.usage.prompt_tokens == 12
+        assert result.usage.total_tokens == 12
+        # Usage should be removed from individual embedding data
+        assert "usage" not in result.data[0]
+        assert "usage" not in result.data[1]
+
+    def test_usage_removed_from_individual_embeddings(self):
+        raw = self._make_gigachat_response([
+            {
+                "object": "embedding",
+                "embedding": [0.5],
+                "index": 0,
+                "usage": {"prompt_tokens": 3},
+            }
+        ])
+        model_response = EmbeddingResponse()
+        result = self.config.transform_embedding_response(
+            model="gigachat/Embeddings",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=self.logging_obj,
+            api_key="key",
+            request_data={"input": ["x"]},
+            optional_params={},
+            litellm_params={},
+        )
+        # usage should NOT be in the final EmbeddingResponse data items
+        for emb in result.data:
+            assert "usage" not in emb
+
+    def test_passes_model_from_response(self):
+        raw = self._make_gigachat_response([
+            {"object": "embedding", "embedding": [0.1], "index": 0},
+        ])
+        model_response = EmbeddingResponse()
+        result = self.config.transform_embedding_response(
+            model="gigachat/Embeddings",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=self.logging_obj,
+            api_key="key",
+            request_data={"input": ["x"]},
+            optional_params={},
+            litellm_params={},
+        )
+        assert result.model == "Embeddings"
+
+    def test_calls_logging_post_call(self):
+        raw = self._make_gigachat_response([
+            {"object": "embedding", "embedding": [0.1], "index": 0},
+        ])
+        model_response = EmbeddingResponse()
+        self.config.transform_embedding_response(
+            model="gigachat/Embeddings",
+            raw_response=raw,
+            model_response=model_response,
+            logging_obj=self.logging_obj,
+            api_key="test-api-key",
+            request_data={"input": ["hello"]},
+            optional_params={},
+            litellm_params={},
+        )
+        self.logging_obj.post_call.assert_called_once()
+        args = self.logging_obj.post_call.call_args.kwargs
+        assert args["api_key"] == "test-api-key"
+        assert args["input"] == ["hello"]
+
+
+class TestValidateEnvironment:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="test-token")
+    def test_sets_oauth_headers(self, mock_get_token):
+        headers = self.config.validate_environment(
+            headers={},
+            model="GigaChat",
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            api_key="creds",
+            api_base="https://api.example.com",
+        )
+        assert headers["Authorization"] == "Bearer test-token"
+        assert headers["Content-Type"] == "application/json"
+        mock_get_token.assert_called_once_with(credentials="creds", litellm_params={})
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="token")
+    def test_merges_custom_headers(self, mock_get_token):
+        headers = self.config.validate_environment(
+            headers={"X-Custom": "value"},
+            model="GigaChat",
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            api_key="creds",
+            api_base="https://api.example.com",
+        )
+        assert headers["Authorization"] == "Bearer token"
+        assert headers["Content-Type"] == "application/json"
+        assert headers["X-Custom"] == "value"
+
+    @patch(f"{TRANSFORM_MODULE}.get_access_token", return_value="token")
+    def test_custom_header_overwrites_default(self, mock_get_token):
+        headers = self.config.validate_environment(
+            headers={"Authorization": "Bearer custom"},
+            model="GigaChat",
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            api_key="creds",
+            api_base="https://api.example.com",
+        )
+        # Merge: default headers first, then custom headers on top
+        assert headers["Authorization"] == "Bearer custom"
+
+
+class TestGetErrorClass:
+    def setup_method(self):
+        self.config = GigaChatEmbeddingConfig()
+
+    def test_returns_gigachat_embedding_error(self):
+        error = self.config.get_error_class(
+            error_message="embedding failed",
+            status_code=400,
+            headers={"x-request-id": "abc"},
+        )
+        assert isinstance(error, GigaChatEmbeddingError)
+        assert error.status_code == 400
+        assert error.message == "embedding failed"

tests/test_litellm/llms/gigachat/passthrough/test_gigachat_passthrough_transformation.py

@@ -0,0 +1,607 @@
+"""
+Unit tests for GigaChatPassthroughConfig transformation.
+
+Tests the GigaChat-specific passthrough configuration including URL construction,
+streaming detection, authentication handling, and logging response transformations.
+"""
+
+import json
+from unittest.mock import MagicMock, patch
+
+import httpx
+import pytest
+
+from litellm.llms.gigachat.passthrough.transformation import GigaChatPassthroughConfig
+from litellm.types.utils import EmbeddingResponse, ModelResponse
+
+
+def _gigachat_chat_completion_body():
+    return {
+        "id": "chatcmpl-test123",
+        "object": "chat.completion",
+        "created": 1700000000,
+        "model": "GigaChat",
+        "choices": [
+            {
+                "index": 0,
+                "message": {
+                    "role": "assistant",
+                    "content": "Hello from GigaChat",
+                },
+                "finish_reason": "stop",
+            }
+        ],
+        "usage": {
+            "prompt_tokens": 5,
+            "completion_tokens": 3,
+            "total_tokens": 8,
+        },
+    }
+
+
+def _gigachat_embedding_body():
+    return {
+        "object": "list",
+        "data": [
+            {
+                "object": "embedding",
+                "embedding": [0.1, 0.2, 0.3],
+                "index": 0,
+                "usage": {"prompt_tokens": 4},
+            }
+        ],
+        "model": "Embeddings",
+    }
+
+
+def _make_httpx_response(body: dict) -> httpx.Response:
+    return httpx.Response(
+        status_code=200,
+        headers={"content-type": "application/json"},
+        content=json.dumps(body).encode("utf-8"),
+        request=httpx.Request(
+            "POST", "https://gigachat.devices.sberbank.ru/api/v1/chat/completions"
+        ),
+    )
+
+
+class TestGigaChatPassthroughConfig:
+    """Tests for GigaChatPassthroughConfig class."""
+
+    def test_is_streaming_request_true(self):
+        """Test streaming is detected when stream=True."""
+        config = GigaChatPassthroughConfig()
+        assert (
+            config.is_streaming_request("chat/completions", {"stream": True}) is True
+        )
+
+    def test_is_streaming_request_false(self):
+        """Test streaming is not detected when stream=False."""
+        config = GigaChatPassthroughConfig()
+        assert (
+            config.is_streaming_request("chat/completions", {"stream": False})
+            is False
+        )
+
+    def test_is_streaming_request_missing_stream_key(self):
+        """Test streaming defaults to False when stream key is missing."""
+        config = GigaChatPassthroughConfig()
+        assert (
+            config.is_streaming_request("chat/completions", {"model": "GigaChat"})
+            is False
+        )
+
+    def test_get_complete_url_with_api_base(self):
+        """Test URL construction with explicit api_base."""
+        config = GigaChatPassthroughConfig()
+        api_base = "https://custom.gigachat.ru/api/v1"
+        endpoint = "chat/completions"
+
+        complete_url, base_target_url = config.get_complete_url(
+            api_base=api_base,
+            api_key=None,
+            model="GigaChat",
+            endpoint=endpoint,
+            request_query_params=None,
+            litellm_params={},
+        )
+
+        assert isinstance(complete_url, httpx.URL)
+        assert str(complete_url) == f"{api_base}/{endpoint}"
+        assert base_target_url == api_base
+
+    def test_get_complete_url_with_leading_slash_endpoint(self):
+        """Test URL construction with endpoint having leading slash."""
+        config = GigaChatPassthroughConfig()
+        api_base = "https://custom.gigachat.ru/api/v1"
+        endpoint = "/chat/completions"
+
+        complete_url, base_target_url = config.get_complete_url(
+            api_base=api_base,
+            api_key=None,
+            model="GigaChat",
+            endpoint=endpoint,
+            request_query_params=None,
+            litellm_params={},
+        )
+
+        assert str(complete_url) == "https://custom.gigachat.ru/api/v1/chat/completions"
+        assert base_target_url == api_base
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_complete_url_with_env_api_base(self, mock_get_secret):
+        """Test URL construction with api_base from environment."""
+        config = GigaChatPassthroughConfig()
+        env_api_base = "https://env.gigachat.ru/api/v1"
+        mock_get_secret.return_value = env_api_base
+
+        complete_url, base_target_url = config.get_complete_url(
+            api_base=None,
+            api_key=None,
+            model="GigaChat",
+            endpoint="embeddings",
+            request_query_params=None,
+            litellm_params={},
+        )
+
+        assert isinstance(complete_url, httpx.URL)
+        assert str(complete_url).startswith(env_api_base)
+        assert base_target_url == env_api_base
+        mock_get_secret.assert_called_once_with("GIGACHAT_API_BASE")
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_complete_url_fallback_to_default(self, mock_get_secret):
+        """Test URL construction falls back to default GIGACHAT_BASE_URL."""
+        config = GigaChatPassthroughConfig()
+        mock_get_secret.return_value = None
+
+        complete_url, base_target_url = config.get_complete_url(
+            api_base=None,
+            api_key=None,
+            model="GigaChat",
+            endpoint="models",
+            request_query_params=None,
+            litellm_params={},
+        )
+
+        assert isinstance(complete_url, httpx.URL)
+        assert "gigachat.devices.sberbank.ru" in str(complete_url)
+        assert base_target_url == "https://gigachat.devices.sberbank.ru/api/v1"
+
+    def test_get_complete_url_no_api_base_raises(self):
+        """Test that exception is raised when no api_base can be resolved."""
+        config = GigaChatPassthroughConfig()
+        with patch(
+            "litellm.llms.gigachat.passthrough.transformation.get_secret_str",  # test-quality-ok: patching litellm internal for unit test isolation
+            return_value=None,
+        ):
+            with patch(
+                "litellm.llms.gigachat.passthrough.transformation.GIGACHAT_BASE_URL",  # test-quality-ok: patching litellm internal for unit test isolation
+                None,
+            ):
+                with pytest.raises(Exception, match="GigaChat api base not found"):
+                    config.get_complete_url(
+                        api_base=None,
+                        api_key=None,
+                        model="GigaChat",
+                        endpoint="chat/completions",
+                        request_query_params=None,
+                        litellm_params={},
+                    )
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_access_token"
+    )
+    def test_validate_environment(self, mock_get_access_token):
+        """Test headers are set correctly with OAuth token."""
+        config = GigaChatPassthroughConfig()
+        mock_get_access_token.return_value = "test-token-123"
+
+        headers = config.validate_environment(
+            headers={},
+            model="GigaChat",
+            messages=[{"role": "user", "content": "hi"}],
+            optional_params={},
+            litellm_params={},
+            api_key="test-credentials",
+            api_base="https://custom.gigachat.ru",
+        )
+
+        assert headers["Authorization"] == "Bearer test-token-123"
+        assert headers["Content-Type"] == "application/json"
+        assert headers["Accept"] == "application/json"
+        mock_get_access_token.assert_called_once_with(
+            credentials="test-credentials",
+            litellm_params={},
+        )
+
+    def test_logging_non_streaming_response_chat_completions(self):
+        """Test chat completions endpoint returns ModelResponse."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        result = config.logging_non_streaming_response(
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            httpx_response=_make_httpx_response(_gigachat_chat_completion_body()),
+            request_data={
+                "model": "gigachat/GigaChat",
+                "messages": [{"role": "user", "content": "hi"}],
+            },
+            logging_obj=logging_obj,
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "Hello from GigaChat"
+        assert result.usage.prompt_tokens == 5
+        assert result.usage.completion_tokens == 3
+        assert result.usage.total_tokens == 8
+
+    def test_logging_non_streaming_response_embeddings(self):
+        """Test embeddings endpoint returns EmbeddingResponse."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        result = config.logging_non_streaming_response(
+            model="gigachat/Embeddings",
+            custom_llm_provider="gigachat",
+            httpx_response=_make_httpx_response(_gigachat_embedding_body()),
+            request_data={"input": ["hello"], "model": "gigachat/Embeddings"},
+            logging_obj=logging_obj,
+            endpoint="embeddings",
+        )
+
+        assert isinstance(result, EmbeddingResponse)
+        assert len(result.data) == 1
+        assert result.data[0]["embedding"] == [0.1, 0.2, 0.3]
+
+    def test_logging_non_streaming_response_unknown_endpoint_returns_none(self):
+        """Test unknown endpoint returns None."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        result = config.logging_non_streaming_response(
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            httpx_response=_make_httpx_response(_gigachat_chat_completion_body()),
+            request_data={},
+            logging_obj=logging_obj,
+            endpoint="images/generations",
+        )
+
+        assert result is None
+
+    def test_handle_logging_collected_chunks_with_string_chunks(self):
+        """Test converting string chunks to model response."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            '{"choices": [{"delta": {"content": "Hello"}, "index": 0}]}',
+            '{"choices": [{"delta": {"content": " world"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 5, "completion_tokens": 2, "total_tokens": 7}}',
+        ]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "Hello world"
+
+    def test_handle_logging_collected_chunks_with_bytes_chunks(self):
+        """Test converting string chunks to model response (bytes pre-decoded upstream)."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            '{"choices": [{"delta": {"content": "Hi"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}}',
+        ]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "Hi"
+
+    def test_handle_logging_collected_chunks_with_done_and_empty(self):
+        """Test that [DONE] and empty chunks are skipped."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            "",
+            "[DONE]",
+            '{"choices": [{"delta": {"content": "test"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}}',
+        ]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "test"
+
+    def test_handle_logging_collected_chunks_with_dict_chunks(self):
+        """Test converting string-serialized dict chunks (dicts pre-serialized upstream)."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            '{"choices": [{"delta": {"content": "direct"}, "index": 0}]}',
+            json.dumps(
+                {
+                    "choices": [
+                        {
+                            "delta": {},
+                            "finish_reason": "stop",
+                            "index": 0,
+                        }
+                    ],
+                    "usage": {
+                        "prompt_tokens": 1,
+                        "completion_tokens": 1,
+                        "total_tokens": 2,
+                    },
+                }
+            ),
+        ]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "direct"
+
+    def test_handle_logging_collected_chunks_empty_list_returns_none(self):
+        """Test empty chunks list returns None."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=[],
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert result is None
+
+    def test_handle_logging_collected_chunks_invalid_json_skipped(self):
+        """Test invalid JSON chunks are skipped gracefully."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            "not-valid-json",
+            '{"choices": [{"delta": {"content": "valid"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}}',
+        ]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "valid"
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_api_base_with_explicit_value(self, mock_get_secret):
+        """Test get_api_base returns explicit value when provided."""
+        explicit_base = "https://custom.gigachat.ru/api/v1"
+        result = GigaChatPassthroughConfig.get_api_base(api_base=explicit_base)
+        assert result == explicit_base
+        mock_get_secret.assert_not_called()
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_api_base_from_environment(self, mock_get_secret):
+        """Test get_api_base retrieves from environment when not provided."""
+        env_base = "https://env.gigachat.ru/api/v1"
+        mock_get_secret.return_value = env_base
+        result = GigaChatPassthroughConfig.get_api_base(api_base=None)
+        assert result == env_base
+        mock_get_secret.assert_called_once_with("GIGACHAT_API_BASE")
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_api_base_fallback_to_default(self, mock_get_secret):
+        """Test get_api_base falls back to GIGACHAT_BASE_URL."""
+        mock_get_secret.return_value = None
+        result = GigaChatPassthroughConfig.get_api_base(api_base=None)
+        assert result == "https://gigachat.devices.sberbank.ru/api/v1"
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_api_key_with_explicit_value(self, mock_get_secret):
+        """Test get_api_key returns explicit value when provided."""
+        explicit_key = "test-api-key"
+        result = GigaChatPassthroughConfig.get_api_key(api_key=explicit_key)
+        assert result == explicit_key
+        mock_get_secret.assert_not_called()
+
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.passthrough.transformation.get_secret_str"
+    )
+    def test_get_api_key_from_environment(self, mock_get_secret):
+        """Test get_api_key retrieves from environment when not provided."""
+        env_key = "env-api-key"
+        mock_get_secret.return_value = env_key
+        result = GigaChatPassthroughConfig.get_api_key(api_key=None)
+        assert result == env_key
+        mock_get_secret.assert_called_once_with("GIGACHAT_API_KEY")
+
+    def test_get_base_model_returns_model(self):
+        """Test get_base_model returns the model as-is."""
+        model = "gigachat/GigaChat"
+        result = GigaChatPassthroughConfig.get_base_model(model)
+        assert result == model
+
+    def test_get_models(self):
+        """Test get_models delegates to base class."""
+        config = GigaChatPassthroughConfig()
+        result = config.get_models()
+        assert result == []
+
+    def test_logging_non_streaming_chat_raises_when_no_config(self):
+        """Test raise when ProviderConfigManager returns None for chat."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        with patch(
+            "litellm.utils.ProviderConfigManager.get_provider_chat_config",  # test-quality-ok: patching litellm internal for unit test isolation
+            return_value=None,
+        ):
+            with pytest.raises(ValueError, match="No provider config found for model"):
+                config.logging_non_streaming_response(
+                    model="gigachat/GigaChat",
+                    custom_llm_provider="gigachat",
+                    httpx_response=_make_httpx_response(_gigachat_chat_completion_body()),
+                    request_data={
+                        "model": "gigachat/GigaChat",
+                        "messages": [{"role": "user", "content": "hi"}],
+                    },
+                    logging_obj=logging_obj,
+                    endpoint="chat/completions",
+                )
+
+    def test_logging_non_streaming_embedding_raises_when_no_config(self):
+        """Test raise when ProviderConfigManager returns None for embeddings."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        with patch(
+            "litellm.utils.ProviderConfigManager.get_provider_embedding_config",  # test-quality-ok: patching litellm internal for unit test isolation
+            return_value=None,
+        ):
+            with pytest.raises(ValueError, match="No provider config found for model"):
+                config.logging_non_streaming_response(
+                    model="gigachat/Embeddings",
+                    custom_llm_provider="gigachat",
+                    httpx_response=_make_httpx_response(_gigachat_embedding_body()),
+                    request_data={
+                        "input": ["hello"],
+                        "model": "gigachat/Embeddings",
+                    },
+                    logging_obj=logging_obj,
+                    endpoint="embeddings",
+                )
+
+    def test_handle_logging_collected_chunks_with_model_response_stream_chunk(self):
+        """Test that a chunk returning ModelResponseStream from chunk_parser is handled.
+
+        Requires patching GigaChatModelResponseIterator.chunk_parser to return
+        a ModelResponseStream so the elif branch is exercised.
+        """
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        from litellm.types.utils import ModelResponseStream
+
+        stream_chunk = ModelResponseStream(
+            choices=[
+                {
+                    "index": 0,
+                    "delta": {"content": "streamed"},
+                    "finish_reason": None,
+                }
+            ]
+        )
+
+        chunks = [
+            '{"choices": [{"delta": {"content": "streamed"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}}',
+        ]
+
+        with patch(
+            "litellm.llms.gigachat.passthrough.transformation.GigaChatModelResponseIterator.chunk_parser",  # test-quality-ok: patching litellm internal for unit test isolation
+            return_value=stream_chunk,
+        ):
+            result = config.handle_logging_collected_chunks(
+                all_chunks=chunks,
+                litellm_logging_obj=logging_obj,
+                model="gigachat/GigaChat",
+                custom_llm_provider="gigachat",
+                endpoint="chat/completions",
+            )
+
+        assert isinstance(result, ModelResponse)
+        assert result.choices[0].message.content == "streamedstreamed"
+
+    def test_handle_logging_collected_chunks_skips_unknown_chunk_type(self):
+        """Test that chunk_parser returning an unknown type is skipped."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        chunks = [
+            '{"choices": [{"delta": {"content": "good"}, "index": 0}]}',
+            '{"choices": [{"delta": {}, "finish_reason": "stop", "index": 0}], "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}}',
+        ]
+
+        with patch(
+            "litellm.llms.gigachat.passthrough.transformation.GigaChatModelResponseIterator.chunk_parser",  # test-quality-ok: patching litellm internal for unit test isolation
+            return_value=12345,  # not dict and not ModelResponseStream
+        ):
+            result = config.handle_logging_collected_chunks(
+                all_chunks=chunks,
+                litellm_logging_obj=logging_obj,
+                model="gigachat/GigaChat",
+                custom_llm_provider="gigachat",
+                endpoint="chat/completions",
+            )
+
+        # All chunks skipped, returns None
+        assert result is None
+
+    def test_handle_logging_collected_chunks_skips_unsupported_chunk_type(self):
+        """Test that unsupported chunk types (non-JSON str) are skipped."""
+        config = GigaChatPassthroughConfig()
+        logging_obj = MagicMock()
+
+        # Both are valid str chunks; "not-a-valid-json" fails json.loads, int is not a str
+        chunks: list[str] = ["not-a-valid-json"]
+
+        result = config.handle_logging_collected_chunks(
+            all_chunks=chunks,
+            litellm_logging_obj=logging_obj,
+            model="gigachat/GigaChat",
+            custom_llm_provider="gigachat",
+            endpoint="chat/completions",
+        )
+
+        assert result is None

