Merge pull request #36725 from BerriAI/litellm_internal_staging

ci: promote staging to main

.github/workflows/test-terraform-modules.yml

@@ -0,0 +1,54 @@
+name: Terraform Modules
+
+on:
+  push:
+    paths:
+      - "terraform/litellm/aws/**"
+      - ".github/workflows/test-terraform-modules.yml"
+  pull_request:
+    branches:
+      - main
+      - litellm_internal_staging
+      - litellm_oss_staging
+      - "litellm_**"
+    paths:
+      - "terraform/litellm/aws/**"
+      - ".github/workflows/test-terraform-modules.yml"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  aws-module:
+    name: fmt, validate, test (aws)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    defaults:
+      run:
+        working-directory: terraform/litellm/aws
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - uses: hashicorp/setup-terraform@b9cd54a3c349d3f38e8881555d616ced269862dd # v3.1.2
+        with:
+          terraform_version: 1.13.3
+          terraform_wrapper: false
+
+      - name: fmt
+        run: terraform fmt -recursive -check -diff
+
+      - name: init
+        run: terraform init -backend=false -input=false
+
+      - name: validate
+        run: terraform validate
+
+      # Plan-only, mock_provider-backed: no AWS credentials, no API calls.
+      - name: test
+        run: terraform test

.github/workflows/test-unit-proxy-db.yml

@@ -135,8 +135,6 @@ jobs:
             test-path: >-
               tests/proxy_unit_tests/test_proxy_server.py
               tests/proxy_unit_tests/test_proxy_server_keys.py
-              tests/proxy_unit_tests/test_proxy_server_caching.py
-              tests/proxy_unit_tests/test_proxy_server_langfuse.py
               tests/proxy_unit_tests/test_proxy_server_spend.py
               tests/proxy_unit_tests/test_aproxy_startup.py
             workers: 4

CLAUDE.md

@@ -29,7 +29,7 @@ End-to-end tests belong in `tests/e2e/` and must follow the harness conventions
 
 When creating PRs, don't set base to `main`. `litellm_internal_staging` is the default base branch and serves that purpose for both internal and external / OSS contributions
 
-When writing a PR body, treat the comments and imperative instructions inside @.github/pull_request_template.md as rules to follow, not just layout. Agent harnesses may strip HTML comments from copies of that file injected into context, so read .github/pull_request_template.md from disk before writing a PR body to make sure you see every comment rule
+When writing a PR body, treat the comments and imperative instructions inside .github/pull_request_template.md as rules to follow, not just layout. Agent harnesses may strip HTML comments from copies of that file injected into context, so read .github/pull_request_template.md from disk before writing a PR body to make sure you see every comment rule
 
 Same applies for filing bug reports and feature requests, with .github/ISSUE_TEMPLATE/bug_report.yml and .github/ISSUE_TEMPLATE/feature_request.yml, respectively
 

basedpyright-code-budget.json

@@ -1,6 +1,6 @@
 {
   "reportAny": {
-    "limit": 23919
+    "limit": 23914
   },
   "reportArgumentType": {
     "limit": 2580

helm/litellm-helm/templates/migrations-job.yaml

@@ -105,6 +105,10 @@ spec:
         {{- toYaml . | nindent 8 }}
       {{- end }}
       restartPolicy: OnFailure
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
       {{- with .Values.affinity }}
       affinity:
         {{- toYaml . | nindent 8 }}

helm/litellm-helm/tests/migrations-job_tests.yaml

@@ -290,3 +290,27 @@ tests:
           value:
             allowPrivilegeEscalation: false
             readOnlyRootFilesystem: true
+  - it: should schedule onto the same nodes as the gateway
+    template: migrations-job.yaml
+    set:
+      migrationJob:
+        enabled: true
+      nodeSelector:
+        karpenter.sh/nodepool: litellm-e2e
+      tolerations:
+        - key: workload
+          operator: Equal
+          value: litellm-e2e
+          effect: NoSchedule
+    asserts:
+      - equal:
+          path: spec.template.spec.nodeSelector
+          value:
+            karpenter.sh/nodepool: litellm-e2e
+      - equal:
+          path: spec.template.spec.tolerations
+          value:
+            - key: workload
+              operator: Equal
+              value: litellm-e2e
+              effect: NoSchedule

litellm/__init__.py

@@ -246,6 +246,7 @@ use_chat_completions_url_for_anthropic_messages: bool = bool(
 # config.yaml.
 strip_anthropic_total_tokens: bool = False
 anthropic_sse_ping_interval_seconds: float = 15.0
+sse_keepalive_ping_interval_seconds: float | None = None
 route_all_chat_openai_to_responses: bool = (
     os.getenv("LITELLM_ROUTE_ALL_CHAT_OPENAI_TO_RESPONSES", "false").lower() == "true"
 )  # When True, routes all OpenAI /chat/completions requests through the Responses API bridge

litellm/constants.py

@@ -472,6 +472,8 @@ EMAIL_BUDGET_ALERT_MAX_SPEND_ALERT_PERCENTAGE: Final = float(
 ### ANTHROPIC CONSTANTS ###
 ANTHROPIC_TOKEN_COUNTING_BETA_VERSION = os.getenv("ANTHROPIC_TOKEN_COUNTING_BETA_VERSION", "token-counting-2024-11-01")
 ANTHROPIC_SKILLS_API_BETA_VERSION: Final = "skills-2025-10-02"
+ANTHROPIC_BATCHES_ROUTE: Final = "/v1/messages/batches"
+VERTEX_BATCH_PREDICTION_JOBS_ROUTE: Final = "batchPredictionJobs"
 ANTHROPIC_WEB_SEARCH_TOOL_MAX_USES: Final = {
     "low": 1,
     "medium": 5,
@@ -1323,6 +1325,7 @@ LITELLM_METADATA_FIELD: Final = "litellm_metadata"
 OLD_LITELLM_METADATA_FIELD: Final = "metadata"
 RETURN_RAW_MODEL_NAME_METADATA_KEY: Final = "_complexity_router_return_raw_model_name"
 SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY: Final = "_session_deployment_affinity_ttl"
+CONSUMED_REQUEST_TAGS_METADATA_KEY: Final = "_consumed_request_tags"
 INTERNAL_CALL_ORIGIN_METADATA_KEY: Final = "internal_call_origin"
 LITELLM_TRUNCATED_PAYLOAD_FIELD: Final = "litellm_truncated"
 LITELLM_TRUNCATION_DB_SAFEGUARD_NOTE: Final = (
@@ -1527,6 +1530,10 @@ APSCHEDULER_REPLACE_EXISTING: Final = os.getenv("APSCHEDULER_REPLACE_EXISTING",
     "1",
 ]  # always replace existing jobs
 