tests/test_litellm/llms/gigachat/test_authenticator.py

@@ -0,0 +1,494 @@
+"""
+Unit tests for GigaChat OAuth authenticator.
+
+Tests get_access_token and get_access_token_async covering token resolution
+from litellm_params/env, credential validation, caching, and error handling.
+"""
+
+import time
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import httpx
+import pytest
+
+from litellm.llms.gigachat import authenticator
+from litellm.llms.gigachat.authenticator import (
+    GigaChatAuthError,
+    TOKEN_EXPIRY_BUFFER_MS,
+    get_access_token,
+    get_access_token_async,
+)
+
+
+AUTH_MODULE = "litellm.llms.gigachat.authenticator"
+
+
+def _future_expires_at_ms(offset_seconds: float = 3600) -> int:
+    return int(time.time() * 1000 + offset_seconds * 1000)
+
+
+def _past_expires_at_ms(offset_seconds: float = 3600) -> int:
+    return int(time.time() * 1000 - offset_seconds * 1000)
+
+
+@pytest.fixture(autouse=True)
+def _isolate_token_cache():
+    """Each test gets a fresh module-level token cache to avoid cross-test leakage."""
+    with patch(f"{AUTH_MODULE}._token_cache", new=MagicMock()):
+        authenticator._token_cache.get_cache.return_value = None
+        authenticator._token_cache.set_cache = MagicMock()
+        yield
+
+
+class TestGetAccessTokenSync:
+    def test_returns_token_from_litellm_params(self):
+        token = get_access_token(litellm_params={"gigachat_access_token": "param-token"})
+        assert token == "param-token"
+        authenticator._token_cache.get_cache.assert_not_called()
+
+    @patch(f"{AUTH_MODULE}.get_secret_str")
+    def test_returns_token_from_env(self, mock_get_secret):
+        mock_get_secret.return_value = "env-access-token"
+        token = get_access_token()
+        assert token == "env-access-token"
+
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value=None)
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_raises_when_no_credentials(self, mock_get_secret, mock_get_creds):
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            get_access_token()
+        assert exc_info.value.status_code == 401
+        assert "credentials not provided" in exc_info.value.message
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value=None)
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_raises_when_no_credentials_even_with_other_resolvers(
+        self, mock_get_secret, mock_get_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            get_access_token()
+        assert exc_info.value.status_code == 401
+        mock_request.assert_not_called()
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds-from-env")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_requests_new_token_and_caches(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        token = "fresh-token"
+        expires_at = _future_expires_at_ms()
+        mock_request.return_value = (token, expires_at)
+
+        result = get_access_token()
+
+        assert result == token
+        mock_request.assert_called_once_with("creds-from-env", "GIGACHAT_API_PERS", "https://auth.example.com")
+        authenticator._token_cache.set_cache.assert_called_once()
+        call_args = authenticator._token_cache.set_cache.call_args
+        assert call_args.args[1] == (token, expires_at)
+        assert call_args.kwargs["ttl"] > 0
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_does_not_cache_when_no_expiry(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        mock_request.return_value = ("token-no-exp", 0)
+
+        result = get_access_token()
+
+        assert result == "token-no-exp"
+        authenticator._token_cache.set_cache.assert_not_called()
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_does_not_cache_when_ttl_non_positive(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        expires_at = int(time.time() * 1000) + TOKEN_EXPIRY_BUFFER_MS - 1000
+        mock_request.return_value = ("token", expires_at)
+
+        result = get_access_token()
+
+        assert result == "token"
+        authenticator._token_cache.set_cache.assert_not_called()
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_returns_cached_valid_token(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        cached_token = "cached-token"
+        cached_expires_at = _future_expires_at_ms(offset_seconds=7200)
+        authenticator._token_cache.get_cache.return_value = (cached_token, cached_expires_at)
+
+        result = get_access_token(credentials="creds")
+
+        assert result == cached_token
+        mock_request.assert_not_called()
+        authenticator._token_cache.set_cache.assert_not_called()
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_requests_new_token_when_cache_expired(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        cached_token = "stale-token"
+        cached_expires_at = _past_expires_at_ms(offset_seconds=10)
+        authenticator._token_cache.get_cache.return_value = (cached_token, cached_expires_at)
+
+        new_token = "refreshed-token"
+        mock_request.return_value = (new_token, _future_expires_at_ms())
+
+        result = get_access_token(credentials="creds")
+
+        assert result == new_token
+        mock_request.assert_called_once()
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://default-auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="env-creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_litellm_params_override_scope_and_auth_url(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):  # test-quality-ok: mock-echo of internal wiring
+        mock_request.return_value = ("token", _future_expires_at_ms())
+
+        get_access_token(
+            litellm_params={
+                "gigachat_scope": "GIGACHAT_API_CORP",
+                "gigachat_auth_url": "https://params-auth.example.com",
+            }
+        )
+
+        mock_request.assert_called_once_with(  # test-quality-ok: mock-echo of internal wiring
+            "env-creds", "GIGACHAT_API_CORP", "https://params-auth.example.com"
+        )
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://default-auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="env-creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_explicit_args_override_everything(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):  # test-quality-ok: mock-echo of internal wiring
+        mock_request.return_value = ("token", _future_expires_at_ms())
+
+        get_access_token(
+            credentials="explicit-creds",
+            scope="EXPLICIT_SCOPE",
+            auth_url="https://explicit.example.com",
+            litellm_params={
+                "gigachat_scope": "PARAM_SCOPE",
+                "gigachat_auth_url": "https://params.example.com",
+            },
+        )
+
+        mock_request.assert_called_once_with(  # test-quality-ok: mock-echo of internal wiring
+            "explicit-creds", "EXPLICIT_SCOPE", "https://explicit.example.com"
+        )
+
+    @patch(f"{AUTH_MODULE}._request_token_sync")
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    def test_propagates_auth_error_from_request(self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request):
+        mock_request.side_effect = GigaChatAuthError(status_code=403, message="forbidden")
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            get_access_token()
+        assert exc_info.value.status_code == 403
+        assert exc_info.value.message == "forbidden"
+
+
+class TestGetAccessTokenAsync:
+    @pytest.mark.asyncio
+    async def test_returns_token_from_litellm_params(self):
+        token = await get_access_token_async(
+            litellm_params={"gigachat_access_token": "param-token"}
+        )
+        assert token == "param-token"
+        authenticator._token_cache.get_cache.assert_not_called()
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}.get_secret_str")
+    async def test_returns_token_from_env(self, mock_get_secret):
+        mock_get_secret.return_value = "env-access-token"
+        token = await get_access_token_async()
+        assert token == "env-access-token"
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value=None)
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_raises_when_no_credentials(self, mock_get_secret, mock_get_creds):
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            await get_access_token_async()
+        assert exc_info.value.status_code == 401
+        assert "credentials not provided" in exc_info.value.message
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds-from-env")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_requests_new_token_and_caches(
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        token = "fresh-token-async"
+        expires_at = _future_expires_at_ms()
+        mock_request.return_value = (token, expires_at)
+
+        result = await get_access_token_async()
+
+        assert result == token
+        mock_request.assert_called_once_with(
+            "creds-from-env", "GIGACHAT_API_PERS", "https://auth.example.com"
+        )
+        authenticator._token_cache.set_cache.assert_called_once()
+        call_args = authenticator._token_cache.set_cache.call_args
+        assert call_args.args[1] == (token, expires_at)
+        assert call_args.kwargs["ttl"] > 0
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_does_not_cache_when_no_expiry(
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        mock_request.return_value = ("token-no-exp", 0)
+
+        result = await get_access_token_async()
+
+        assert result == "token-no-exp"
+        authenticator._token_cache.set_cache.assert_not_called()
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_returns_cached_valid_token(
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        cached_token = "cached-token-async"
+        cached_expires_at = _future_expires_at_ms(offset_seconds=7200)
+        authenticator._token_cache.get_cache.return_value = (cached_token, cached_expires_at)
+
+        result = await get_access_token_async(credentials="creds")
+
+        assert result == cached_token
+        mock_request.assert_not_called()
+        authenticator._token_cache.set_cache.assert_not_called()
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_requests_new_token_when_cache_expired(
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        cached_expires_at = _past_expires_at_ms(offset_seconds=10)
+        authenticator._token_cache.get_cache.return_value = ("stale", cached_expires_at)
+
+        new_token = "refreshed-token-async"
+        mock_request.return_value = (new_token, _future_expires_at_ms())
+
+        result = await get_access_token_async(credentials="creds")
+
+        assert result == new_token
+        mock_request.assert_called_once()
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://default-auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="env-creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_litellm_params_override_scope_and_auth_url(  # test-quality-ok: mock-echo of internal wiring
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        mock_request.return_value = ("token", _future_expires_at_ms())
+
+        await get_access_token_async(
+            litellm_params={
+                "gigachat_scope": "GIGACHAT_API_CORP",
+                "gigachat_auth_url": "https://params-auth.example.com",
+            }
+        )
+
+        mock_request.assert_called_once_with(  # test-quality-ok: mock-echo of internal wiring
+            "env-creds", "GIGACHAT_API_CORP", "https://params-auth.example.com"
+        )
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://default-auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="env-creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_explicit_args_override_everything(  # test-quality-ok: mock-echo of internal wiring
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        mock_request.return_value = ("token", _future_expires_at_ms())
+
+        await get_access_token_async(
+            credentials="explicit-creds",
+            scope="EXPLICIT_SCOPE",
+            auth_url="https://explicit.example.com",
+            litellm_params={
+                "gigachat_scope": "PARAM_SCOPE",
+                "gigachat_auth_url": "https://params.example.com",
+            },
+        )
+
+        mock_request.assert_called_once_with(  # test-quality-ok: mock-echo of internal wiring
+            "explicit-creds", "EXPLICIT_SCOPE", "https://explicit.example.com"
+        )
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}._request_token_async", new_callable=AsyncMock)
+    @patch(f"{AUTH_MODULE}._get_auth_url", return_value="https://auth.example.com")
+    @patch(f"{AUTH_MODULE}._get_scope", return_value="GIGACHAT_API_PERS")
+    @patch(f"{AUTH_MODULE}._get_credentials", return_value="creds")
+    @patch(f"{AUTH_MODULE}.get_secret_str", return_value=None)
+    async def test_propagates_auth_error_from_request(
+        self, mock_get_secret, mock_creds, mock_scope, mock_auth_url, mock_request
+    ):
+        mock_request.side_effect = GigaChatAuthError(status_code=403, message="forbidden")
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            await get_access_token_async()
+        assert exc_info.value.status_code == 403
+        assert exc_info.value.message == "forbidden"
+
+
+class TestRequestTokenSyncErrorMapping:
+    @patch(f"{AUTH_MODULE}._get_http_client")
+    def test_http_status_error_maps_to_auth_error(self, mock_get_client):
+        client = MagicMock()
+        request = httpx.Request("POST", "https://auth.example.com")
+        response = httpx.Response(status_code=401, content=b"bad creds", request=request)
+        http_error = httpx.HTTPStatusError("unauthorized", request=request, response=response)
+        client.post.side_effect = http_error
+        mock_get_client.return_value = client
+
+        from litellm.llms.gigachat.authenticator import _request_token_sync
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            _request_token_sync("creds", "GIGACHAT_API_PERS", "https://auth.example.com")
+        assert exc_info.value.status_code == 401
+        assert "bad creds" in exc_info.value.message
+
+    @patch(f"{AUTH_MODULE}._get_http_client")
+    def test_request_error_maps_to_auth_error(self, mock_get_client):
+        client = MagicMock()
+        client.post.side_effect = httpx.ConnectError("connection refused")
+        mock_get_client.return_value = client
+
+        from litellm.llms.gigachat.authenticator import _request_token_sync
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            _request_token_sync("creds", "GIGACHAT_API_PERS", "https://auth.example.com")
+        assert exc_info.value.status_code == 500
+        assert "connection refused" in exc_info.value.message
+
+
+class TestRequestTokenAsyncErrorMapping:
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}.get_async_httpx_client")
+    async def test_http_status_error_maps_to_auth_error(self, mock_get_client):
+        client = MagicMock()
+        request = httpx.Request("POST", "https://auth.example.com")
+        response = httpx.Response(status_code=401, content=b"bad creds", request=request)
+        http_error = httpx.HTTPStatusError("unauthorized", request=request, response=response)
+        client.post = AsyncMock(side_effect=http_error)
+        mock_get_client.return_value = client
+
+        from litellm.llms.gigachat.authenticator import _request_token_async
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            await _request_token_async("creds", "GIGACHAT_API_PERS", "https://auth.example.com")
+        assert exc_info.value.status_code == 401
+        assert "bad creds" in exc_info.value.message
+
+    @pytest.mark.asyncio
+    @patch(f"{AUTH_MODULE}.get_async_httpx_client")
+    async def test_request_error_maps_to_auth_error(self, mock_get_client):
+        client = MagicMock()
+        client.post = AsyncMock(side_effect=httpx.ConnectError("connection refused"))
+        mock_get_client.return_value = client
+
+        from litellm.llms.gigachat.authenticator import _request_token_async
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            await _request_token_async("creds", "GIGACHAT_API_PERS", "https://auth.example.com")
+        assert exc_info.value.status_code == 500
+        assert "connection refused" in exc_info.value.message
+
+
+class TestParseTokenResponse:
+    def _make_response(self, body: dict) -> httpx.Response:
+        import json
+
+        return httpx.Response(
+            status_code=200,
+            content=json.dumps(body).encode("utf-8"),
+            request=httpx.Request("POST", "https://auth.example.com"),
+        )
+
+    def test_parses_tok_exp_fields(self):
+        from litellm.llms.gigachat.authenticator import _parse_token_response
+
+        token, expires_at = _parse_token_response(
+            self._make_response({"tok": "abc", "exp": 1700000000000})
+        )
+        assert token == "abc"
+        assert expires_at == 1700000000000
+
+    def test_parses_access_token_expires_at_fields(self):
+        from litellm.llms.gigachat.authenticator import _parse_token_response
+
+        token, expires_at = _parse_token_response(
+            self._make_response({"access_token": "xyz", "expires_at": 1700000000000})
+        )
+        assert token == "xyz"
+        assert expires_at == 1700000000000
+
+    def test_parses_string_expires_at(self):
+        from litellm.llms.gigachat.authenticator import _parse_token_response
+
+        token, expires_at = _parse_token_response(
+            self._make_response({"tok": "abc", "exp": "1700000000000"})
+        )
+        assert token == "abc"
+        assert expires_at == 1700000000000
+        assert isinstance(expires_at, int)
+
+    def test_raises_when_no_access_token(self):
+        from litellm.llms.gigachat.authenticator import _parse_token_response
+
+        with pytest.raises(GigaChatAuthError) as exc_info:
+            _parse_token_response(self._make_response({"exp": 1700000000000}))
+        assert exc_info.value.status_code == 500
+        assert "Invalid token response" in exc_info.value.message
+
+
+class TestGetHttpClient:
+    def test_reuses_cached_client_across_calls(self):
+        """Regression: the sync OAuth path must use the shared cached httpx client,
+        not construct a fresh HTTPHandler per token request."""
+        assert authenticator._get_http_client() is authenticator._get_http_client()