+# Width of the window scheduled background jobs are spread across, so they do not all fire
+# on one instant on every replica. Tunable per deployment via general_settings.
+DEFAULT_STAGGER_WINDOW_SECONDS: Final = 300
+
 # The number of tag entries are higher than number of user, team entries. This leads to a higher QPS.
 # This will run tag spcific tasks at a later time to smooth QPS
 DAILY_TAG_SPEND_BATCH_MULTIPLIER: Final = 2.3

litellm/experimental_mcp_client/client.py

@@ -6,10 +6,11 @@ import asyncio
 import base64
 import os
 from collections.abc import Awaitable, Callable, Generator
+from datetime import timedelta
 from typing import Any, Final, TypeVar
 
 import httpx
-from mcp import ClientSession, ReadResourceResult, Resource, StdioServerParameters
+from mcp import ClientSession, McpError, ReadResourceResult, Resource, StdioServerParameters
 from mcp.client.sse import sse_client
 from mcp.client.stdio import stdio_client
 
@@ -69,6 +70,29 @@ def _first_non_cancelled_cause(exc: BaseException) -> BaseException | None:
     return None
 
 
+_SDK_READ_TIMEOUT_CODE: Final = int(httpx.codes.REQUEST_TIMEOUT)
+"""The code the MCP SDK puts on its own elapsed read timeout, an HTTP status in a field that
+otherwise carries JSON-RPC error codes."""
+
+
+def _as_read_timeout(exc: BaseException) -> TimeoutError | None:
+    """The session read timeout elapsing, re-expressed as a ``TimeoutError``, or ``None``.
+
+    The SDK reports its own elapsed read timeout as ``McpError`` carrying an HTTP status code in a
+    field that otherwise holds JSON-RPC error codes, and it relays an upstream's JSON-RPC error
+    through that same class and field. The numeric code alone therefore cannot separate the two, and
+    an upstream answering with application code 408 would be reported as a gateway timeout it never
+    caused. The SDK raises its own from inside an ``except TimeoutError``, so the elapsed timeout is
+    on the context chain, while a relayed error is built from a received message and has no such
+    chain; that is the discriminator.
+    """
+    if not isinstance(exc, McpError) or exc.error.code != _SDK_READ_TIMEOUT_CODE:
+        return None
+    if not isinstance(exc.__context__, TimeoutError):
+        return None
+    return TimeoutError(exc.error.message)
+
+
 TSessionResult = TypeVar("TSessionResult")
 
 
@@ -347,7 +371,14 @@ class MCPClient:
                 session_kwargs["elicitation_callback"] = self._elicitation_callback
             if self._logging_callback is not None:
                 session_kwargs["logging_callback"] = self._logging_callback
-            session_ctx: Final = ClientSession(read_stream, write_stream, **session_kwargs)
+            # The SDK drops a response stream that ends without a JSON-RPC reply, so nothing else
+            # ever fails the request.
+            session_ctx: Final = ClientSession(
+                read_stream,
+                write_stream,
+                read_timeout_seconds=timedelta(seconds=self.timeout),
+                **session_kwargs,
+            )
             session: Final = await session_ctx.__aenter__()
             try:
                 init_result: Final = await session.initialize()
@@ -390,7 +421,16 @@ class MCPClient:
             self._last_initialize_instructions = None
             transport_ctx, http_client = self._create_transport_context()
             return await self._execute_session_operation(transport_ctx, operation)
-        except Exception:
+        except Exception as e:
+            read_timeout: Final = _as_read_timeout(e)
+            if read_timeout is not None:
+                verbose_logger.warning(
+                    "MCP client timed out after %ss waiting for %s to answer; the server accepted the "
+                    "request and ended its response stream without a JSON-RPC reply",
+                    self.timeout,
+                    self.server_url or "stdio",
+                )
+                raise read_timeout from e
             _log: Final = verbose_logger.debug if quiet_on_error else verbose_logger.warning
             _log("MCP client run_with_session failed for %s", self.server_url or "stdio")
             raise

litellm/integrations/langfuse/langfuse.py

@@ -2,7 +2,7 @@
 #    On success, logs events to Langfuse
 import os
 import traceback
-from collections.abc import Callable
+from collections.abc import Callable, Iterable
 from datetime import datetime
 from typing import TYPE_CHECKING, Any, Final, cast
 
@@ -75,6 +75,22 @@ def _extract_cache_read_input_tokens(usage_obj) -> int:
     return cache_read_input_tokens
 
 
+def _as_steering_flag(value: object) -> bool:
+    """A string ``str_to_bool`` does not recognise falls back to its truthiness."""
+    if isinstance(value, str):
+        parsed: Final = str_to_bool(value)
+        return bool(value) if parsed is None else parsed
+    return bool(value)
+
+
+def _as_steering_key_sequence(value: object) -> tuple[str, ...]:
+    if isinstance(value, str):
+        return tuple(key.strip() for key in value.split(",") if key.strip())
+    if isinstance(value, Iterable):
+        return tuple(str(key) for key in value)
+    return ()
+
+
 def resolve_langfuse_credentials(
     langfuse_public_key=None,
     langfuse_secret=None,
@@ -552,10 +568,10 @@ class LangFuseLogger:
             # This allows continuing an existing trace while still returning the correct trace_id
             if existing_trace_id is not None:
                 trace_id = existing_trace_id
-            update_trace_keys: Final = cast(list, clean_metadata.pop("update_trace_keys", []))
+            update_trace_keys: Final = _as_steering_key_sequence(clean_metadata.pop("update_trace_keys", ()))
             debug: Final = clean_metadata.pop("debug_langfuse", None)
-            mask_input: Final = clean_metadata.pop("mask_input", False)
-            mask_output: Final = clean_metadata.pop("mask_output", False)
+            mask_input: Final = _as_steering_flag(clean_metadata.pop("mask_input", False))
+            mask_output: Final = _as_steering_flag(clean_metadata.pop("mask_output", False))
             # Look for masking function in the dedicated location first (set by scrub_sensitive_keys_in_metadata)
             # Fall back to metadata for backwards compatibility
             masking_function: Final = litellm_params.get("_langfuse_masking_function") or clean_metadata.pop(

litellm/integrations/langfuse/langfuse_otel.py

@@ -90,7 +90,6 @@ class LangfuseOtelLogger(OpenTelemetry):
             "generation_name": LangfuseSpanAttributes.GENERATION_NAME,
             "generation_id": LangfuseSpanAttributes.GENERATION_ID,
             "parent_observation_id": LangfuseSpanAttributes.PARENT_OBSERVATION_ID,
-            "version": LangfuseSpanAttributes.GENERATION_VERSION,
             "mask_input": LangfuseSpanAttributes.MASK_INPUT,
             "mask_output": LangfuseSpanAttributes.MASK_OUTPUT,
             "trace_user_id": LangfuseSpanAttributes.TRACE_USER_ID,
@@ -99,13 +98,18 @@ class LangfuseOtelLogger(OpenTelemetry):
             "trace_name": LangfuseSpanAttributes.TRACE_NAME,
             "trace_id": LangfuseSpanAttributes.TRACE_ID,
             "trace_metadata": LangfuseSpanAttributes.TRACE_METADATA,
-            "trace_version": LangfuseSpanAttributes.TRACE_VERSION,
-            "trace_release": LangfuseSpanAttributes.TRACE_RELEASE,
+            "trace_release": LangfuseSpanAttributes.RELEASE,
             "existing_trace_id": LangfuseSpanAttributes.EXISTING_TRACE_ID,
             "update_trace_keys": LangfuseSpanAttributes.UPDATE_TRACE_KEYS,
             "debug_langfuse": LangfuseSpanAttributes.DEBUG_LANGFUSE,
         }
 
+        version: Final = (
+            metadata.get("trace_version") if metadata.get("trace_version") is not None else metadata.get("version")
+        )
+        if version is not None:
+            safe_set_attribute(span, LangfuseSpanAttributes.VERSION.value, version)
+
         for key, enum_attr in mapping.items():
             if key in metadata and metadata[key] is not None:
                 value = metadata[key]

litellm/integrations/otel/presets/langfuse.py

@@ -42,9 +42,7 @@ def langfuse_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str,
     public_key: Final = params.get("langfuse_public_key")
     secret_key: Final = params.get("langfuse_secret_key")
     if public_key and secret_key:
-        return {
-            "Authorization": _V1Langfuse._get_langfuse_authorization_header(
-                public_key=public_key, secret_key=secret_key
-            )
-        }
+        return _V1Langfuse._build_langfuse_otel_headers(
+            _V1Langfuse._get_langfuse_authorization_header(public_key=public_key, secret_key=secret_key)
+        )
     return {}

litellm/litellm_core_utils/core_helpers.py

@@ -1,7 +1,7 @@
 # What is this?
 ## Helper utilities
 import copy
-from collections.abc import Iterable
+from collections.abc import Iterable, Mapping
 from typing import TYPE_CHECKING, Any, Final, Literal
 
 import httpx
@@ -181,7 +181,7 @@ def add_missing_spend_metadata_to_litellm_metadata(litellm_metadata: dict, metad
 
 
 def get_metadata_variable_name_from_kwargs(
-    kwargs: dict,
+    kwargs: Mapping[str, object],
 ) -> Literal["metadata", "litellm_metadata"]:
     """
     Helper to return what the "metadata" field should be called in the request data

litellm/litellm_core_utils/litellm_logging.py

@@ -3462,7 +3462,7 @@ class Logging(LiteLLMLoggingBaseClass):
                 model=self.model,
                 messages=[],
                 logging_obj=self,
-                optional_params={},
+                optional_params=self.optional_params or {},
                 api_key="",
                 request_data={},
                 encoding=litellm.encoding,
@@ -3483,6 +3483,7 @@ class Logging(LiteLLMLoggingBaseClass):
                 ),
                 model_response=litellm.ModelResponse(),
                 json_mode=None,
+                speed=self.optional_params.get("speed") if self.optional_params else None,
             )
         return result
 

litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py

@@ -2,6 +2,7 @@
 Helper utilities for tracking the cost of built-in tools.
 """
 
+from collections.abc import Mapping
 from typing import Any, Final, Literal
 
 import litellm
@@ -23,6 +24,14 @@ from litellm.types.utils import (
 )
 
 
+def _usage_reports_server_side_web_search_calls(usage: Usage) -> bool:
+    details: Final = getattr(usage, "server_side_tool_usage_details", None)
+    if not isinstance(details, Mapping):
+        return False
+    calls: Final = details.get("web_search_calls")
+    return isinstance(calls, int) and calls > 0
+
+
 class StandardBuiltInToolCostTracking:
     """
     Helper class for tracking the cost of built-in tools
@@ -351,6 +360,10 @@ class StandardBuiltInToolCostTracking:
                 # and _handle_web_search_cost() is never called.
                 if hasattr(usage, "server_tool_use") and _get_web_search_requests(usage.server_tool_use) is not None:
                     return True
+                # xAI reports usage.server_side_tool_usage_details.web_search_calls; a searched
+                # answer with no url_citation annotations has no other chat-path signal
+                if _usage_reports_server_side_web_search_calls(usage):
+                    return True
             return False
         elif isinstance(response_object, ResponsesAPIResponse):
             # response api explicitly includes web_search_call in the output
@@ -370,6 +383,8 @@ class StandardBuiltInToolCostTracking:
                 )
             ):
                 return True
+            if _usage_reports_server_side_web_search_calls(usage):
+                return True
 
         return False
 
@@ -432,7 +447,9 @@ class StandardBuiltInToolCostTracking:
         """
         output: Final = response_object.output
         for output_item in output:
-            _output_type: str | None = getattr(output_item, "type", None)
+            _output_type: str | None = (
+                output_item.get("type") if isinstance(output_item, dict) else getattr(output_item, "type", None)
+            )
             if _output_type == output_type:
                 return True
         return False

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -803,8 +803,28 @@ class ChunkProcessor:
             completion_tokens_details=completion_tokens_details,
             prompt_tokens_details=prompt_tokens_details,
             cost=cost,
+            inference_geo=self._last_provider_pricing_field(chunks, "inference_geo"),
+            speed=self._last_provider_pricing_field(chunks, "speed"),
         )
 
+    def _last_provider_pricing_field(
+        self,
+        chunks: Sequence["_UsageBearingChunk | ModelResponse"],
+        field: str,
+    ) -> str | None:
+        """
+        Last value of a provider-specific usage field that changes pricing but is not a
+        declared ``Usage`` field, e.g. Anthropic's ``speed`` (fast mode multiplies
+        non-cache token cost) and ``inference_geo``.
+        """
+        values: Final = [
+            value
+            for chunk in chunks
+            if (usage_chunk := self._extract_usage_chunk(chunk)) is not None
+            and isinstance(value := getattr(usage_chunk, field, None), str)
+        ]
+        return values[-1] if values else None
+
     @staticmethod
     def _reset_anthropic_cursor_completion_tokens(
         chunks: Sequence["_UsageBearingChunk | ModelResponse"],
@@ -934,7 +954,16 @@ class ChunkProcessor:
 
         # Return a new usage object with the new values
 
-        returned_usage = Usage(**returned_usage.model_dump())
+        provider_pricing_fields: Final = {
+            field: value
+            for field, value in (
+                ("inference_geo", calculated_usage_per_chunk["inference_geo"]),
+                ("speed", calculated_usage_per_chunk["speed"]),
+            )
+            if value is not None
+        }
+
+        returned_usage = Usage(**returned_usage.model_dump(), **provider_pricing_fields)
 
         return returned_usage
 

litellm/llms/nvidia_nim/rerank/ranking_transformation.py

@@ -8,13 +8,23 @@ Reference: https://build.nvidia.com/nvidia/llama-3_2-nv-rerankqa-1b-v2/deploy
 
 from typing import Final
 
+import httpx
+
+from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.llms.nvidia_nim.rerank.transformation import NvidiaNimRerankConfig
+from litellm.types.rerank import RerankResponse
 
 
 class NvidiaNimRankingConfig(NvidiaNimRerankConfig):
     """
     Configuration for NVIDIA NIM models that use the /v1/ranking endpoint.
-    
+
+    The native /v1/ranking request schema accepts only 'model', 'query',
+    'passages', and 'truncate' -- requests containing 'top_k' are rejected
+    with a 400 validation error. Cohere-compatible 'top_n' is therefore
+    applied client-side by truncating the converted response instead of
+    being forwarded to the endpoint.
+
     Example:
         curl -X "POST" 'https://ai.api.nvidia.com/v1/ranking' \
             -H 'Accept: application/json' \
@@ -27,6 +37,16 @@ class NvidiaNimRankingConfig(NvidiaNimRerankConfig):
             }'
     """
 