tests/test_litellm/llms/gigachat/test_file_handler.py

@@ -0,0 +1,504 @@
+"""
+Unit tests for GigaChat file handler.
+
+Tests _get_url_hash, _parse_data_url, _download_image_sync, _download_image_async,
+upload_file_sync, and upload_file_async covering caching, base64 data URL decoding,
+network errors, and the full upload flow.
+"""
+
+import json
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import httpx
+import pytest
+
+from litellm.llms.gigachat import file_handler
+from litellm.llms.gigachat.file_handler import (
+    _file_cache,
+    _get_url_hash,
+    _parse_data_url,
+    upload_file_async,
+    upload_file_sync,
+)
+
+FILE_MODULE = "litellm.llms.gigachat.file_handler"
+
+# A valid 1x1 red PNG as base64
+_RED_PNG_B64 = (
+    "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAA"
+    "DUlEQVQI12NgYPgPAAEDAQAR3X3ZAAAASUVORK5CYII="
+)
+_RED_PNG_DATA_URL = f"data:image/png;base64,{_RED_PNG_B64}"
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture(autouse=True)
+def _isolate_file_cache():
+    """Each test gets a fresh module-level file cache to avoid cross-test leakage."""
+    _file_cache.clear()
+    yield
+    _file_cache.clear()
+
+
+# ---------------------------------------------------------------------------
+# _get_url_hash
+# ---------------------------------------------------------------------------
+
+
+class TestGetUrlHash:
+    def test_returns_hex_string(self):
+        h = _get_url_hash("https://example.com/image.png")
+        assert isinstance(h, str)
+        assert len(h) == 64  # SHA-256
+
+    def test_different_urls_different_hashes(self):
+        h1 = _get_url_hash("https://example.com/a.png")
+        h2 = _get_url_hash("https://example.com/b.png")
+        assert h1 != h2
+
+    def test_same_url_same_hash(self):
+        h1 = _get_url_hash("https://example.com/image.png")
+        h2 = _get_url_hash("https://example.com/image.png")
+        assert h1 == h2
+
+
+# ---------------------------------------------------------------------------
+# _parse_data_url
+# ---------------------------------------------------------------------------
+
+
+class TestParseDataUrl:
+    def test_valid_base64_png(self):
+        result = _parse_data_url(_RED_PNG_DATA_URL)
+        assert result is not None
+        content_bytes, content_type, ext = result
+        assert content_type == "image/png"
+        assert ext == "png"
+        assert len(content_bytes) > 0
+
+    def test_valid_base64_jpeg(self):
+        # Simple valid base64 (24 chars, properly padded, no + or / chars)
+        valid_b64 = "aGVsbG8gd29ybGQhISEhIQ=="
+        data_url = f"data:image/jpeg;base64,{valid_b64}"
+        result = _parse_data_url(data_url)
+        assert result is not None
+        _, content_type, ext = result
+        assert content_type == "image/jpeg"
+        assert ext == "jpeg"
+
+    def test_valid_base64_with_semicolon_in_type(self):
+        """Data URLs with charset before base64 segment do not match the regex."""
+        # The regex `data:([^;]+);base64,(.+)` requires the pattern to be
+        # `data:<type>;base64,<data>`. If `;charset=utf-8` appears before
+        # `;base64,`, the regex sees `data:image/png` as group 1 but then
+        # looks for `;base64,` immediately after — which isn't there because
+        # `;charset=utf-8;base64,` has extra text before `;base64,`
+        data_url = "data:image/png;charset=utf-8;base64," + _RED_PNG_B64
+        result = _parse_data_url(data_url)
+        assert result is None
+
+    def test_invalid_data_url_returns_none(self):
+        assert _parse_data_url("not-a-data-url") is None
+
+    def test_empty_base64_returns_none(self):
+        """Empty base64 data (nothing after comma) does not match regex `(.+)`."""
+        assert _parse_data_url("data:image/png;base64,") is None
+
+    def test_missing_base64_segment(self):
+        assert _parse_data_url("data:image/png;base64") is None
+
+    def test_unknown_extension_falls_back_to_jpg(self):
+        data_url = "data:application/octet-stream;base64," + _RED_PNG_B64
+        result = _parse_data_url(data_url)
+        assert result is not None
+        _, content_type, ext = result
+        assert content_type == "application/octet-stream"
+        # The extension is derived from content_type.split("/")[-1].split(";")[0]
+        # which gives "octet-stream", not "jpg"
+        assert ext == "octet-stream"
+
+
+# ---------------------------------------------------------------------------
+# _download_image_sync
+# ---------------------------------------------------------------------------
+
+
+class TestDownloadImageSync:
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_downloads_image_successfully(self, mock_http_handler_cls):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.content = b"fake-image-bytes"
+        mock_response.headers = {"content-type": "image/jpeg"}
+        mock_client.get.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        content_bytes, content_type, ext = file_handler._download_image_sync("https://example.com/img.jpg")
+
+        assert content_bytes == b"fake-image-bytes"
+        assert content_type == "image/jpeg"
+        assert ext == "jpeg"
+        mock_client.get.assert_called_once_with("https://example.com/img.jpg")
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_raises_on_http_error(self, mock_http_handler_cls):
+        mock_client = MagicMock()
+        mock_client.get.side_effect = httpx.HTTPStatusError(
+            "Not Found",
+            request=httpx.Request("GET", "https://example.com/404"),
+            response=httpx.Response(status_code=404, request=httpx.Request("GET", "https://example.com/404")),
+        )
+        mock_http_handler_cls.return_value = mock_client
+
+        with pytest.raises(httpx.HTTPStatusError):
+            file_handler._download_image_sync("https://example.com/404")
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_parse_content_type_fallback(self, mock_http_handler_cls):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.content = b"data"
+        mock_response.headers = {}
+        mock_client.get.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        _, content_type, ext = file_handler._download_image_sync("https://example.com/img")
+
+        assert content_type == "image/jpeg"
+        assert ext == "jpeg"
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_extracts_extension_from_parametrized_type(self, mock_http_handler_cls):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.content = b"data"
+        mock_response.headers = {"content-type": "image/png; charset=utf-8"}
+        mock_client.get.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        _, _, ext = file_handler._download_image_sync("https://example.com/img.png")
+
+        assert ext == "png"
+
+
+# ---------------------------------------------------------------------------
+# _download_image_async
+# ---------------------------------------------------------------------------
+
+
+class TestDownloadImageAsync:
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    async def test_downloads_image_successfully(self, mock_get_client):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.content = b"fake-image-bytes"
+        mock_response.headers = {"content-type": "image/webp"}
+        mock_client.get = AsyncMock(return_value=mock_response)
+        mock_get_client.return_value = mock_client
+
+        content_bytes, content_type, ext = await file_handler._download_image_async(
+            "https://example.com/img.webp"
+        )
+
+        assert content_bytes == b"fake-image-bytes"
+        assert content_type == "image/webp"
+        assert ext == "webp"
+        mock_client.get.assert_called_once_with("https://example.com/img.webp")
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    async def test_raises_on_http_error(self, mock_get_client):
+        mock_client = MagicMock()
+        mock_client.get = AsyncMock(
+            side_effect=httpx.HTTPStatusError(
+                "Forbidden",
+                request=httpx.Request("GET", "https://example.com/403"),
+                response=httpx.Response(status_code=403, request=httpx.Request("GET", "https://example.com/403")),
+            )
+        )
+        mock_get_client.return_value = mock_client
+
+        with pytest.raises(httpx.HTTPStatusError):
+            await file_handler._download_image_async("https://example.com/403")
+
+
+# ---------------------------------------------------------------------------
+# upload_file_sync
+# ---------------------------------------------------------------------------
+
+
+class TestUploadFileSync:
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_uploads_base64_image_and_caches(
+        self, mock_http_handler_cls, mock_get_token, mock_get_api_base
+    ):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json.return_value = {"id": "file-12345"}
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        result = upload_file_sync(
+            image_url=_RED_PNG_DATA_URL,
+            credentials="creds",
+            api_base="https://custom.example.com",
+        )
+
+        assert result == "file-12345"
+        # Verify it was cached
+        url_hash = _get_url_hash(_RED_PNG_DATA_URL)
+        assert _file_cache[url_hash] == "file-12345"
+
+        # Check the upload request — url is passed as first positional arg
+        call_args = mock_client.post.call_args
+        assert call_args.args[0] == "https://api.example.com/files"
+        assert call_args.kwargs["headers"]["Authorization"] == "Bearer test-token"
+        # Verify purpose
+        assert call_args.kwargs["data"] == {"purpose": "general"}
+        # Verify a file was attached
+        assert "file" in call_args.kwargs["files"]
+
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_returns_cached_file_id(
+        self, mock_http_handler_cls, mock_get_token, mock_get_api_base
+    ):
+        # Pre-populate the cache
+        url_hash = _get_url_hash(_RED_PNG_DATA_URL)
+        _file_cache[url_hash] = "cached-file-id"
+
+        result = upload_file_sync(image_url=_RED_PNG_DATA_URL, credentials="creds")
+
+        assert result == "cached-file-id"
+        # No upload call was made
+        mock_http_handler_cls.return_value.post.assert_not_called()
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}._download_image_sync")
+    def test_downloads_and_uploads_url_image(
+        self, mock_download, mock_get_api_base, mock_get_token, mock_http_handler_cls
+    ):
+        mock_download.return_value = (b"remote-bytes", "image/png", "png")
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json.return_value = {"id": "file-remote"}
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        result = upload_file_sync(
+            image_url="https://example.com/remote.png", credentials="creds"
+        )
+
+        assert result == "file-remote"
+        mock_download.assert_called_once_with("https://example.com/remote.png")
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    def test_returns_none_on_upload_failure(
+        self, mock_get_api_base, mock_get_token, mock_http_handler_cls
+    ):
+        mock_client = MagicMock()
+        mock_client.post.side_effect = httpx.HTTPStatusError(
+            "Bad Request",
+            request=httpx.Request("POST", "https://api.example.com/files"),
+            response=httpx.Response(status_code=400, request=httpx.Request("POST", "https://api.example.com/files")),
+        )
+        mock_http_handler_cls.return_value = mock_client
+
+        # upload_file_sync catches all exceptions and returns None
+        result = upload_file_sync(
+            image_url=_RED_PNG_DATA_URL, credentials="creds"
+        )
+
+        assert result is None
+
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    def test_returns_none_when_response_missing_id(
+        self, mock_get_api_base, mock_get_token, mock_http_handler_cls
+    ):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json.return_value = {"status": "ok"}  # no "id" key
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        result = upload_file_sync(
+            image_url=_RED_PNG_DATA_URL, credentials="creds"
+        )
+
+        assert result is None
+
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token", return_value="test-token")
+    @patch(f"{FILE_MODULE}._get_httpx_client")
+    def test_uploads_without_optional_args(
+        self, mock_http_handler_cls, mock_get_token, mock_get_api_base
+    ):
+        """Verify that credentials, api_base, and litellm_params are optional."""
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json.return_value = {"id": "file-no-args"}
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post.return_value = mock_response
+        mock_http_handler_cls.return_value = mock_client
+
+        result = upload_file_sync(image_url=_RED_PNG_DATA_URL)
+
+        assert result == "file-no-args"
+        # Should still have called get_access_token without args
+        mock_get_token.assert_called_once_with(credentials=None, litellm_params=None)
+
+
+# ---------------------------------------------------------------------------
+# upload_file_async
+# ---------------------------------------------------------------------------
+
+
+class TestUploadFileAsync:
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    async def test_uploads_base64_image_and_caches(
+        self, mock_get_client, mock_get_token, mock_get_api_base
+    ):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json = MagicMock(return_value={"id": "async-file-1"})
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post = AsyncMock(return_value=mock_response)
+        mock_get_client.return_value = mock_client
+
+        result = await upload_file_async(
+            image_url=_RED_PNG_DATA_URL,
+            credentials="creds",
+            api_base="https://custom.example.com",
+        )
+
+        assert result == "async-file-1"
+        # Verify cache
+        url_hash = _get_url_hash(_RED_PNG_DATA_URL)
+        assert _file_cache[url_hash] == "async-file-1"
+
+        # Check upload request details — url is first positional arg
+        call_args = mock_client.post.call_args
+        assert call_args.args[0] == "https://api.example.com/files"
+        assert call_args.kwargs["headers"]["Authorization"] == "Bearer test-token-async"
+        assert "purpose" in str(call_args.kwargs["data"])
+        assert "file" in call_args.kwargs["files"]
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    async def test_returns_cached_file_id(
+        self, mock_get_client, mock_get_token, mock_get_api_base
+    ):
+        url_hash = _get_url_hash(_RED_PNG_DATA_URL)
+        _file_cache[url_hash] = "cached-async-id"
+
+        result = await upload_file_async(image_url=_RED_PNG_DATA_URL, credentials="creds")
+
+        assert result == "cached-async-id"
+        mock_get_client.return_value.post.assert_not_called()
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}._download_image_async")
+    async def test_downloads_and_uploads_url_image(
+        self, mock_download, mock_get_api_base, mock_get_token, mock_get_client
+    ):
+        mock_download.return_value = (b"remote-bytes-async", "image/png", "png")
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json = MagicMock(return_value={"id": "async-file-remote"})
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post = AsyncMock(return_value=mock_response)
+        mock_get_client.return_value = mock_client
+
+        result = await upload_file_async(
+            image_url="https://example.com/remote.png", credentials="creds"
+        )
+
+        assert result == "async-file-remote"
+        mock_download.assert_called_once_with("https://example.com/remote.png")
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    async def test_returns_none_on_upload_failure(
+        self, mock_get_api_base, mock_get_token, mock_get_client
+    ):
+        mock_client = MagicMock()
+        mock_client.post = AsyncMock(
+            side_effect=httpx.HTTPStatusError(
+                "Bad Request",
+                request=httpx.Request("POST", "https://api.example.com/files"),
+                response=httpx.Response(status_code=400, request=httpx.Request("POST", "https://api.example.com/files")),
+            )
+        )
+        mock_get_client.return_value = mock_client
+
+        result = await upload_file_async(
+            image_url=_RED_PNG_DATA_URL, credentials="creds"
+        )
+
+        assert result is None
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    async def test_returns_none_when_response_missing_id(
+        self, mock_get_api_base, mock_get_token, mock_get_client
+    ):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json = MagicMock(return_value={"status": "ok"})
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post = AsyncMock(return_value=mock_response)
+        mock_get_client.return_value = mock_client
+
+        result = await upload_file_async(
+            image_url=_RED_PNG_DATA_URL, credentials="creds"
+        )
+
+        assert result is None
+
+    @pytest.mark.asyncio
+    @patch(f"{FILE_MODULE}.get_api_base", return_value="https://api.example.com")
+    @patch(f"{FILE_MODULE}.get_access_token_async", return_value="test-token-async")
+    @patch(f"{FILE_MODULE}.get_async_httpx_client")
+    async def test_uploads_without_optional_args(
+        self, mock_get_client, mock_get_token, mock_get_api_base
+    ):
+        mock_client = MagicMock()
+        mock_response = MagicMock()
+        mock_response.json = MagicMock(return_value={"id": "async-no-args"})
+        mock_response.raise_for_status = MagicMock()
+        mock_client.post = AsyncMock(return_value=mock_response)
+        mock_get_client.return_value = mock_client
+
+        result = await upload_file_async(image_url=_RED_PNG_DATA_URL)
+
+        assert result == "async-no-args"
+        mock_get_token.assert_called_once_with(credentials=None, litellm_params=None)

tests/test_litellm/llms/gigachat/test_utils.py

@@ -0,0 +1,79 @@
+"""
+Tests for litellm.llms.gigachat.utils
+"""
+
+import pytest
+from litellm.llms.gigachat.utils import convert_usage
+from litellm.types.utils import PromptTokensDetailsWrapper, Usage
+
+
+class TestConvertUsage:
+    def test_basic_usage_without_precached(self):
+        """Test convert_usage with standard tokens, no precached prompt tokens."""
+        result = convert_usage(
+            {
+                "prompt_tokens": 10,
+                "completion_tokens": 5,
+                "total_tokens": 15,
+            }
+        )
+
+        assert result == Usage(
+            prompt_tokens=10,
+            completion_tokens=5,
+            total_tokens=15,
+            prompt_tokens_details=None,
+        )
+
+    def test_usage_with_precached_prompt_tokens(self):
+        """GigaChat's prompt_tokens and total_tokens exclude cached tokens (docs example:
+        prompt_tokens=1, precached_prompt_tokens=37, total_tokens=5), so OpenAI-convention
+        usage adds precached back in and surfaces it as cached_tokens."""
+        result = convert_usage(
+            {
+                "prompt_tokens": 10,
+                "completion_tokens": 5,
+                "precached_prompt_tokens": 3,
+                "total_tokens": 15,
+            }
+        )
+
+        assert result == Usage(
+            prompt_tokens=13,
+            completion_tokens=5,
+            total_tokens=18,
+            prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=3),
+        )
+
+    def test_zero_precached_prompt_tokens(self):
+        """Test convert_usage with zero precached_prompt_tokens does not create details wrapper."""
+        result = convert_usage(
+            {
+                "prompt_tokens": 10,
+                "completion_tokens": 5,
+                "precached_prompt_tokens": 0,
+                "total_tokens": 15,
+            }
+        )
+
+        assert result == Usage(
+            prompt_tokens=10,
+            completion_tokens=5,
+            total_tokens=15,
+            prompt_tokens_details=None,
+        )
+
+    def test_missing_optional_fields(self):
+        """Test convert_usage with missing optional fields defaults to zero."""
+        result = convert_usage(
+            {
+                "prompt_tokens": 10,
+                "completion_tokens": 5,
+                "total_tokens": 15,
+            }
+        )
+
+        assert result.prompt_tokens == 10
+        assert result.completion_tokens == 5
+        assert result.total_tokens == 15
+        assert result.prompt_tokens_details is None

tests/test_litellm/models/test_models.py

@@ -5,6 +5,7 @@ Tests for backend domain models.
 from datetime import datetime
 
 import pytest
+from pydantic import BaseModel, TypeAdapter
 
 from litellm.models.access_group import LiteLLM_AccessGroupTable
 from litellm.models.budget import (
@@ -130,6 +131,33 @@ class TestModel:
         assert model.litellm_params == {"model": "gpt-4"}
         assert model.model_info == {"team_id": "t1"}
 
+    def test_response_type_adapter_accepts_pydantic_row(self):
+        class PrismaModelRow(BaseModel):
+            model_id: str
+            model_name: str
+            litellm_params: dict[str, str]
+            model_info: dict[str, str] | None = None
+            blocked: bool = False
+
+        row = PrismaModelRow(
+            model_id="m1",
+            model_name="gpt-4",
+            litellm_params={"model": "gpt-4"},
+            model_info={"team_id": "t1"},
+            blocked=True,
+        )
+
+        model = TypeAdapter(LiteLLM_ProxyModelTable | None).validate_python(
+            row,
+            from_attributes=True,
+        )
+
+        assert model is not None
+        assert model.model_id == "m1"
+        assert model.litellm_params == {"model": "gpt-4"}
+        assert model.model_info == {"team_id": "t1"}
+        assert model.blocked is True
+
     def test_team_helpers_none_when_no_model_info(self):
         model = LiteLLM_ProxyModelTable(
             model_id="m1", model_name="gpt-4", litellm_params={}, model_info=None

tests/test_litellm/passthrough/test_async_streaming_error_propagation.py

@@ -1,12 +1,12 @@
 """
-Tests for error propagation in _async_streaming passthrough routes.
+Tests for error propagation in async passthrough streaming routes.
 
-Verifies that HTTP 4xx/5xx errors from upstream (e.g. Azure 429 rate limits)
-raise exceptions instead of being silently forwarded as raw bytes under HTTP 200.
-
-See: litellm/passthrough/main.py _async_streaming()
+Verifies that streaming passthrough wrappers preserve the previous guarantees:
+HTTP 4xx/5xx failures must raise instead of being silently forwarded as bytes,
+and successful streaming responses should still yield chunks normally.
 """
 
+import asyncio
 import json
 from unittest.mock import AsyncMock, MagicMock
 
@@ -54,19 +54,19 @@ def _make_mock_logging_obj():
 @pytest.mark.asyncio
 async def test_async_streaming_429_raises():
     """429 from upstream should raise HTTPStatusError, not yield error bytes."""
-    from litellm.passthrough.main import _async_streaming
-
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
+    
     error_body = json.dumps(
         {"error": {"code": "429", "message": "Rate limit exceeded."}}
     ).encode()
     mock_response = _make_mock_response(429, error_body)
-
+    
     async def response_coro():
         return mock_response
-
+    
     chunks = []
     async def _drain():
-        async for chunk in _async_streaming(
+        async for chunk in AsyncPassthroughStreamingResponse(
             response=response_coro(),
             litellm_logging_obj=_make_mock_logging_obj(),
             provider_config=MagicMock(),
@@ -83,45 +83,78 @@ async def test_async_streaming_429_raises():
 @pytest.mark.asyncio
 async def test_async_streaming_500_raises():
     """500 from upstream should also raise, not yield error bytes."""
-    from litellm.passthrough.main import _async_streaming
-
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
+    
     error_body = json.dumps(
         {"error": {"code": "500", "message": "Internal server error"}}
     ).encode()
     mock_response = _make_mock_response(500, error_body)
-
+    
     async def response_coro():
         return mock_response
-
+    
     with pytest.raises(httpx.HTTPStatusError) as exc_info:
-        async for _ in _async_streaming(
+        async for _ in AsyncPassthroughStreamingResponse(
             response=response_coro(),
             litellm_logging_obj=_make_mock_logging_obj(),
             provider_config=MagicMock(),
         ):
             pass
-
+    
     assert exc_info.value.response.status_code == 500
 
 
 @pytest.mark.asyncio
-async def test_async_streaming_200_yields_chunks():
+async def test_async_passthrough_wrapper_200_yields_chunks():
     """Successful 200 streaming responses should continue to work normally."""
-    from litellm.passthrough.main import _async_streaming
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
 
     sse_data = b'data: {"type":"response.created"}\n\ndata: [DONE]\n\n'
     mock_response = _make_mock_response(200, sse_data)
+    mock_logging_obj = _make_mock_logging_obj()
 
     async def response_coro():
         return mock_response
 
-    chunks = []
-    async for chunk in _async_streaming(
+    async_stream = AsyncPassthroughStreamingResponse(
         response=response_coro(),
-        litellm_logging_obj=_make_mock_logging_obj(),
+        litellm_logging_obj=mock_logging_obj,
         provider_config=MagicMock(),
-    ):
+    )
+
+    chunks = []
+    async for chunk in async_stream:
         chunks.append(chunk)
 
+    await asyncio.sleep(0)
+
     assert len(chunks) == 1
     assert b"response.created" in chunks[0]
+    mock_logging_obj.async_flush_passthrough_collected_chunks.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_error_body_readable_after_failed_await():
+    """The upstream error body must stay readable so the proxy can map the real status and message."""
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
+
+    error_body = b'{"message":"model not found"}'
+
+    async def byte_stream():
+        yield error_body
+
+    request = httpx.Request("POST", "https://bedrock.example.com/model/x/converse-stream")
+    response = httpx.Response(400, content=byte_stream(), request=request)
+
+    async def response_coro():
+        return response
+
+    with pytest.raises(httpx.HTTPStatusError) as exc_info:
+        await AsyncPassthroughStreamingResponse(
+            response=response_coro(),
+            litellm_logging_obj=_make_mock_logging_obj(),
+            provider_config=MagicMock(),
+        )
+
+    assert exc_info.value.response.status_code == 400
+    assert await exc_info.value.response.aread() == error_body

tests/test_litellm/passthrough/test_passthrough_main.py

@@ -645,13 +645,14 @@ async def test_allm_passthrough_route_429_streaming_raises():
     Regression test: Azure 429 during streaming must raise HTTPStatusError,
     not be silently forwarded as raw bytes under HTTP 200.
 
-    Before the fix, _async_streaming() would yield the 429 error JSON as
-    chunks and allm_passthrough_route returned an async generator.  The
-    caller (azure_proxy_route) wrapped it in StreamingResponse(status_code=200),
+    Before the fix, the async passthrough streaming path would yield the 429
+    error JSON as chunks and allm_passthrough_route returned a streaming
+    iterator. The caller (azure_proxy_route) wrapped it in
+    StreamingResponse(status_code=200),
     so the client saw HTTP 200 + unparseable SSE body → silent task_complete(null).
 
-    After the fix, raise_for_status() fires inside _async_streaming() before
-    any chunks are yielded, so the exception propagates all the way up.
+    After the fix, raise_for_status() fires before the streaming wrapper is
+    returned, so the exception propagates all the way up.
     """
     mock_provider_config = MagicMock()
     mock_provider_config.get_complete_url.return_value = (
@@ -679,6 +680,7 @@ async def test_allm_passthrough_route_429_streaming_raises():
     mock_logging_obj = MagicMock()
     mock_logging_obj.update_environment_variables = MagicMock()
     mock_logging_obj.async_flush_passthrough_collected_chunks = AsyncMock()
+    mock_logging_obj.async_failure_handler = AsyncMock()
 
     with (
         patch(
@@ -701,29 +703,101 @@ async def test_allm_passthrough_route_429_streaming_raises():
         patch.object(async_client.client, "send", mock_send),
         patch.object(async_client.client, "build_request", mock_build_request),
     ):
-        result = await allm_passthrough_route(
-            model="azure/gpt-4",
-            endpoint="openai/deployments/gpt-4/responses",
-            method="POST",
-            custom_llm_provider="azure",
-            api_base="https://my-azure.openai.azure.com",
-            api_key="fake-azure-key",
-            json={"model": "gpt-4", "input": "hello", "stream": True},
-            client=async_client,
-            litellm_logging_obj=mock_logging_obj,
-        )
-
-        # result is an async generator — consuming it must raise, not silently yield error bytes
-        chunks = []
-        async def _drain():
-            async for chunk in result:  # type: ignore[union-attr]
-                chunks.append(chunk)
-
         with pytest.raises(httpx.HTTPStatusError) as exc_info:
-            await _drain()
+            await allm_passthrough_route(
+                model="azure/gpt-4",
+                endpoint="openai/deployments/gpt-4/responses",
+                method="POST",
+                custom_llm_provider="azure",
+                api_base="https://my-azure.openai.azure.com",
+                api_key="fake-azure-key",
+                json={"model": "gpt-4", "input": "hello", "stream": True},
+                client=async_client,
+                litellm_logging_obj=mock_logging_obj,
+            )
 