+    SUPPORTED_PASSAGE_FIELDS: tuple[str, ...] = ("text", "image")
+
+    def __init__(self) -> None:
+        super().__init__()
+        # top_n captured in transform_rerank_request and applied in
+        # transform_rerank_response. The provider config is instantiated
+        # per-request (see ProviderConfigManager.get_provider_rerank_config),
+        # so this does not leak across requests.
+        self._client_side_top_n: int | None = None
+
     def _get_clean_model_name(self, model: str) -> str:
         """Strip 'nvidia_nim/' and 'ranking/' prefixes from model name."""
         # First strip nvidia_nim/ prefix if present
@@ -58,6 +78,47 @@ class NvidiaNimRankingConfig(NvidiaNimRerankConfig):
 
         return f"{api_base}/v1/ranking"
 
+    def map_cohere_rerank_params(
+        self,
+        non_default_params: dict | None,  # mutable-ok: matches BaseRerankConfig's request contract
+        model: str,
+        drop_params: bool,
+        query: str,
+        documents: list[str | dict[str, object]],  # mutable-ok: matches BaseRerankConfig's document contract
+        custom_llm_provider: str | None = None,
+        top_n: int | None = None,
+        rank_fields: list[str] | None = None,  # mutable-ok: matches BaseRerankConfig's field contract
+        return_documents: bool | None = True,
+        max_chunks_per_doc: int | None = None,
+        max_tokens_per_doc: int | None = None,
+        instruction: str | None = None,
+    ) -> dict:  # mutable-ok: LiteLLM provider transforms return mutable request dictionaries
+        """
+        Keep Cohere's top_n as-is instead of mapping it to top_k.
+
+        The native /v1/ranking endpoint rejects top_k, so top_n is applied
+        client-side after the response is converted.
+        """
+        optional_params: Final = super().map_cohere_rerank_params(
+            non_default_params=non_default_params,
+            model=model,
+            drop_params=drop_params,
+            query=query,
+            documents=documents,
+            custom_llm_provider=custom_llm_provider,
+            top_n=None,  # do not map top_n -> top_k for /v1/ranking
+            rank_fields=rank_fields,
+            return_documents=return_documents,
+            max_chunks_per_doc=max_chunks_per_doc,
+            max_tokens_per_doc=max_tokens_per_doc,
+            instruction=instruction,
+        )
+        # /v1/ranking rejects top_k even when passed as a provider-specific param
+        optional_params.pop("top_k", None)
+        if top_n is not None:
+            optional_params["top_n"] = top_n
+        return optional_params
+
     def transform_rerank_request(
         self,
         model: str,
@@ -67,11 +128,66 @@ class NvidiaNimRankingConfig(NvidiaNimRerankConfig):
     ) -> dict:
         """
         Transform request, using clean model name without 'ranking/' prefix.
+
+        top_n / top_k are stripped from the outgoing request: the native
+        /v1/ranking endpoint accepts only model, query, passages, and
+        truncate. top_n is stashed and applied client-side in
+        transform_rerank_response.
         """
+        top_n: Final = optional_rerank_params.get("top_n")
+        if top_n is not None:
+            if isinstance(top_n, bool) or not isinstance(top_n, int) or top_n < 1:
+                raise ValueError(f"top_n must be a positive integer, got: {top_n!r}")
+            self._client_side_top_n = top_n
+
         clean_model: Final = self._get_clean_model_name(model)
+        filtered_params: Final = {  # mutable-ok: the base transformer requires a mutable request dictionary
+            k: v for k, v in optional_rerank_params.items() if k not in ("top_n", "top_k")
+        }
         return super().transform_rerank_request(
             model=clean_model,
-            optional_rerank_params=optional_rerank_params,
+            optional_rerank_params=filtered_params,
             headers=headers,
             litellm_params=litellm_params,
         )
+
+    def transform_rerank_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        model_response: RerankResponse,
+        logging_obj: LiteLLMLoggingObj,
+        api_key: str | None = None,
+        request_data: dict | None = None,  # mutable-ok: matches BaseRerankConfig's response contract
+        optional_params: dict | None = None,  # mutable-ok: matches BaseRerankConfig's response contract
+        litellm_params: dict | None = None,  # mutable-ok: matches BaseRerankConfig's response contract
+    ) -> RerankResponse:
+        """
+        Convert the native ranking response, then apply top_n client-side.
+
+        /v1/ranking returns rankings sorted by relevance, but sort before
+        truncating in case a server returns them unsorted.
+        """
+        resolved_request_data: Final = request_data or {}  # mutable-ok: the base transformer requires a dictionary
+        resolved_optional_params: Final = optional_params or {}  # mutable-ok: response options are keyed lookups
+        resolved_litellm_params: Final = litellm_params or {}  # mutable-ok: the base transformer requires a dictionary
+
+        response: Final = super().transform_rerank_response(
+            model=model,
+            raw_response=raw_response,
+            model_response=model_response,
+            logging_obj=logging_obj,
+            api_key=api_key,
+            request_data=resolved_request_data,
+            optional_params=resolved_optional_params,
+            litellm_params=resolved_litellm_params,
+        )
+
+        top_n: Final = resolved_optional_params.get("top_n") or self._client_side_top_n
+        if top_n is not None and response.results is not None and len(response.results) > top_n:
+            response.results = sorted(
+                response.results,
+                key=lambda result: result["relevance_score"],
+                reverse=True,
+            )[:top_n]
+        return response

litellm/llms/nvidia_nim/rerank/transformation.py

@@ -21,8 +21,9 @@ class NvidiaNimQueryObject(TypedDict):
     text: Required[str]
 
 
-class NvidiaNimPassageObject(TypedDict):
-    text: Required[str]
+class NvidiaNimPassageObject(TypedDict, total=False):
+    text: str
+    image: str
 
 
 class NvidiaNimRerankRequest(TypedDict, total=False):
@@ -53,6 +54,10 @@ class NvidiaNimRerankConfig(BaseRerankConfig):
 
     DEFAULT_NIM_RERANK_API_BASE = "https://ai.api.nvidia.com"
 
+    # The legacy retrieval rerank route accepts text passages only. The native
+    # ranking subclass expands this tuple for VL models that accept images.
+    SUPPORTED_PASSAGE_FIELDS: tuple[str, ...] = ("text",)
+
     def __init__(self) -> None:
         pass
 
@@ -206,11 +211,17 @@ class NvidiaNimRerankConfig(BaseRerankConfig):
             if isinstance(doc, str):
                 passages.append({"text": doc})
             elif isinstance(doc, dict):
-                # If document is already a dict, check if it has 'text' field
-                if "text" in doc:
-                    passages.append({"text": doc["text"]})
+                # Preserve only the structured passage fields supported by the
+                # selected rerank route.
+                supported_fields: NvidiaNimPassageObject = {}  # mutable-ok: assembling a request TypedDict
+                if "text" in self.SUPPORTED_PASSAGE_FIELDS and "text" in doc:
+                    supported_fields["text"] = doc["text"]
+                if "image" in self.SUPPORTED_PASSAGE_FIELDS and "image" in doc:
+                    supported_fields["image"] = doc["image"]
+                if supported_fields:
+                    passages.append(supported_fields)
                 else:
-                    # Otherwise, stringify the dict
+                    # No supported fields - stringify the dict
                     import json
 
                     passages.append({"text": json.dumps(doc)})
@@ -304,9 +315,10 @@ class NvidiaNimRerankConfig(BaseRerankConfig):
                 "relevance_score": ranking["logit"],
             }
 