     assert exc_info.value.response.status_code == 429
-    assert len(chunks) == 0, "No chunks should be yielded before the 429 raises"
+
+
+def test_llm_passthrough_route_sync_streaming_error_maps_upstream_status():
+    """
+    Regression test: a sync streaming passthrough whose upstream answers an
+    error status must surface the mapped provider error, not
+    httpx.ResponseNotRead.
+
+    Before the fix, raise_for_status() raised on the still-unread streamed
+    response, and _handle_error then touched e.response.text, which raises
+    ResponseNotRead on a streamed-but-unread body, masking the real upstream
+    error entirely.
+    """
+    from litellm.llms.base_llm.chat.transformation import BaseLLMException
+
+    error_body = json.dumps(
+        {
+            "error": {
+                "code": "429",
+                "message": "Rate limit exceeded. Retry after 10 seconds.",
+            }
+        }
+    ).encode()
+
+    class _UnreadErrorStream(httpx.SyncByteStream):
+        def __iter__(self):
+            yield error_body
+
+    def _handler(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(
+            429,
+            stream=_UnreadErrorStream(),
+            headers={"content-type": "application/json"},
+        )
+
+    sync_client = HTTPHandler(
+        client=httpx.Client(transport=httpx.MockTransport(_handler))
+    )
+
+    mock_provider_config = MagicMock()
+    mock_provider_config.get_complete_url.return_value = (
+        httpx.URL("https://gigachat.devices.sberbank.ru/api/v1/chat/completions"),
+        "https://gigachat.devices.sberbank.ru/api/v1",
+    )
+    mock_provider_config.get_api_key.return_value = "fake-key"
+    mock_provider_config.validate_environment.return_value = {
+        "Authorization": "Bearer fake-key"
+    }
+    mock_provider_config.sign_request.return_value = (
+        {"Authorization": "Bearer fake-key"},
+        None,
+    )
+    mock_provider_config.is_streaming_request.return_value = True
+    mock_provider_config.get_error_class.side_effect = (
+        lambda error_message, status_code, headers: BaseLLMException(
+            status_code=status_code, message=error_message, headers=headers
+        )
+    )
+
+    mock_logging_obj = MagicMock()
+
+    with pytest.raises(BaseLLMException) as exc_info:
+        llm_passthrough_route(
+            model="gigachat/GigaChat-2",
+            endpoint="chat/completions",
+            method="POST",
+            custom_llm_provider="gigachat",
+            api_base="https://gigachat.devices.sberbank.ru/api/v1",
+            api_key="fake-key",
+            json={
+                "model": "GigaChat-2",
+                "messages": [{"role": "user", "content": "hi"}],
+                "stream": True,
+            },
+            client=sync_client,
+            litellm_logging_obj=mock_logging_obj,
+            provider_config=mock_provider_config,
+        )
+
+    assert exc_info.value.status_code == 429
+    assert "Rate limit exceeded" in str(exc_info.value)
 
 
 def test_llm_passthrough_route_propagates_allm_passthrough_route_to_logging_obj():

tests/test_litellm/passthrough/test_streaming_interrupt_spend_tracking.py

@@ -35,11 +35,14 @@ class _ImmediateExecutor:
 
 
 @pytest.mark.asyncio
-async def test_async_streaming_flushes_on_normal_completion():
-    from litellm.passthrough.main import _async_streaming
+async def test_asyncpassthroughstreamingresponse_flushes_on_normal_completion():
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
 
     chunks = [b"chunk-1", b"chunk-2", b"chunk-3"]
     mock_response = _make_streaming_response(chunks)
+    mock_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream", "x-request-id": "req-123"}
+    )
 
     async def response_coro():
         return mock_response
@@ -48,14 +51,19 @@ async def test_async_streaming_flushes_on_normal_completion():
     provider_config = MagicMock()
 
     received = []
-    async for chunk in _async_streaming(
+    received_response = AsyncPassthroughStreamingResponse(
         response=response_coro(),
         litellm_logging_obj=mock_logging_obj,
         provider_config=provider_config,
-    ):
+    )
+
+    async for chunk in received_response:
         received.append(chunk)
 
     assert received == chunks
+    
+    assert received_response.headers["content-type"] == "application/octet-stream"
+    assert received_response.headers["x-request-id"] == "req-123"
 
     await asyncio.sleep(0)
 
@@ -68,8 +76,8 @@ async def test_async_streaming_flushes_on_normal_completion():
 
 
 @pytest.mark.asyncio
-async def test_async_streaming_flushes_on_client_disconnect():
-    from litellm.passthrough.main import _async_streaming
+async def test_asyncpassthroughstreamingresponse_flushes_on_client_disconnect():
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
 
     chunks = [
         b'{"chunk": 1, "outputTokens": 10}',
@@ -77,6 +85,9 @@ async def test_async_streaming_flushes_on_client_disconnect():
         b'{"chunk": 3, "outputTokens": 8}',
     ]
     mock_response = _make_streaming_response(chunks)
+    mock_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream", "x-request-id": "req-123"}
+    )
 
     async def response_coro():
         return mock_response
@@ -84,7 +95,7 @@ async def test_async_streaming_flushes_on_client_disconnect():
     mock_logging_obj = _make_logging_obj()
     provider_config = MagicMock()
 
-    gen = _async_streaming(
+    gen = AsyncPassthroughStreamingResponse(
         response=response_coro(),
         litellm_logging_obj=mock_logging_obj,
         provider_config=provider_config,
@@ -105,11 +116,14 @@ async def test_async_streaming_flushes_on_client_disconnect():
 
 
 @pytest.mark.asyncio
-async def test_async_streaming_does_not_flush_on_4xx():
-    from litellm.passthrough.main import _async_streaming
+async def test_asyncpassthroughstreamingresponse_does_not_flush_on_4xx():
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
 
     err_response = MagicMock(spec=httpx.Response)
     err_response.status_code = 429
+    err_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream"}
+    )
 
     def _raise():
         raise httpx.HTTPStatusError(
@@ -129,7 +143,7 @@ async def test_async_streaming_does_not_flush_on_4xx():
     mock_logging_obj = _make_logging_obj()
 
     with pytest.raises(httpx.HTTPStatusError):
-        async for _ in _async_streaming(
+        async for _ in AsyncPassthroughStreamingResponse(
             response=response_coro(),
             litellm_logging_obj=mock_logging_obj,
             provider_config=MagicMock(),
@@ -140,8 +154,8 @@ async def test_async_streaming_does_not_flush_on_4xx():
 
 
 @pytest.mark.asyncio
-async def test_async_streaming_flushes_on_upstream_exception_with_partial_data():
-    from litellm.passthrough.main import _async_streaming
+async def test_asyncpassthroughstreamingresponse_flushes_on_upstream_exception_with_partial_data():
+    from litellm.passthrough.main import AsyncPassthroughStreamingResponse
 
     partial_chunks = [b"partial-chunk-1", b"partial-chunk-2"]
 
@@ -149,6 +163,9 @@ async def test_async_streaming_flushes_on_upstream_exception_with_partial_data()
     mock_response.status_code = 200
     mock_response.raise_for_status = MagicMock(return_value=None)
     mock_response.aclose = AsyncMock()
+    mock_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream", "x-request-id": "req-123"}
+    )
 
     async def _aiter_bytes_then_raise():
         for c in partial_chunks:
@@ -165,7 +182,7 @@ async def test_async_streaming_flushes_on_upstream_exception_with_partial_data()
 
     received = []
     async def _drain():
-        async for chunk in _async_streaming(
+        async for chunk in AsyncPassthroughStreamingResponse(
             response=response_coro(),
             litellm_logging_obj=mock_logging_obj,
             provider_config=provider_config,
@@ -186,12 +203,16 @@ async def test_async_streaming_flushes_on_upstream_exception_with_partial_data()
     assert call_kwargs["raw_bytes"] == partial_chunks
 
 
-def test_sync_streaming_flushes_on_normal_completion():
-    from litellm.passthrough.main import _sync_streaming
+def test_passthroughstreamingresponse_flushes_on_normal_completion():
+    from litellm.passthrough.main import PassthroughStreamingResponse
 
     chunks = [b"a", b"b", b"c"]
 
     mock_response = MagicMock(spec=httpx.Response)
+    mock_response.status_code = 200
+    mock_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream", "x-request-id": "req-123"}
+    )
 
     def _iter_bytes():
         yield from chunks
@@ -202,25 +223,33 @@ def test_sync_streaming_flushes_on_normal_completion():
     mock_logging_obj.flush_passthrough_collected_chunks = MagicMock()
     provider_config = MagicMock()
 
+    received_responce = PassthroughStreamingResponse(
+        response=mock_response,
+        litellm_logging_obj=mock_logging_obj,
+        provider_config=provider_config,
+    )
+
     with patch("litellm.utils.executor", _ImmediateExecutor()):
-        received = list(
-            _sync_streaming(
-                response=mock_response,
-                litellm_logging_obj=mock_logging_obj,
-                provider_config=provider_config,
-            )
-        )
+        received = list(received_responce)
 
     assert received == chunks
+
+    assert received_responce.headers["content-type"] == "application/octet-stream"
+    assert received_responce.headers["x-request-id"] == "req-123"
+
     mock_logging_obj.flush_passthrough_collected_chunks.assert_called_once()
 
 
-def test_sync_streaming_flushes_on_early_close():
-    from litellm.passthrough.main import _sync_streaming
+def test_passthroughstreamingresponse_flushes_on_early_close():
+    from litellm.passthrough.main import PassthroughStreamingResponse
 
     chunks = [b"first", b"second", b"third"]
 
     mock_response = MagicMock(spec=httpx.Response)
+    mock_response.status_code = 200
+    mock_response.headers = httpx.Headers(
+        {"content-type": "application/octet-stream", "x-request-id": "req-123"}
+    )
 
     def _iter_bytes():
         yield from chunks
@@ -232,7 +261,7 @@ def test_sync_streaming_flushes_on_early_close():
     provider_config = MagicMock()
 
     with patch("litellm.utils.executor", _ImmediateExecutor()):
-        gen = _sync_streaming(
+        gen = PassthroughStreamingResponse(
             response=mock_response,
             litellm_logging_obj=mock_logging_obj,
             provider_config=provider_config,

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

@@ -35,6 +35,22 @@ def mock_mcp_client_ip():
         yield
 
 
+@pytest.fixture(autouse=True)
+def isolate_global_mcp_registry():
+    """Restore the module-global MCP server registry after each test.
+
+    Tests here register servers on ``global_mcp_server_manager`` directly; without a
+    restore, entries leak into other test modules sharing the same worker and break
+    assertions over the full registry contents.
+    """
+    from litellm.proxy._experimental.mcp_server.mcp_server_manager import global_mcp_server_manager
+
+    snapshot = dict(global_mcp_server_manager.registry)
+    yield
+    global_mcp_server_manager.registry.clear()
+    global_mcp_server_manager.registry.update(snapshot)
+
+
 def _mock_callback_request(base_url: str = "http://localhost:3000/"):
     """Return a MagicMock Request for callback/authorize same-origin tests.
 

tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py

@@ -8312,15 +8312,17 @@ async def test_key_does_not_override_explicit_budget_duration():
 @patch(
     "litellm.proxy.management_endpoints.key_management_endpoints.rotate_mcp_server_credentials_master_key"
 )
-async def test_rotate_master_key_model_data_valid_for_prisma(
+async def test_rotate_master_key_reencrypts_model_params_in_place(
     mock_rotate_mcp,
 ):
     """
-    Test that _rotate_master_key produces valid data for Prisma create_many().
-
-    Regression test for: master key rotation fails with Prisma validation error
-    because created_at/updated_at are None (non-nullable DateTime) and
-    litellm_params/model_info are JSON strings (create_many expects dicts).
+    Regression test for: master key rotation wipes every non-credential column
+    on LiteLLM_ProxyModelTable. Rotation used to rebuild the table via
+    delete_many + create_many from Deployment objects, which carry no
+    blocked/created_at/created_by/updated_at/updated_by, so every rotation
+    reset blocked to False (silently unblocking blocked models) and rewrote the
+    audit columns. Rotation must instead update only litellm_params (the sole
+    encrypted column) on each existing row, keyed by model_id.
     """
     from unittest.mock import AsyncMock, MagicMock
 
@@ -8352,6 +8354,7 @@ async def test_rotate_master_key_model_data_valid_for_prisma(
     mock_tx.litellm_proxymodeltable = MagicMock()
     mock_tx.litellm_proxymodeltable.delete_many = AsyncMock()
     mock_tx.litellm_proxymodeltable.create_many = AsyncMock()
+    mock_tx.litellm_proxymodeltable.update_many = AsyncMock()
     mock_prisma_client.db.tx = MagicMock(
         return_value=AsyncMock(
             __aenter__=AsyncMock(return_value=mock_tx),
@@ -8400,36 +8403,33 @@ async def test_rotate_master_key_model_data_valid_for_prisma(
             new_master_key="sk-new-master-key",
         )
 
-    # Verify create_many was called
-    mock_tx.litellm_proxymodeltable.create_many.assert_called_once()
+    # Rotation must never rewrite whole rows: no delete + recreate
+    mock_tx.litellm_proxymodeltable.delete_many.assert_not_called()
+    mock_tx.litellm_proxymodeltable.create_many.assert_not_called()
 
-    # Get the data passed to create_many
-    call_args = mock_tx.litellm_proxymodeltable.create_many.call_args
-    created_models = call_args.kwargs.get("data") or call_args[1].get("data")
+    mock_tx.litellm_proxymodeltable.update_many.assert_called_once()
+    call_args = mock_tx.litellm_proxymodeltable.update_many.call_args
 
-    assert len(created_models) == 1
-    model_data = created_models[0]
+    assert call_args.kwargs["where"] == {
+        "model_id": "model-1"
+    }, "the re-encrypted params must land on the same row, keyed by model_id"
 
-    # Verify timestamps are NOT present (Prisma @default(now()) should apply)
-    assert (
-        "created_at" not in model_data
-    ), "created_at should be excluded so Prisma @default(now()) applies"
-    assert (
-        "updated_at" not in model_data
-    ), "updated_at should be excluded so Prisma @default(now()) applies"
+    update_data = call_args.kwargs["data"]
+    assert set(update_data.keys()) == {"litellm_params"}, (
+        "rotation must touch only the encrypted litellm_params column; writing any "
+        f"other column wipes it (blocked, audit columns), got {sorted(update_data.keys())}"
+    )
 
-    # Verify litellm_params and model_info are prisma.Json wrappers, NOT JSON strings
     import prisma
 
     assert isinstance(
-        model_data["litellm_params"], prisma.Json
-    ), f"litellm_params should be prisma.Json for create_many(), got {type(model_data['litellm_params'])}"
-    assert isinstance(
-        model_data["model_info"], prisma.Json
-    ), f"model_info should be prisma.Json for create_many(), got {type(model_data['model_info'])}"
-
-    # Verify delete_many was called inside the transaction (before create_many)
-    mock_tx.litellm_proxymodeltable.delete_many.assert_called_once()
+        update_data["litellm_params"], prisma.Json
+    ), f"litellm_params should be prisma.Json for update_many(), got {type(update_data['litellm_params'])}"
+    reencrypted_params = update_data["litellm_params"].data
+    assert set(reencrypted_params.keys()) >= {"model", "api_key"}
+    assert (
+        reencrypted_params["api_key"] != "sk-decrypted-key"
+    ), "api_key must be stored re-encrypted under the new master key, not in plaintext"
 
 
 async def test_default_key_generate_params_duration(monkeypatch):

tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py

@@ -4466,3 +4466,65 @@ class TestEnforceRpmTpmOnModelAdd:
             _raise_if_rate_limits_required_but_missing(litellm_params=params, enforced=True)
         assert expected_missing in str(exc_info.value.message)
         assert exc_info.value.code == "400"
+
+
+class TestBlockModelResponseSerialization:
+    @pytest.mark.parametrize(
+        ("route", "blocked"), [("/model/block", True), ("/model/unblock", False)]
+    )
+    def test_block_routes_serialize_prisma_row_to_200(self, route, blocked):
+        from datetime import datetime, timezone
+
+        from prisma import models as prisma_models
+
+        import litellm.proxy.proxy_server as ps
+        from litellm.proxy.proxy_server import app
+
+        written_at = datetime(2026, 8, 29, tzinfo=timezone.utc)
+        row_fields = {
+            "model_id": "m-block-1",
+            "model_name": "gpt-4o-mini",
+            "litellm_params": json.dumps({"model": "openai/gpt-4o-mini", "api_key": "encrypted-value"}),
+            "model_info": json.dumps({"id": "m-block-1"}),
+            "created_at": written_at,
+            "created_by": "admin",
+            "updated_at": written_at,
+            "updated_by": "admin",
+        }
+        existing_row = prisma_models.LiteLLM_ProxyModelTable(blocked=not blocked, **row_fields)
+        updated_row = prisma_models.LiteLLM_ProxyModelTable(blocked=blocked, **row_fields)
+
+        mock_prisma = MagicMock()
+        mock_prisma.db.litellm_proxymodeltable.find_unique = AsyncMock(return_value=existing_row)
+        mock_prisma.db.litellm_proxymodeltable.update = AsyncMock(return_value=updated_row)
+
+        admin = UserAPIKeyAuth(user_id="admin", user_role=LitellmUserRoles.PROXY_ADMIN)
+        app.dependency_overrides[ps.user_api_key_auth] = lambda: admin
+        try:
+            with (
+                patch("litellm.proxy.proxy_server.prisma_client", mock_prisma),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+                patch("litellm.proxy.proxy_server.store_model_in_db", True),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+                patch(  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+                    "litellm.proxy.proxy_server.llm_router",
+                    MagicMock(**{"get_model_ids.return_value": ["m-block-1"]}),
+                ),
+                patch("litellm.proxy.proxy_server.redis_usage_cache", None),  # test-quality-ok: proxy_server module global is the endpoint's only injection point
+                patch(  # test-quality-ok: stubs the cache write so the test observes only response serialization
+                    "litellm.proxy.management_endpoints.model_management_endpoints.clear_cache",
+                    new=AsyncMock(return_value=ReconcileOutcome(still_desired=None, live_after=None)),
+                ),
+                patch(  # test-quality-ok: audit logging is a background side effect outside this test's contract
+                    "litellm.proxy.management_endpoints.model_management_endpoints.create_object_audit_log",
+                    new=AsyncMock(return_value=None),
+                ),
+            ):
+                client = TestClient(app)
+                response = client.post(route, json={"model_id": "m-block-1"})
+        finally:
+            app.dependency_overrides.pop(ps.user_api_key_auth, None)
+
+        assert response.status_code == 200, response.text
+        body = response.json()
+        assert body["model_id"] == "m-block-1"
+        assert body["blocked"] is blocked
+        assert body["litellm_params"] == {"model": "openai/gpt-4o-mini", "api_key": "encrypted-value"}

tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py

@@ -12,11 +12,13 @@ from unittest.mock import AsyncMock, MagicMock, Mock, patch
 import httpx
 import pytest
 from fastapi import HTTPException, Request, Response
+from fastapi.responses import StreamingResponse
 from fastapi.testclient import TestClient
 from starlette.datastructures import FormData
 
 
 import litellm
+from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
 from litellm.constants import LITELLM_PROXY_MASTER_KEY_ALIAS
 from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
     BaseOpenAIPassThroughHandler,
@@ -30,6 +32,7 @@ from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
     get_azure_ai_search_index_from_endpoint,
     get_vertex_base_url,
     is_azure_ai_search_service_level_index_create,
+    gigachat_proxy_route,
     llm_passthrough_factory_proxy_route,
     milvus_proxy_route,
     mistral_proxy_route,
@@ -178,7 +181,7 @@ class TestBaseOpenAIPassThroughHandler:
             assert result["api-key"] == "test_api_key"
             assert result["test-header"] == "value"
 
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.create_pass_through_route"
     )
     async def test_base_openai_pass_through_handler(self, mock_create_pass_through):
@@ -2022,15 +2025,15 @@ class TestLLMPassthroughFactoryProxyRoute:
 
 class TestVLLMProxyRoute:
     @pytest.mark.asyncio
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.get_request_body",
         return_value={"model": "router-model", "stream": False},
     )
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_passthrough_request_using_router_model",
         return_value=True,
     )
-    @patch("litellm.proxy.proxy_server.llm_router")
+    @patch("litellm.proxy.proxy_server.llm_router")  # test-quality-ok: patching litellm internal for unit test isolation
     async def test_vllm_proxy_route_with_router_model(
         self, mock_llm_router, mock_is_router, mock_get_body
     ):
@@ -2055,15 +2058,15 @@ class TestVLLMProxyRoute:
         mock_llm_router.allm_passthrough_route.assert_awaited_once()
 
     @pytest.mark.asyncio
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.get_request_body",
         return_value={"model": "other-model"},
     )
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_passthrough_request_using_router_model",
         return_value=False,
     )
-    @patch(
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
         "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.llm_passthrough_factory_proxy_route"
     )
     async def test_vllm_proxy_route_fallback_to_factory(
@@ -2085,6 +2088,312 @@ class TestVLLMProxyRoute:
         mock_factory_route.assert_awaited_once()
 
 
+class TestGigachatProxyRoute:
+    @pytest.mark.asyncio
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.get_request_body",
+        return_value={"model": "router-model", "stream": False},
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_passthrough_request_using_router_model",
+        return_value=True,
+    )
+    @patch("litellm.proxy.proxy_server.llm_router")  # test-quality-ok: patching litellm internal for unit test isolation
+    async def test_gigachat_proxy_route_with_router_model(
+        self, mock_llm_router, mock_is_router, mock_get_body
+    ):
+        mock_request = MagicMock(spec=Request)
+        mock_request.method = "POST"
+        mock_request.headers = {"content-type": "application/json"}
+        mock_request.query_params = {}
+        mock_fastapi_response = MagicMock(spec=Response)
+        mock_user_api_key_dict = MagicMock()
+        mock_llm_router.allm_passthrough_route = AsyncMock(
+            return_value=httpx.Response(200, json={"response": "success"})
+        )
+
+        result = await gigachat_proxy_route(
+            endpoint="/chat/completions",
+            request=mock_request,
+            fastapi_response=mock_fastapi_response,
+            user_api_key_dict=mock_user_api_key_dict,
+        )
+
+        mock_is_router.assert_called_once()
+        mock_llm_router.allm_passthrough_route.assert_awaited_once()
+        assert isinstance(result, Response)
+
+    @pytest.mark.asyncio
+    async def test_gigachat_router_handler_keeps_cached_body_and_payload_metadata_pristine(self):
+        """Regression: auth-metadata injection must not leak into the cached parsed body or the upstream payload."""
+        from litellm.proxy.common_utils.http_parsing_utils import get_request_body
+        from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
+            handle_gigachat_passthrough_router_model,
+        )
+
+        body = json.dumps(
+            {
+                "model": "gigachat-router",
+                "messages": [{"role": "user", "content": "hi"}],
+                "metadata": {"client_tag": "user-supplied"},
+            }
+        ).encode()
+        scope = {
+            "type": "http",
+            "method": "POST",
+            "headers": [(b"content-type", b"application/json")],
+            "query_string": b"",
+            "path": "/gigachat/chat/completions",
+        }
+
+        async def receive():
+            return {"type": "http.request", "body": body, "more_body": False}
+
+        request = Request(scope, receive)
+        request_body = await get_request_body(request)
+
+        captured: dict = {}
+
+        class _CapturingProcessor:
+            def __init__(self, data: dict):
+                captured["data"] = data
+
+            async def base_passthrough_process_llm_request(self, **kwargs):
+                return Response(content=b"{}", status_code=200)
+
+        with patch(  # test-quality-ok: patching litellm internal for unit test isolation
+            "litellm.proxy.common_request_processing.ProxyBaseLLMRequestProcessing",
+            _CapturingProcessor,
+        ):
+            await handle_gigachat_passthrough_router_model(
+                model="gigachat-router",
+                endpoint="/chat/completions",
+                request=request,
+                request_body=request_body,
+                fastapi_response=Response(),
+                llm_router=MagicMock(),
+                user_api_key_dict=UserAPIKeyAuth(user_id="user-1", team_id="team-1"),
+                proxy_logging_obj=MagicMock(),
+                general_settings={},
+                proxy_config=MagicMock(),
+                select_data_generator=MagicMock(),
+                user_model=None,
+                user_temperature=None,
+                user_request_timeout=None,
+                user_max_tokens=None,
+                user_api_base=None,
+                version=None,
+            )
+
+        data = captured["data"]
+        assert data["json"] is request_body
+        assert request_body["metadata"] == {"client_tag": "user-supplied"}
+        assert data["metadata"]["client_tag"] == "user-supplied"
+        assert data["metadata"]["user_api_key_user_id"] == "user-1"
+        assert data["metadata"]["user_api_key_team_id"] == "team-1"
+        cached_reread = await get_request_body(request)
+        assert cached_reread["metadata"] == {"client_tag": "user-supplied"}
+
+    @pytest.mark.asyncio
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.get_request_body",
+        return_value={"model": "other-model"},
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_passthrough_request_using_router_model",
+        return_value=False,
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_streaming_request_fn",
+        new_callable=AsyncMock,
+        return_value=False,
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.authenticator.get_access_token",
+        return_value="gigachat-test-token",
+    )
+    async def test_gigachat_proxy_route_fallback_forwards_to_gigachat_api(
+        self,
+        mock_get_token,
+        mock_is_streaming,
+        mock_is_router,
+        mock_get_body,
+        monkeypatch,
+    ):
+        monkeypatch.delenv("GIGACHAT_API_BASE", raising=False)
+        mock_request = MagicMock(spec=Request)
+        mock_fastapi_response = MagicMock(spec=Response)
+        mock_user_api_key_dict = MagicMock()
+
+        captured_kwargs = {}
+
+        async def fake_endpoint(request, fastapi_response, user_api_key_dict):
+            return Response(content=b'{"response": "success"}', status_code=200)
+
+        def fake_create_pass_through_route(**kwargs):
+            captured_kwargs.update(kwargs)
+            return fake_endpoint
+
+        with patch(  # test-quality-ok: patching litellm internal for unit test isolation
+            "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.create_pass_through_route",
+            side_effect=fake_create_pass_through_route,
+        ):
+            result = await gigachat_proxy_route(
+                endpoint="/chat/completions",
+                request=mock_request,
+                fastapi_response=mock_fastapi_response,
+                user_api_key_dict=mock_user_api_key_dict,
+            )
+
+        assert isinstance(result, Response)
+        assert result.status_code == 200
+        assert captured_kwargs["target"] == "https://gigachat.devices.sberbank.ru/api/v1/chat/completions"
+        assert captured_kwargs["custom_headers"] == {"Authorization": "Bearer gigachat-test-token"}
+
+    @pytest.mark.asyncio
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.get_request_body",
+        return_value={},
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.is_streaming_request_fn",
+        new_callable=AsyncMock,
+        return_value=False,
+    )
+    @patch(  # test-quality-ok: patching litellm internal for unit test isolation
+        "litellm.llms.gigachat.authenticator.get_access_token",
+        return_value="gigachat-test-token",
+    )
+    async def test_gigachat_proxy_route_models_endpoint_without_model(
+        self,
+        mock_get_token,
+        mock_is_streaming,
+        mock_get_body,
+        monkeypatch,
+    ):
+        monkeypatch.delenv("GIGACHAT_API_BASE", raising=False)
+        mock_request = MagicMock(spec=Request)
+        mock_fastapi_response = MagicMock(spec=Response)
+        mock_user_api_key_dict = MagicMock()
+
+        captured_kwargs = {}
+
+        async def fake_endpoint(request, fastapi_response, user_api_key_dict):
+            return Response(content=b'{"data": []}', status_code=200)
+
+        def fake_create_pass_through_route(**kwargs):
+            captured_kwargs.update(kwargs)
+            return fake_endpoint
+
+        with patch(  # test-quality-ok: patching litellm internal for unit test isolation
+            "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.create_pass_through_route",
+            side_effect=fake_create_pass_through_route,
+        ):
+            result = await gigachat_proxy_route(
+                endpoint="models",
+                request=mock_request,
+                fastapi_response=mock_fastapi_response,
+                user_api_key_dict=mock_user_api_key_dict,
+            )
+
+        assert isinstance(result, Response)
+        assert result.status_code == 200
+        assert captured_kwargs["target"] == "https://gigachat.devices.sberbank.ru/api/v1/models"
+        assert captured_kwargs["custom_headers"] == {"Authorization": "Bearer gigachat-test-token"}
+
+    @pytest.mark.asyncio
+    async def test_allm_passthrough_streaming_preserves_upstream_headers(self):
+        async def _stream() -> bytes:
+            yield b'data: {"id":"1"}\n\n'
+
+        class MockPassthroughStreamingResponse:
+            def __init__(self):
+                self.status_code = 201
+                self.headers = {
+                    "content-type": "text/event-stream; charset=utf-8",
+                    "x-request-id": "req-123",
+                    "x-ratelimit-remaining-requests": "77",
+                    "transfer-encoding": "chunked",
+                    "content-encoding": "gzip",
+                }
+                self._iterator = _stream()
+
+            def __aiter__(self):
+                return self
+
+            async def __anext__(self):
+                return await self._iterator.__anext__()
+
+        processor = ProxyBaseLLMRequestProcessing(
+            data={
+                "model": "some-provider/model",
+                "stream": True,
+                "litellm_call_id": "call-123",
+                "litellm_logging_obj": MagicMock(litellm_call_id="call-123"),
+            }
+        )
+
+        mock_request = MagicMock(spec=Request)
+        mock_request.headers = {"content-type": "application/json"}
+        mock_fastapi_response = MagicMock(spec=Response)
+        mock_user_api_key_dict = MagicMock()
+        mock_user_api_key_dict.allowed_model_region = ""
+        mock_user_api_key_dict.spend = 0.0
+        mock_proxy_logging_obj = MagicMock()
+        mock_proxy_logging_obj.during_call_hook = AsyncMock(return_value=None)
+        mock_proxy_logging_obj.update_request_status = AsyncMock(return_value=None)
+        mock_proxy_logging_obj.post_call_response_headers_hook = AsyncMock(
+            return_value={"x-test-callback-header": "callback-value"}
+        )
+
+        streaming_response = MockPassthroughStreamingResponse()
+
+        async def _fake_route_request(*args, **kwargs):
+            async def _inner():
+                return streaming_response
+
+            return _inner()
+
+        with patch.object(
+            processor,
+            "common_processing_pre_call_logic",
+            new=AsyncMock(
+                return_value=(
+                    processor.data,
+                    processor.data["litellm_logging_obj"],
+                )
+            ),
+        ), patch(  # test-quality-ok: patching litellm internal for unit test isolation
+            "litellm.proxy.common_request_processing.route_request",
+            new=_fake_route_request,
+        ), patch(  # test-quality-ok: patching litellm internal for unit test isolation
+            "litellm.proxy.common_request_processing.ProxyBaseLLMRequestProcessing.get_custom_headers",
+            return_value={"x-litellm-call-id": "call-123"},
+        ):
+            result = await processor.base_passthrough_process_llm_request(
+                request=mock_request,
+                fastapi_response=mock_fastapi_response,
+                user_api_key_dict=mock_user_api_key_dict,
+                proxy_logging_obj=mock_proxy_logging_obj,
+                general_settings={},
+                proxy_config=MagicMock(),
+                select_data_generator=MagicMock(),
+                llm_router=None,
+                model="some-provider/model",
+                version="test-version",
+            )
+
+        assert isinstance(result, StreamingResponse)
+        assert result.status_code == 201
+        assert result.headers["content-type"] == "text/event-stream; charset=utf-8"
+        assert result.headers["x-request-id"] == "req-123"
+        assert result.headers["x-ratelimit-remaining-requests"] == "77"
+        assert result.headers["x-litellm-call-id"] == "call-123"
+        assert result.headers["x-test-callback-header"] == "callback-value"
+        assert "transfer-encoding" not in result.headers
+        assert "content-encoding" not in result.headers
+
+
 class TestForwardHeaders:
     """
     Test cases for _forward_headers parameter in passthrough endpoints
@@ -4627,3 +4936,76 @@ class TestPassthroughRouterModelBudgetReservation:
         )
 