-            # Include document if it was in the original request
+            # Include document if it was in the original request.
+            # Image-only passages carry no 'text' field, so guard the lookup.
             index: int = ranking["index"]
-            if index < len(original_passages):
+            if index < len(original_passages) and "text" in original_passages[index]:
                 result_item["document"] = {"text": original_passages[index]["text"]}
 
             results.append(result_item)

litellm/llms/openai/responses/transformation.py

@@ -353,6 +353,16 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
             )
             return event_pydantic_model.model_construct(**parsed_chunk)
 
+    @staticmethod
+    def parse_terminal_response_from_stream_chunks(all_chunks: list[str]) -> ResponsesAPIResponse | None:
+        for chunk_str in reversed(all_chunks):
+            for event_model in (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent):
+                try:
+                    return event_model.model_validate_json(chunk_str.removeprefix("data: ")).response
+                except ValueError:
+                    continue
+        return None
+
     @staticmethod
     def get_event_model_class(event_type: str) -> Any:
         """

litellm/llms/vertex_ai/vertex_gemma_models/transformation.py

@@ -14,6 +14,11 @@ from typing import Any, Final, cast
 import httpx
 
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.custom_httpx.http_handler import (
+    AsyncHTTPHandler,
+    HTTPHandler,
+    _get_httpx_client,
+)
 from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import ModelResponse
@@ -123,6 +128,79 @@ class VertexGemmaConfig(OpenAIGPTConfig):
 
         return response_json["predictions"]
 
+    @staticmethod
+    def _sync_post(
+        client: HTTPHandler | httpx.Client | None,
+        api_base: str,
+        headers: dict[str, str],  # mutable-ok: forwarded to post(headers: dict | None)
+        request_data: dict[str, Any],  # mutable-ok: forwarded to post(json: dict | ...)
+        timeout: float | httpx.Timeout | None,
+    ) -> httpx.Response:
+        if isinstance(client, HTTPHandler):
+            return client.post(
+                url=api_base,
+                headers=headers,
+                json=request_data,
+                timeout=timeout,
+            )
+        if isinstance(client, httpx.Client):
+            if timeout is None:
+                return client.post(
+                    url=api_base,
+                    headers=headers,
+                    json=request_data,
+                )
+            return client.post(
+                url=api_base,
+                headers=headers,
+                json=request_data,
+                timeout=timeout,
+            )
+        return _get_httpx_client().post(
+            url=api_base,
+            headers=headers,
+            json=request_data,
+            timeout=timeout,
+        )
+
+    @staticmethod
+    async def _async_post(
+        client: AsyncHTTPHandler | httpx.AsyncClient | None,
+        api_base: str,
+        headers: dict[str, str],  # mutable-ok: forwarded to post(headers: dict | None)
+        request_data: dict[str, Any],  # mutable-ok: forwarded to post(json: dict | ...)
+        timeout: float | httpx.Timeout | None,
+    ) -> httpx.Response:
+        from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
+        from litellm.types.utils import LlmProviders
+
+        if isinstance(client, AsyncHTTPHandler):
+            return await client.post(
+                url=api_base,
+                headers=headers,
+                json=request_data,
+                timeout=timeout,
+            )
+        if isinstance(client, httpx.AsyncClient):
+            if timeout is None:
+                return await client.post(
+                    url=api_base,
+                    headers=headers,
+                    json=request_data,
+                )
+            return await client.post(
+                url=api_base,
+                headers=headers,
+                json=request_data,
+                timeout=timeout,
+            )
+        return await get_async_httpx_client(llm_provider=LlmProviders.VERTEX_AI).post(
+            url=api_base,
+            headers=headers,
+            json=request_data,
+            timeout=timeout,
+        )
+
     def completion(
         self,
         model: str,
@@ -137,7 +215,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         acompletion: bool,
         litellm_params: dict,
         logger_fn: Callable | None = None,
-        client: httpx.Client | None = None,
+        client: HTTPHandler | AsyncHTTPHandler | httpx.Client | httpx.AsyncClient | None = None,
         timeout: float | httpx.Timeout | None = None,
         encoding=None,
         custom_llm_provider: str = "vertex_ai",
@@ -147,6 +225,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         Supports both sync and async requests with fake streaming.
         """
         if acompletion:
+            async_client = client if isinstance(client, (AsyncHTTPHandler, httpx.AsyncClient)) else None
             return self._async_completion(
                 model=model,
                 messages=messages,
@@ -157,10 +236,12 @@ class VertexGemmaConfig(OpenAIGPTConfig):
                 logging_obj=logging_obj,
                 optional_params=optional_params,
                 litellm_params=litellm_params,
+                client=async_client,
                 timeout=timeout,
                 encoding=encoding,
             )
         else:
+            sync_client = client if isinstance(client, (HTTPHandler, httpx.Client)) else None
             return self._sync_completion(
                 model=model,
                 messages=messages,
@@ -171,6 +252,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
                 logging_obj=logging_obj,
                 optional_params=optional_params,
                 litellm_params=litellm_params,
+                client=sync_client,
                 timeout=timeout,
                 encoding=encoding,
             )
@@ -186,11 +268,11 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         logging_obj: Any,
         optional_params: dict,
         litellm_params: dict,
-        timeout: float | httpx.Timeout | None,
-        encoding: Any,
+        client: HTTPHandler | httpx.Client | None = None,
+        timeout: float | httpx.Timeout | None = None,
+        encoding: Any = None,
     ):
         """Synchronous completion request"""
-        from litellm.llms.custom_httpx.http_handler import HTTPHandler
         from litellm.utils import convert_to_model_response_object
 
         # Check if streaming is requested (will be faked)
@@ -222,11 +304,11 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         )
 
         # Make the HTTP request
-        http_handler: Final = HTTPHandler(concurrent_limit=1)
-        response: Final = http_handler.post(
-            url=api_base,
+        response: Final = self._sync_post(
+            client=client,
+            api_base=api_base,
             headers=headers,
-            json=request_data,
+            request_data=request_data,
             timeout=timeout,
         )
 
@@ -276,12 +358,11 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         logging_obj: Any,
         optional_params: dict,
         litellm_params: dict,
-        timeout: float | httpx.Timeout | None,
-        encoding: Any,
+        client: AsyncHTTPHandler | httpx.AsyncClient | None = None,
+        timeout: float | httpx.Timeout | None = None,
+        encoding: Any = None,
     ):
         """Asynchronous completion request"""
-        from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
-        from litellm.types.utils import LlmProviders
         from litellm.utils import convert_to_model_response_object
 
         # Check if streaming is requested (will be faked)
@@ -313,13 +394,11 @@ class VertexGemmaConfig(OpenAIGPTConfig):
         )
 
         # Make the HTTP request
-        http_handler: Final = get_async_httpx_client(
-            llm_provider=LlmProviders.VERTEX_AI,
-        )
-        response: Final = await http_handler.post(
-            url=api_base,
+        response: Final = await self._async_post(
+            client=client,
+            api_base=api_base,
             headers=headers,
-            json=request_data,
+            request_data=request_data,
             timeout=timeout,
         )
 

litellm/llms/xai/chat/transformation.py

@@ -1,4 +1,4 @@
-from collections.abc import AsyncIterator, Iterator
+from collections.abc import AsyncIterator, Iterator, Mapping
 from typing import Any, Final
 
 import httpx
@@ -12,13 +12,15 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     strip_name_from_messages,
 )
 from litellm.llms.xai.common_utils import XAIModelInfo
+from litellm.llms.xai.cost_calculator import (
+    apply_server_side_tool_usage_details_to_usage,
+)
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import (
     Choices,
     ModelResponse,
     ModelResponseStream,
-    PromptTokensDetailsWrapper,
     Usage,
 )
 
@@ -248,7 +250,7 @@ class XAIChatConfig(OpenAIGPTConfig):
         XAI API returns empty string for finish_reason when using tools,
         so we need to fix this after the standard OpenAI transformation.
 
-        Also handles X.AI web search usage tracking by extracting num_sources_used.
+        Also handles X.AI web search usage tracking.
         """
 
         # First, let the parent class handle the standard transformation
@@ -351,25 +353,20 @@ class XAIChatConfig(OpenAIGPTConfig):
 
     def _enhance_usage_with_xai_web_search_fields(self, model_response: ModelResponse, raw_response_json: dict) -> None:
         """
-        Extract num_sources_used from X.AI response and map it to web_search_requests.
+        Copy usage.server_side_tool_usage_details from the provider usage block
+        onto model_response.usage for tool cost calculation.
         """
         if not hasattr(model_response, "usage") or model_response.usage is None:
             return
 
         usage: Final[Usage] = model_response.usage
-        num_sources_used = None
-        response_usage: Final = raw_response_json.get("usage", {})
-        if isinstance(response_usage, dict) and "num_sources_used" in response_usage:
-            num_sources_used = response_usage.get("num_sources_used")
-
-        # Map num_sources_used to web_search_requests for cost detection
-        if num_sources_used is not None and num_sources_used > 0:
-            if usage.prompt_tokens_details is None:
-                usage.prompt_tokens_details = PromptTokensDetailsWrapper()
-
-            usage.prompt_tokens_details.web_search_requests = int(num_sources_used)
-            setattr(usage, "num_sources_used", int(num_sources_used))
-            verbose_logger.debug("X.AI web search sources used: %s", num_sources_used)
+        response_usage: Final = raw_response_json.get("usage")
+        if not isinstance(response_usage, dict):
+            return
+        details: Final = response_usage.get("server_side_tool_usage_details")
+        if isinstance(details, Mapping):
+            apply_server_side_tool_usage_details_to_usage(usage, details)
+            verbose_logger.debug("X.AI server_side_tool_usage_details: %s", details)
 
     @staticmethod
     def _normalize_openai_compatible_usage_totals(

litellm/llms/xai/cost_calculator.py

@@ -4,14 +4,37 @@ Helper util for handling XAI-specific cost calculation
 - Handles XAI-specific reasoning token billing (billed as part of completion tokens)
 """
 
+from collections.abc import Mapping
 from typing import TYPE_CHECKING, Final
 
 from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token
-from litellm.types.utils import Usage
+from litellm.types.utils import PromptTokensDetailsWrapper, Usage
 
 if TYPE_CHECKING:
     from litellm.types.utils import ModelInfo
 
+# https://docs.x.ai/developers/pricing#tools-pricing — default when unset in model map
+_DEFAULT_WEB_SEARCH_COST_PER_CALL: Final = 5.0 / 1000.0
+
+
+def apply_server_side_tool_usage_details_to_usage(usage: Usage, details: Mapping[str, object] | None) -> None:
+    """
+    Attach server_side_tool_usage_details and mirror web_search_calls onto
+    prompt_tokens_details.web_search_requests for built-in tool cost gating.
+    """
+    if details is None:
+        return
+    usage.server_side_tool_usage_details = details  # pyright: ignore[reportAttributeAccessIssue]  # extra  # rebind-ok: extras
+    try:
+        web_search_calls: Final = int(details.get("web_search_calls") or 0)
+    except (TypeError, ValueError):
+        return
+    if web_search_calls <= 0:
+        return
+    prompt_tokens_details: Final = usage.prompt_tokens_details or PromptTokensDetailsWrapper()
+    prompt_tokens_details.web_search_requests = web_search_calls
+    usage.prompt_tokens_details = prompt_tokens_details  # rebind-ok: write details onto caller usage
+
 
 def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
     """
@@ -32,9 +55,11 @@ def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
     prompt_tokens: Final = int(getattr(usage, "prompt_tokens", 0) or 0)
     completion_tokens: Final = int(getattr(usage, "completion_tokens", 0) or 0)
     total_tokens: Final = int(getattr(usage, "total_tokens", 0) or 0)
-    reasoning_tokens = 0
-    if hasattr(usage, "completion_tokens_details") and usage.completion_tokens_details:
-        reasoning_tokens = int(getattr(usage.completion_tokens_details, "reasoning_tokens", 0) or 0)
+    reasoning_tokens: Final = (
+        int(getattr(usage.completion_tokens_details, "reasoning_tokens", 0) or 0)
+        if hasattr(usage, "completion_tokens_details") and usage.completion_tokens_details
+        else 0
+    )
 
     already_normalised: Final = total_tokens == prompt_tokens + completion_tokens
     total_completion_tokens: Final = completion_tokens if already_normalised else completion_tokens + reasoning_tokens
@@ -52,33 +77,48 @@ def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
     return prompt_cost, completion_cost
 
 
+def _web_search_cost_per_call_from_model_info(model_info: "ModelInfo") -> float:
+    """
+    Per-invocation web_search price from model_info when configured.
+
+    Prefer ``search_context_cost_per_query`` (same shape as Gemini/Anthropic web
+    search pricing in the model cost map). Fall back to current xAI list pricing.
+    """
+    search_costs: Final = model_info.get("search_context_cost_per_query")
+    if not isinstance(search_costs, Mapping):
+        return _DEFAULT_WEB_SEARCH_COST_PER_CALL
+    for key in (
+        "search_context_size_medium",
+        "search_context_size_low",
+        "search_context_size_high",
+    ):
+        value = search_costs.get(key)
+        if value is None:
+            continue
+        try:
+            cost = float(value)
+        except (TypeError, ValueError):
+            continue
+        if cost > 0:
+            return cost
+    return _DEFAULT_WEB_SEARCH_COST_PER_CALL
+
+
 def cost_per_web_search_request(usage: "Usage", model_info: "ModelInfo") -> float:
     """
     Calculate the cost of web search requests for X.AI models.
 