         self._assert_metadata_carries_attribution(captured, user_api_key_dict)
+
+
+class TestAzureRouterModelStreamingDispatch:
+    """
+    Regression: ``llm_router.allm_passthrough_route`` returns an awaited
+    ``AsyncPassthroughStreamingResponse`` for streaming calls, which is no
+    longer an async generator under ``inspect.isasyncgen``. The dispatch's
+    else branch therefore calls ``.aiter_bytes()`` / ``.status_code`` /
+    ``.headers`` on it. The router's ``set_response_headers`` also runs the
+    result through ``prepare_response_for_header_attachment``, which used to
+    wrap it in ``HiddenParamsAsyncIteratorWrapper`` (no ``aiter_bytes``), so
+    every streaming Azure router-model request 500'd with
+    ``AttributeError: aiter_bytes``; ``_hidden_params`` on the streaming
+    response keeps it unwrapped.
+    """
+
+    @pytest.mark.asyncio
+    async def test_azure_router_model_streaming_returns_streaming_response(self, monkeypatch):
+        import litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints as ep
+        import litellm.proxy.proxy_server as proxy_server
+        from litellm.passthrough.main import AsyncPassthroughStreamingResponse
+
+        upstream_body = b"data: hello\n\n"
+
+        async def _upstream_response() -> httpx.Response:
+            upstream_request = httpx.Request(
+                "POST",
+                "https://my-azure.openai.azure.com/openai/deployments/gpt-5/chat/completions",
+            )
+            return httpx.Response(
+                200,
+                headers={"content-type": "text/event-stream"},
+                content=upstream_body,
+                request=upstream_request,
+            )
+
+        logging_obj = MagicMock()
+        logging_obj.async_flush_passthrough_collected_chunks = AsyncMock()
+
+        from litellm.router_utils.add_retry_fallback_headers import prepare_response_for_header_attachment
+
+        class StreamingRouter:
+            async def allm_passthrough_route(self, **kwargs):
+                streaming_response = await AsyncPassthroughStreamingResponse(
+                    response=_upstream_response(),
+                    litellm_logging_obj=logging_obj,
+                    provider_config=MagicMock(),
+                )
+                return prepare_response_for_header_attachment(streaming_response)
+
+        async def fake_get_request_body(_request):
+            return {"model": "gpt-5", "stream": True}
+
+        monkeypatch.setattr(proxy_server, "llm_router", StreamingRouter())
+        monkeypatch.setattr(ep, "get_request_body", fake_get_request_body)
+        monkeypatch.setattr(ep, "is_passthrough_request_using_router_model", lambda *a, **k: True)
+
+        request = MagicMock(spec=Request)
+        request.method = "POST"
+        request.headers = {"content-type": "application/json"}
+        request.query_params = {}
+
+        result = await azure_proxy_route(
+            endpoint="openai/deployments/gpt-5/chat/completions",
+            request=request,
+            fastapi_response=MagicMock(spec=Response),
+            user_api_key_dict=UserAPIKeyAuth(api_key="hashed-token"),
+        )
+
+        assert isinstance(result, StreamingResponse)
+        assert result.status_code == 200
+        body = b"".join([chunk async for chunk in result.body_iterator])
+        assert body == upstream_body

tests/test_litellm/proxy/test_common_request_processing.py

@@ -4694,7 +4694,7 @@ class TestAllmPassthroughStreamingProviderGate:
         }
         return ProxyBaseLLMRequestProcessing(data=data)
 
-    async def _run(self, processing_obj, monkeypatch, chunks):
+    async def _run(self, processing_obj, monkeypatch, chunks, stream=None):
         import litellm.proxy.common_request_processing as crp
         from litellm.proxy._types import UserAPIKeyAuth as RealUserAPIKeyAuth
 
@@ -4702,9 +4702,11 @@ class TestAllmPassthroughStreamingProviderGate:
             for chunk in chunks:
                 yield chunk
 
+        upstream_stream = stream if stream is not None else streaming_response()
+
         async def fake_route_request(**kwargs):
             async def _llm_call():
-                return streaming_response()
+                return upstream_stream
 
             return _llm_call()
 
@@ -4729,6 +4731,40 @@ class TestAllmPassthroughStreamingProviderGate:
             skip_pre_call_logic=True,
         )
 
+    @pytest.mark.asyncio
+    async def test_client_disconnect_closes_unbuffered_passthrough_stream(self, monkeypatch):
+        """Starlette abandons the body iterator when the client disconnects, so the
+        unbuffered passthrough branch must return _UpstreamClosingStreamingResponse,
+        whose shielded cleanup closes the upstream stream; that close is what flushes
+        buffered passthrough usage into spend logs."""
+        processing_obj = self._build_processing_obj("gigachat")
+        monkeypatch.setattr(litellm, "callbacks", [])
+        upstream_closed = asyncio.Event()
+
+        async def hanging_stream():
+            try:
+                yield b"chunk-1"
+                await asyncio.Event().wait()
+            finally:
+                upstream_closed.set()
+
+        result = await self._run(processing_obj, monkeypatch, [], stream=hanging_stream())
+
+        assert isinstance(result, _UpstreamClosingStreamingResponse)
+
+        first_chunk_sent = asyncio.Event()
+
+        async def receive():
+            await first_chunk_sent.wait()
+            return {"type": "http.disconnect"}
+
+        async def send(message):
+            if message["type"] == "http.response.body" and message.get("body"):
+                first_chunk_sent.set()
+
+        await result({"type": "http"}, receive, send)
+        await asyncio.wait_for(upstream_closed.wait(), timeout=5)
+
     @pytest.mark.asyncio
     async def test_non_bedrock_stream_is_not_buffered(self, monkeypatch):
         processing_obj = self._build_processing_obj("anthropic")

tests/test_litellm/router_strategy/test_complexity_router.py

@@ -4425,6 +4425,265 @@ class _DummyPlugin:
         return context
 
 
+class TestClassificationMode:
+    """Test classification_mode='user_turn': classify only requests whose newest turn is a new
+    human ask; tool-loop continuation turns replay the session's held routing decision."""
+
+    REASONING_ASK = {
+        "role": "user",
+        "content": "Let's think step by step and reason through this problem carefully.",
+    }
+    SIMPLE_ASK = {"role": "user", "content": "Hello!"}
+    ASSISTANT_ANSWER = {"role": "assistant", "content": "the answer"}
+    TOOL_CALL_1 = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read_file", "arguments": "{}"}}],
+    }
+    TOOL_RESULT_1 = {"role": "tool", "tool_call_id": "call_1", "content": "file contents"}
+    TOOL_CALL_2 = {
+        "role": "assistant",
+        "content": None,
+        "tool_calls": [{"id": "call_2", "type": "function", "function": {"name": "run_tests", "arguments": "{}"}}],
+    }
+    TOOL_RESULT_2 = {"role": "tool", "tool_call_id": "call_2", "content": "3 passed"}
+
+    @pytest.fixture
+    def user_turn_config(self, basic_config) -> dict:
+        return {**basic_config, "classification_mode": "user_turn"}
+
+    @staticmethod
+    def _request_kwargs(session_id: str) -> dict:
+        return {"metadata": {"session_id": session_id}}
+
+    def _router(self, mock_router_instance, config: dict) -> ComplexityRouter:
+        mock_router_instance.cache = DualCache()
+        return ComplexityRouter(
+            model_name="test-router",
+            litellm_router_instance=mock_router_instance,
+            complexity_router_config=config,
+        )
+
+    def _tool_loop_turns(self) -> list[list[dict]]:
+        return [
+            [self.REASONING_ASK],
+            [self.REASONING_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1],
+            [self.REASONING_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1, self.TOOL_CALL_2, self.TOOL_RESULT_2],
+        ]
+
+    def test_default_mode_is_every_request(self, complexity_router):
+        assert complexity_router.config.classification_mode == "every_request"
+
+    def test_invalid_classification_mode_rejected(self, mock_router_instance, basic_config):
+        with pytest.raises(ValidationError):
+            ComplexityRouter(
+                model_name="test-router",
+                litellm_router_instance=mock_router_instance,
+                complexity_router_config={**basic_config, "classification_mode": "sometimes"},
+            )
+
+    @pytest.mark.asyncio
+    async def test_user_turn_mode_classifies_tool_loop_once(self, mock_router_instance, user_turn_config):
+        """The mutation check: a 3-request tool loop drives exactly one classification, and both
+        continuation turns hold the classified model under the user_turn_continuation cause."""
+        router = self._router(mock_router_instance, user_turn_config)
+        with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy:
+            responses = [
+                await router.async_pre_routing_hook(
+                    model="test-model", request_kwargs=self._request_kwargs("loop-1"), messages=turn
+                )
+                for turn in self._tool_loop_turns()
+            ]
+        assert spy.call_count == 1
+        assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"]
+        assert [r.routing_decision["cause"] for r in responses[1:]] == [
+            "user_turn_continuation",
+            "user_turn_continuation",
+        ]
+
+    @pytest.mark.asyncio
+    async def test_every_request_default_classifies_every_tool_loop_turn(self, mock_router_instance, basic_config):
+        """Pins today's default: every request classifies, including tool-loop continuations."""
+        router = self._router(mock_router_instance, basic_config)
+        with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy:
+            responses = [
+                await router.async_pre_routing_hook(
+                    model="test-model", request_kwargs=self._request_kwargs("loop-2"), messages=turn
+                )
+                for turn in self._tool_loop_turns()
+            ]
+        assert spy.call_count == 3
+        assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"]
+        assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses)
+
+    @pytest.mark.asyncio
+    async def test_continuation_without_session_id_still_classifies(self, mock_router_instance, user_turn_config):
+        """No resolvable session id means no held decision to replay, so every request classifies."""
+        router = self._router(mock_router_instance, user_turn_config)
+        with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy:
+            responses = [
+                await router.async_pre_routing_hook(model="test-model", request_kwargs={}, messages=turn)
+                for turn in self._tool_loop_turns()
+            ]
+        assert spy.call_count == 3
+        assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"]
+        assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses)
+
+    @pytest.mark.asyncio
+    async def test_plugins_suppress_user_turn_gate(self, mock_router_instance, basic_config):
+        """A replayed decision would bypass the plugin pipeline, so plugins force every request
+        through _classify_and_route, exactly as they do for session_affinity."""
+        router = self._router(
+            mock_router_instance,
+            {**basic_config, "classification_mode": "user_turn", "plugins": [_DummyPlugin()]},
+        )
+        with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy:
+            responses = [
+                await router.async_pre_routing_hook(
+                    model="test-model", request_kwargs=self._request_kwargs("loop-3"), messages=turn
+                )
+                for turn in self._tool_loop_turns()
+            ]
+        assert spy.call_count == 3
+        assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"]
+        assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses)
+
+    @pytest.mark.asyncio
+    async def test_new_human_ask_reclassifies_and_repins(self, mock_router_instance, user_turn_config):
+        """Unlike session_affinity, a new human ask never short-circuits on the pin: the session
+        re-classifies, moves tier, and the moved decision becomes the next held decision."""
+        router = self._router(mock_router_instance, user_turn_config)
+        first = await router.async_pre_routing_hook(
+            model="test-model", request_kwargs=self._request_kwargs("s-repin"), messages=[self.REASONING_ASK]
+        )
+        second = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-repin"),
+            messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK],
+        )
+        third = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-repin"),
+            messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1],
+        )
+        assert first.model == "o1-preview"
+        assert second.model == "gpt-4o-mini"
+        assert third.model == "gpt-4o-mini"
+        assert third.routing_decision["cause"] == "user_turn_continuation"
+
+    @pytest.mark.asyncio
+    async def test_new_ask_with_trailing_system_reminder_reclassifies(self, mock_router_instance, user_turn_config):
+        """Claude Code appends a system-role reminder after the human turn; that trailing plumbing
+        must not turn a new ask into a continuation, and a continuation turn carrying the same
+        trailing reminder stays a continuation."""
+        router = self._router(mock_router_instance, user_turn_config)
+        reminder = {"role": "system", "content": "<total_tokens>100 tokens left</total_tokens>"}
+        first = await router.async_pre_routing_hook(
+            model="test-model", request_kwargs=self._request_kwargs("s-reminder"), messages=[self.REASONING_ASK]
+        )
+        second = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-reminder"),
+            messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK, reminder],
+        )
+        third = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-reminder"),
+            messages=[
+                self.REASONING_ASK,
+                self.ASSISTANT_ANSWER,
+                self.SIMPLE_ASK,
+                reminder,
+                self.TOOL_CALL_1,
+                self.TOOL_RESULT_1,
+                reminder,
+            ],
+        )
+        assert first.model == "o1-preview"
+        assert second.model == "gpt-4o-mini"
+        assert second.routing_decision["cause"] != "user_turn_continuation"
+        assert third.model == "gpt-4o-mini"
+        assert third.routing_decision["cause"] == "user_turn_continuation"
+
+    @pytest.mark.asyncio
+    async def test_escalation_keyword_turn_is_a_new_ask(self, mock_router_instance, user_turn_config):
+        """An escalation keyword arrives as human text, so the turn classifies and escalates
+        instead of replaying the held decision."""
+        router = self._router(mock_router_instance, user_turn_config)
+        first = await router.async_pre_routing_hook(
+            model="test-model", request_kwargs=self._request_kwargs("s-esc"), messages=[self.SIMPLE_ASK]
+        )
+        second = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-esc"),
+            messages=[self.SIMPLE_ASK, self.ASSISTANT_ANSWER, {"role": "user", "content": "LITELLM ESCALATE"}],
+        )
+        assert first.model == "gpt-4o-mini"
+        assert second.model == "gpt-4o"
+        assert second.routing_decision["escalated"] is True
+
+    @pytest.mark.asyncio
+    async def test_messages_surface_tool_result_shapes(self, mock_router_instance, user_turn_config):
+        """Messages-surface shapes: a tool_result-only user turn is a continuation, while an ask
+        riding alongside a tool_result in the same turn is a new ask."""
+        router = self._router(mock_router_instance, user_turn_config)
+        tool_use = {"role": "assistant", "content": [{"type": "tool_use", "id": "x", "name": "t", "input": {}}]}
+        tool_result = {"type": "tool_result", "tool_use_id": "x", "content": "ok"}
+        first = await router.async_pre_routing_hook(
+            model="test-model", request_kwargs=self._request_kwargs("s-msgs"), messages=[self.REASONING_ASK]
+        )
+        pure = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-msgs"),
+            messages=[self.REASONING_ASK, tool_use, {"role": "user", "content": [tool_result]}],
+        )
+        hybrid = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-msgs"),
+            messages=[
+                self.REASONING_ASK,
+                tool_use,
+                {"role": "user", "content": [tool_result, {"type": "text", "text": "Hello!"}]},
+            ],
+        )
+        assert first.model == "o1-preview"
+        assert pure.model == "o1-preview"
+        assert pure.routing_decision["cause"] == "user_turn_continuation"
+        assert hybrid.model == "gpt-4o-mini"
+
+    @pytest.mark.asyncio
+    async def test_session_affinity_wins_when_both_knobs_are_on(self, mock_router_instance, user_turn_config):
+        """With session_affinity also on, the pin short-circuits new asks too and keeps its own
+        cause, so the session stays on turn 1's model."""
+        router = self._router(mock_router_instance, {**user_turn_config, "session_affinity": True})
+        first = await router.async_pre_routing_hook(
+            model="test-model", request_kwargs=self._request_kwargs("s-both"), messages=[self.REASONING_ASK]
+        )
+        second = await router.async_pre_routing_hook(
+            model="test-model",
+            request_kwargs=self._request_kwargs("s-both"),
+            messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK],
+        )
+        assert first.model == "o1-preview"
+        assert second.model == "o1-preview"
+        assert second.routing_decision["cause"] == "session_affinity_pin"
+
+    def test_user_turn_mode_enables_tier_and_deployment_pins(self, mock_router_instance, basic_config):
+        """user_turn implies the tier pin machinery (the pin write is what gives a continuation
+        a held decision) and the tier pin implies the deployment pin; plugins suppress both."""
+        default = self._router(mock_router_instance, basic_config)
+        enabled = self._router(mock_router_instance, {**basic_config, "classification_mode": "user_turn"})
+        suppressed = self._router(
+            mock_router_instance,
+            {**basic_config, "classification_mode": "user_turn", "plugins": [_DummyPlugin()]},
+        )
+        assert default._uses_tier_pin is False
+        assert enabled._uses_tier_pin is True
+        assert enabled._uses_deployment_pin is True
+        assert suppressed._uses_tier_pin is False
+        assert suppressed._uses_deployment_pin is False
+
+
 class TestRoutingPlugins:
     """Test the `complexity_router_config.plugins` field: narrows the classified
     tier's candidate pool before a model is picked. Discussion:

tests/test_litellm/test_friendli_glm_5_3_flash_model_metadata.py

@@ -0,0 +1,35 @@
+import json
+from pathlib import Path
+
+from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
+
+
+def test_friendli_glm_5_3_flash_model_info():
+    model = "friendliai/zai-org/GLM-5.3-Flash"
+    json_path = Path(__file__).parents[2] / "model_prices_and_context_window.json"
+    with open(json_path) as f:
+        model_cost = json.load(f)
+
+    info = model_cost.get(model)
+    assert (
+        info is not None
+    ), f"{model} not found in model_prices_and_context_window.json"
+    assert info["litellm_provider"] == "friendliai"
+    assert info["mode"] == "chat"
+    assert info["input_cost_per_token"] == 1.5e-07
+    assert info["output_cost_per_token"] == 5e-07
+    assert info["cache_read_input_token_cost"] == 3e-08
+    assert info["max_input_tokens"] == 1048576
+    assert info["max_output_tokens"] == 1048576
+    assert info["supports_function_calling"] is True
+    assert info["supports_reasoning"] is True
+    assert info["reasoning_effort_levels"] == ["low", "high", "max"]
+    assert info["supports_tool_choice"] is True
+    assert info["supports_prompt_caching"] is True
+    assert info["supports_vision"] is True
+    assert info["supports_image_input"] is True
+    assert info["supports_video_input"] is True
+
+    routed_model, provider, _, _ = get_llm_provider(model=model)
+    assert routed_model == "zai-org/GLM-5.3-Flash"
+    assert provider == "friendliai"

tests/test_litellm/test_friendli_glm_5_3_model_metadata.py

@@ -0,0 +1,34 @@
+import json
+from pathlib import Path
+
+from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider
+
+
+def test_friendli_glm_5_3_model_info():
+    model = "friendliai/zai-org/GLM-5.3"
+    json_path = Path(__file__).parents[2] / "model_prices_and_context_window.json"
+    with open(json_path) as f:
+        model_cost = json.load(f)
+
+    info = model_cost.get(model)
+    assert (
+        info is not None
+    ), f"{model} not found in model_prices_and_context_window.json"
+    assert info["litellm_provider"] == "friendliai"
+    assert info["mode"] == "chat"
+    assert info["input_cost_per_token"] == 1.26e-06
+    assert info["output_cost_per_token"] == 3.96e-06
+    assert info["cache_read_input_token_cost"] == 2.34e-07
+    assert info["max_input_tokens"] == 1048576
+    assert info["max_output_tokens"] == 1048576
+    assert info["supports_function_calling"] is True
+    assert info["supports_reasoning"] is True
+    assert info["reasoning_effort_levels"] == ["low", "high", "max"]
+    assert info["supports_tool_choice"] is True
+    assert info["supports_prompt_caching"] is True
+    assert info["supports_vision"] is False
+    assert info["supports_image_input"] is False
+
+    routed_model, provider, _, _ = get_llm_provider(model=model)
+    assert routed_model == "zai-org/GLM-5.3"
+    assert provider == "friendliai"

tests/test_litellm/test_router.py

@@ -8155,6 +8155,71 @@ class TestUpsertDeploymentRollback:
         assert len(router.model_list) == 1
 
 
+class TestUpsertDeploymentRename:
+    """
+    Issue #38360: renaming a model wrote the new `model_name` to the db, but the reload's
+    `upsert_deployment` compared only `litellm_params` and `model_info`. A rename with no
+    other edit therefore compared equal and the router kept the old name until a restart,
+    so `/model/info` and `/v1/models` served the stale name and the new one was unroutable.
+    """
+
+    @staticmethod
+    def _router() -> "litellm.Router":
+        return litellm.Router(
+            model_list=[
+                {
+                    "model_name": "old-name",
+                    "litellm_params": {"model": "openai/gpt-4o", "api_key": "sk-test"},
+                    "model_info": {"id": "rename-1", "db_model": True},
+                }
+            ]
+        )
+
+    @staticmethod
+    def _deployment(model_name: str, tpm: int | None = None):
+        from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo
+
+        return Deployment(
+            model_name=model_name,
+            litellm_params=LiteLLM_Params(model="openai/gpt-4o", api_key="sk-test", tpm=tpm),
+            model_info=ModelInfo(id="rename-1", db_model=True),
+        )
+
+    def test_rename_only_updates_the_router(self):
+        router = self._router()
+
+        assert router.upsert_deployment(deployment=self._deployment("new-name")) is not None
+
+        assert [model["model_name"] for model in router.model_list] == ["new-name"]
+        renamed = router.get_deployment(model_id="rename-1")
+        assert renamed is not None
+        assert renamed.model_name == "new-name"
+
+    def test_rename_only_makes_the_new_name_routable(self):
+        router = self._router()
+
+        router.upsert_deployment(deployment=self._deployment("new-name"))
+
+        assert router.get_model_ids(model_name="new-name") == ["rename-1"]
+        assert router.get_model_ids(model_name="old-name") == []
+
+    def test_rename_alongside_another_edit_still_updates(self):
+        router = self._router()
+
+        router.upsert_deployment(deployment=self._deployment("new-name", tpm=1234))
+
+        assert router.get_model_ids(model_name="new-name") == ["rename-1"]
+        renamed = router.get_deployment(model_id="rename-1")
+        assert renamed is not None
+        assert renamed.litellm_params.tpm == 1234
+
+    def test_unchanged_deployment_is_still_a_no_op(self):
+        router = self._router()
+
+        assert router.upsert_deployment(deployment=self._deployment("old-name")) is None
+        assert [model["model_name"] for model in router.model_list] == ["old-name"]
+
+
 class TestConsumedRequestTagsStamp:
     """Issue #36621: when a request's tags select a tagged pre-routing strategy, those
     tags are consumed by the selection; the hook must stamp the rewritten model group so

type-discipline-budget.json

@@ -9,7 +9,7 @@
     "limit": 269
   },
   "LIT004": {
-    "limit": 43
+    "limit": 40
   },
   "LIT005": {
     "limit": 0

ui/litellm-dashboard/public/assets/logos/gigachat.svg

@@ -0,0 +1,27 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!-- Generator: Adobe Illustrator 28.0.0, SVG Export Plug-In . SVG Version: 6.00 Build 0)  -->
+<svg version="1.1" id="Слой_1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" x="0px" y="0px"
+	 viewBox="0 0 1000 1000" enable-background="new 0 0 1000 1000" xml:space="preserve">
+<path fill-rule="evenodd" clip-rule="evenodd" fill="#000001" d="M500,0c276.1218262,0,500,223.8403168,500,500
+	c0,276.1593628-223.8781738,500-500,500C223.8771973,1000,0,776.1218262,0,500C0.0000314,223.8771973,223.8771973,0.0000314,500,0z
+	 M911.703125,378.8250122c-12.6343384,87.6312561-57.1812744,159.9468384-94.3469238,207.0031128
+	c-57.359314,70.65625-134.203064,131.828125-222.3842773,177.03125
+	c-87.2250366,44.7218628-181.578186,71.8875122-272.8406677,78.5562744
+	c-15.4475098,0.890625-29.7947083,1.3311768-43.2412415,1.3311768c-16.0025024,0-30.7562714-0.640625-44.759079-1.8999023
+	c73.1806183,57.3218384,165.3878326,91.499939,265.5440674,91.499939l-0.0562744-0.0562744
+	c238.0687561,0,431.1000061-192.999939,431.1000061-431.0686951c0-41.3156433-5.828125-81.2562561-16.6843872-119.0875244
+	C913.2562256,381.0218811,912.4812622,379.8999939,911.703125,378.8250122z M561.3343506,45.6000023
+	c-147.2843323,0-281.1668701,60.73843-358.1281128,162.4078064l-0.4881287,0.5968781
+	c-33.2709351,40.5328217-55.3865662,84.9996796-63.910614,128.6078033
+	c-4.8540802,26.53125-3.9568939,50.8968811,2.7127991,72.1312561v-0.0562439
+	c15.8215637,47.5031128,54.3062439,85.5968628,97.9543762,96.7875061
+	c18.1562653,4.5218506,35.6100006,6.4124756,51.8393707,5.671875l4.2859497-0.3812256
+	C405.9468689,502.4718628,504.34375,432.9312439,586.046875,366.046875
+	c51.1312256-42.7937622,100.0405884-99.4146729,103.1906128-135.9850006
+	C617.9468384,264.6322021,544.3500366,309.6121826,459.5,370.4156189
+	c-9.4125061,6.7062378-22.2406311,4.828125-29.2437439-4.2875061
+	c-42.1281128-54.5405884-78.8125305-98.8321838-115.53125-139.2943726
+	c-3.9968872-4.4449921-5.9906311-10.4062347-5.3987427-16.3302917c0.555603-5.9284515,3.6299744-11.4168854,8.4081116-14.9740753
+	C430.2281189,111.1612473,542.125,63.3568764,650.6875,53.2768745
+	C621.5718994,48.1650009,591.7124634,45.6006279,561.3343506,45.6000023z"/>
+</svg>

ui/litellm-dashboard/src/components/provider_info_helpers.tsx

@@ -20,6 +20,7 @@ import falAiLogo from "../../public/assets/logos/fal_ai.jpg";
 import featherlessLogo from "../../public/assets/logos/featherless.svg";
 import fireworksLogo from "../../public/assets/logos/fireworks.svg";
 import friendliLogo from "../../public/assets/logos/friendli.svg";
+import gigachatLogo from "../../public/assets/logos/gigachat.svg";
 import githubCopilotLogo from "../../public/assets/logos/github_copilot.svg";
 import googleLogo from "../../public/assets/logos/google.svg";
 import groqLogo from "../../public/assets/logos/groq.svg";
@@ -107,6 +108,7 @@ export enum Providers {
   FireworksAI = "Fireworks AI",
   FRIENDLIAI = "Friendliai",
   GALADRIEL = "Galadriel",
+  GIGACHAT = "GigaChat",
   GITHUB_COPILOT = "Github Copilot",
   Google_AI_Studio = "Google AI Studio",
   GradientAI = "GradientAI",
@@ -218,6 +220,7 @@ export const provider_map: Record<string, string> = {
   FireworksAI: "fireworks_ai",
   FRIENDLIAI: "friendliai",
   GALADRIEL: "galadriel",
+  GIGACHAT: "gigachat",
   GITHUB_COPILOT: "github_copilot",
   Google_AI_Studio: "gemini",
   GradientAI: "gradient_ai",
@@ -323,6 +326,7 @@ export const providerLogoMap: Partial<Record<Providers, string>> = {
   [Providers.FEATHERLESS_AI]: featherlessLogo.src,
   [Providers.FireworksAI]: fireworksLogo.src,
   [Providers.FRIENDLIAI]: friendliLogo.src,
+  [Providers.GIGACHAT]: gigachatLogo.src,
   [Providers.GITHUB_COPILOT]: githubCopilotLogo.src,
   [Providers.Google_AI_Studio]: googleLogo.src,
   [Providers.Groq]: groqLogo.src,

ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx

@@ -89,6 +89,7 @@ const CONSTANT_CAUSE_LABELS: Record<string, string> = {
   semantic_keyword_match: "Semantic keyword match",
   session_affinity_pin: "Pinned to session",
   session_affinity_escalation: "Escalated from session pin",
+  user_turn_continuation: "Continuation turn, classifier skipped",
   quality_tier: "Quality tier mapping",
   bandit: "Adaptive bandit",
   default_fallback: "Default model, no route matched",

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -5258,6 +5258,42 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/gigachat/{endpoint}": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        /**
+         * Gigachat Proxy Route
+         * @description [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+         */
+        get: operations["gigachat_proxy_route_gigachat__endpoint__get"];
+        /**
+         * Gigachat Proxy Route
+         * @description [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+         */
+        put: operations["gigachat_proxy_route_gigachat__endpoint__put"];
+        /**
+         * Gigachat Proxy Route
+         * @description [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+         */
+        post: operations["gigachat_proxy_route_gigachat__endpoint__post"];
+        /**
+         * Gigachat Proxy Route
+         * @description [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+         */
+        delete: operations["gigachat_proxy_route_gigachat__endpoint__delete"];
+        options?: never;
+        head?: never;
+        /**
+         * Gigachat Proxy Route
+         * @description [Docs](https://docs.litellm.ai/docs/pass_through/gigachat)
+         */
+        patch: operations["gigachat_proxy_route_gigachat__endpoint__patch"];
+        trace?: never;
+    };
     "/global/activity": {
         parameters: {
             query?: never;
@@ -34282,6 +34318,13 @@ export interface components {
             adaptive_eligible: "all" | "classified_tier";
             /** @description Quality vs cost weights for adaptive selection (used when adaptive=True) */
             adaptive_weights?: components["schemas"]["AdaptiveRouterWeights"];
+            /**
+             * Classification Mode
+             * @description When to run the complexity classifier. 'every_request' (the default) classifies every inference request, including the tool-result continuation turns of an agentic loop. 'user_turn' classifies only requests whose newest turn is a new human ask and replays the session's held routing decision on continuation turns, which cuts classifier spend and eliminates mid-loop model switches. Continuations with no held decision to replay (no resolvable session_id, expired pin, fresh restart) still classify. Unlike session_affinity, a new human ask always re-classifies, so a session can still move tiers between asks. Suppressed when plugins are configured, for the same reason session_affinity is: a replayed decision would bypass the plugin pipeline.
+             * @default every_request
+             * @enum {string}
+             */
+            classification_mode: "every_request" | "user_turn";
             /**
              * Classification Prompt
              * @description Replaces the opening instructions of the LLM classifier rubric (the judging-criteria prose) for a custom tier set. The per-tier bullets and the trust-boundary paragraph telling the classifier to ignore tier requests embedded in quoted caller text are always appended after it and cannot be overridden. Requires tier_definitions; a built-in-tier router customizes its prompt via classifier_llm_config.system_prompt or classification_rubric instead.
@@ -35554,7 +35597,7 @@ export interface components {
              * Cause
              * @enum {string}
              */
-            cause?: "heuristic_scorer" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "session_affinity_pin" | "session_affinity_escalation" | "default_fallback" | "keyword" | "quality_tier" | "bandit";
+            cause?: "heuristic_scorer" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit";
             /** Classifier Cost */
             classifier_cost?: number;
             /** Classifier Model */
@@ -46550,6 +46593,161 @@ export interface operations {
             };
         };
     };
+    gigachat_proxy_route_gigachat__endpoint__get: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                endpoint: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
+    gigachat_proxy_route_gigachat__endpoint__put: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                endpoint: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
+    gigachat_proxy_route_gigachat__endpoint__post: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                endpoint: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
+    gigachat_proxy_route_gigachat__endpoint__delete: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                endpoint: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
+    gigachat_proxy_route_gigachat__endpoint__patch: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path: {
+                endpoint: string;
+            };
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+            /** @description Validation Error */
+            422: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": components["schemas"]["HTTPValidationError"];
+                };
+            };
+        };
+    };
     get_global_activity_global_activity_get: {
         parameters: {
             query?: {