-    X.AI Live Search costs $25 per 1,000 sources used.
-    Each source costs $0.025.
-
-    The number of sources is stored in prompt_tokens_details.web_search_requests
-    by the transformation layer to be compatible with the existing detection system.
+    Counts invocations from usage.server_side_tool_usage_details.web_search_calls.
+    Per-call rate comes from model_info.search_context_cost_per_query when set,
+    otherwise the default xAI tools rate ($5 / 1k calls).
     """
-    # Cost per source used: $25 per 1,000 sources = $0.025 per source
-    cost_per_source: Final = 25.0 / 1000.0  # $0.025
-
-    num_sources_used = 0
-
-    if (
-        hasattr(usage, "prompt_tokens_details")
-        and usage.prompt_tokens_details is not None
-        and hasattr(usage.prompt_tokens_details, "web_search_requests")
-        and usage.prompt_tokens_details.web_search_requests is not None
-    ):
-        num_sources_used = int(usage.prompt_tokens_details.web_search_requests)
-
-    # Fallback: try to get from num_sources_used if set directly
-    elif hasattr(usage, "num_sources_used") and usage.num_sources_used is not None:
-        num_sources_used = int(usage.num_sources_used)
-
-    total_cost: Final = cost_per_source * num_sources_used
-
-    return total_cost
+    details: Final = getattr(usage, "server_side_tool_usage_details", None)
+    if not isinstance(details, Mapping):
+        return 0.0
+    try:
+        web_search_calls: Final = int(details.get("web_search_calls") or 0)
+    except (TypeError, ValueError):
+        return 0.0
+    if web_search_calls <= 0:
+        return 0.0
+    return _web_search_cost_per_call_from_model_info(model_info) * web_search_calls

litellm/llms/xai/responses/transformation.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, Any, Final
+from typing import Any, Final
 
 import litellm
 from litellm._logging import verbose_logger
@@ -12,13 +12,6 @@ from litellm.types.llms.xai import XAIWebSearchTool, XAIXSearchTool
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import LlmProviders
 
-if TYPE_CHECKING:
-    from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
-
-    LiteLLMLoggingObj = _LiteLLMLoggingObj
-else:
-    LiteLLMLoggingObj = Any
-
 
 class XAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
     """

litellm/model_prices_and_context_window_backup.json

@@ -9707,6 +9707,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -9830,6 +9831,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -9922,6 +9924,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10007,6 +10010,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10410,6 +10414,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10624,6 +10629,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10701,6 +10707,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -11262,6 +11269,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -12940,6 +12948,103 @@
         "supports_system_messages": true,
         "supports_tool_choice": false
     },
+    "dashscope/deepseek-v4-flash": {
+        "cache_read_input_token_cost": 4e-08,
+        "input_cost_per_token": 2e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4e-07,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/deepseek-v4-flash-0731": {
+        "cache_read_input_token_cost": 4e-08,
+        "input_cost_per_token": 2e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4e-07,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/deepseek-v4-pro": {
+        "cache_read_input_token_cost": 2e-07,
+        "input_cost_per_token": 2.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4.8e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/glm-5.1": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 202745,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/glm-5.2": {
+        "cache_read_input_token_cost": 2.8e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/kimi-k2.7-code": {
+        "cache_read_input_token_cost": 1.9e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 16384,
+        "max_tokens": 16384,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "dashscope/qwen-coder": {
         "input_cost_per_token": 3e-07,
         "litellm_provider": "dashscope",
@@ -13733,6 +13838,23 @@
             }
         ]
     },
+    "dashscope/qwen3.8-max": {
+        "cache_read_input_token_cost": 2.5e-07,
+        "input_cost_per_token": 2e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 991808,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 6e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "dashscope/qwq-plus": {
         "input_cost_per_token": 8e-07,
         "litellm_provider": "dashscope",
@@ -15430,6 +15552,17 @@
         "supports_tool_choice": true,
         "supports_function_calling": true
     },
+    "deepinfra/nvidia/NVIDIA-Nemotron-3.5-Lightning": {
+        "max_input_tokens": 262144,
+        "input_cost_per_token": 5e-08,
+        "output_cost_per_token": 2e-07,
+        "litellm_provider": "deepinfra",
+        "mode": "chat",
+        "source": "https://deepinfra.com/nvidia/NVIDIA-Nemotron-3.5-Lightning",
+        "supports_tool_choice": true,
+        "supports_function_calling": true,
+        "supports_reasoning": true
+    },
     "deepinfra/nvidia/NVIDIA-Nemotron-Nano-9B-v2": {
         "max_tokens": 131072,
         "max_input_tokens": 131072,
@@ -18888,6 +19021,60 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "vertex_ai/gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "vertex_ai",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "vertex_ai/gemini-3.1-pro-preview": {
         "cache_read_input_token_cost": 2e-07,
         "cache_read_input_token_cost_above_200k_tokens": 4e-07,
@@ -20563,6 +20750,63 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "gemini/gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "gemini",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "rpm": 2000,
+        "source": "https://ai.google.dev/pricing/gemini-3",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_output": false,
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "tpm": 800000,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "gemini/gemini-omni-flash-preview": {
         "input_cost_per_audio_token": 1.5e-06,
         "input_cost_per_token": 1.5e-06,
@@ -20898,6 +21142,61 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "vertex_ai-language-models",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "source": "https://ai.google.dev/pricing/gemini-3",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_output": false,
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "gemini/gemini-2.5-pro-preview-tts": {
         "cache_read_input_token_cost": 1.25e-07,
         "cache_read_input_token_cost_above_200k_tokens": 2.5e-07,
@@ -25987,11 +26286,12 @@
         "supports_vision": true
     },
     "groq/llama-3.1-8b-instant": {
+        "deprecation_date": "2026-08-16",
         "input_cost_per_token": 5e-08,
         "litellm_provider": "groq",
-        "max_input_tokens": 128000,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
+        "max_input_tokens": 131072,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
         "mode": "chat",
         "output_cost_per_token": 8e-08,
         "supports_function_calling": true,
@@ -25999,9 +26299,10 @@
         "supports_tool_choice": true
     },
     "groq/llama-3.3-70b-versatile": {
+        "deprecation_date": "2026-08-16",
         "input_cost_per_token": 5.9e-07,
         "litellm_provider": "groq",
-        "max_input_tokens": 128000,
+        "max_input_tokens": 131072,
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
@@ -26022,7 +26323,28 @@
         "supports_response_schema": false,
         "supports_tool_choice": true
     },
+    "groq/meta-llama/llama-prompt-guard-2-22m": {
+        "input_cost_per_token": 3e-08,
+        "litellm_provider": "groq",
+        "max_input_tokens": 512,
+        "max_output_tokens": 512,
+        "max_tokens": 512,
+        "mode": "chat",
+        "output_cost_per_token": 3e-08,
+        "source": "https://console.groq.com/docs/models"
+    },
+    "groq/meta-llama/llama-prompt-guard-2-86m": {
+        "input_cost_per_token": 4e-08,
+        "litellm_provider": "groq",
+        "max_input_tokens": 512,
+        "max_output_tokens": 512,
+        "max_tokens": 512,
+        "mode": "chat",
+        "output_cost_per_token": 4e-08,
+        "source": "https://console.groq.com/docs/model/meta-llama/llama-prompt-guard-2-86m"
+    },
     "groq/meta-llama/llama-guard-4-12b": {
+        "deprecation_date": "2026-03-05",
         "input_cost_per_token": 2e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 8192,
@@ -26032,6 +26354,7 @@
         "output_cost_per_token": 2e-07
     },
     "groq/meta-llama/llama-4-maverick-17b-128e-instruct": {
+        "deprecation_date": "2026-03-09",
         "input_cost_per_token": 2e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
@@ -26045,6 +26368,7 @@
         "supports_vision": true
     },
     "groq/meta-llama/llama-4-scout-17b-16e-instruct": {
+        "deprecation_date": "2026-07-17",
         "input_cost_per_token": 1.1e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
@@ -26058,6 +26382,7 @@
         "supports_vision": true
     },
     "groq/moonshotai/kimi-k2-instruct-0905": {
+        "deprecation_date": "2026-04-15",
         "input_cost_per_token": 1e-06,
         "output_cost_per_token": 3e-06,
         "cache_read_input_token_cost": 5e-07,
@@ -26075,8 +26400,8 @@
         "input_cost_per_token": 1.5e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
-        "max_output_tokens": 32766,
-        "max_tokens": 32766,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
         "mode": "chat",
         "output_cost_per_token": 6e-07,
         "search_context_cost_per_query": {
@@ -26096,8 +26421,8 @@
         "input_cost_per_token": 7.5e-08,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
-        "max_output_tokens": 32768,
-        "max_tokens": 32768,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
         "mode": "chat",
         "output_cost_per_token": 3e-07,
         "search_context_cost_per_query": {
@@ -26132,7 +26457,26 @@
         "supports_tool_choice": true,
         "supports_web_search": true
     },
+    "groq/canopylabs/orpheus-v1-english": {
+        "input_cost_per_character": 2.2e-05,
+        "litellm_provider": "groq",
+        "max_input_tokens": 4000,
+        "max_output_tokens": 50000,
+        "max_tokens": 50000,
+        "mode": "audio_speech",
+        "source": "https://console.groq.com/docs/model/canopylabs/orpheus-v1-english"
+    },
+    "groq/canopylabs/orpheus-arabic-saudi": {
+        "input_cost_per_character": 4e-05,
+        "litellm_provider": "groq",
+        "max_input_tokens": 4000,
+        "max_output_tokens": 50000,
+        "max_tokens": 50000,
+        "mode": "audio_speech",
+        "source": "https://console.groq.com/docs/models"
+    },
     "groq/playai-tts": {
+        "deprecation_date": "2025-12-31",
         "input_cost_per_character": 5e-05,
         "litellm_provider": "groq",
         "max_input_tokens": 10000,
@@ -26140,7 +26484,23 @@
         "max_tokens": 10000,
         "mode": "audio_speech"
     },
+    "groq/qwen/qwen3.6-27b": {
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "groq",
+        "max_input_tokens": 131072,
+        "max_output_tokens": 16384,
+        "max_tokens": 16384,
+        "mode": "chat",
+        "output_cost_per_token": 3e-06,
+        "source": "https://console.groq.com/docs/model/qwen/qwen3.6-27b",
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": false,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "groq/qwen/qwen3-32b": {
+        "deprecation_date": "2026-07-17",
         "input_cost_per_token": 2.9e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131000,
@@ -31632,6 +31992,17 @@
         "supports_video_input": true,
         "supports_vision": true
     },
+    "openrouter/nvidia/nemotron-3.5-lightning": {
+        "input_cost_per_token": 5e-08,
+        "litellm_provider": "openrouter",
+        "max_input_tokens": 262144,
+        "mode": "chat",
+        "output_cost_per_token": 2e-07,
+        "source": "https://openrouter.ai/nvidia/nemotron-3.5-lightning",
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_tool_choice": true
+    },
     "openrouter/openai/gpt-3.5-turbo": {
         "input_cost_per_token": 1.5e-06,
         "litellm_provider": "openrouter",
@@ -35769,6 +36140,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true
     },
     "eu.deepseek.v3.2": {
@@ -35781,6 +36153,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true
     },
     "us.meta.llama3-1-405b-instruct-v1:0": {
@@ -45680,11 +46053,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-sol": {
         "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
         "cache_creation_input_token_cost": 6.875e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
         "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
         "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -45708,11 +46085,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-terra": {
         "input_cost_per_token": 2.2e-06,
+        "input_cost_per_token_above_272k_tokens": 4.4e-06,
         "cache_creation_input_token_cost": 2.75e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06,
         "cache_read_input_token_cost": 2.2e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 4.4e-07,
         "output_cost_per_token": 1.32e-05,
+        "output_cost_per_token_above_272k_tokens": 1.98e-05,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -45736,11 +46117,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-luna": {
         "input_cost_per_token": 2.2e-07,
+        "input_cost_per_token_above_272k_tokens": 4.4e-07,
         "cache_creation_input_token_cost": 2.75e-07,
+        "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07,
         "cache_read_input_token_cost": 2.2e-08,
+        "cache_read_input_token_cost_above_272k_tokens": 4.4e-08,
         "output_cost_per_token": 1.32e-06,
+        "output_cost_per_token_above_272k_tokens": 1.98e-06,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -46057,6 +46442,7 @@
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_system_messages": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -46071,6 +46457,7 @@
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_system_messages": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },

litellm/proxy/_types.py

@@ -3,7 +3,7 @@ import json
 import os
 from collections.abc import Callable, Mapping
 from datetime import datetime
-from typing import TYPE_CHECKING, Any, Final, Literal
+from typing import TYPE_CHECKING, Any, Final, Literal, NamedTuple
 
 import httpx
 from pydantic import (
@@ -18,7 +18,7 @@ from pydantic import (
 from typing_extensions import NotRequired, Required, TypedDict
 
 from litellm._uuid import uuid
-from litellm.constants import MCP_STDIO_ALLOWED_COMMANDS
+from litellm.constants import DEFAULT_STAGGER_WINDOW_SECONDS, MCP_STDIO_ALLOWED_COMMANDS
 from litellm.litellm_core_utils.initialize_dynamic_callback_params import (
     validate_no_callback_env_reference,
 )
@@ -73,6 +73,27 @@ else:
     Span = Any
 
 
+class ReconcileOutcome(NamedTuple):
+    """What a model reconcile observed, captured while it still held the reconcile
+    lock.
+
+    Both fields have to be read under that lock to be worth anything. ``live_after``
+    in particular is the router's serving state the instant this reconcile finished,
+    which is NOT the same as what a later snapshot would see: any other model write
+    admitted in between briefly un-serves every db model (see ``clear_cache``), so a
+    caller that re-snapshots at verdict time can observe that hole and blame its own
+    reload for it.
+
+    - ``still_desired``: the db + config ids the reconcile reconciled against, or None
+      when no reconcile ran and the desired set is therefore unknown.
+    - ``live_after``: the ids the router served immediately after the reconcile, or
+      None when no reconcile ran.
+    """
+
+    still_desired: frozenset[str] | None
+    live_after: frozenset[str] | None
+
+
 class SupportedDBObjectType(str, enum.Enum):
     """
     Supported database object types for fine-grained DB storage control.
@@ -2251,6 +2272,39 @@ class CoordinationRedisParams(LiteLLMPydanticObjectBase):
         return any(value is not None for value in (self.host, self.url, self.startup_nodes, self.sentinel_nodes))
 
 
+class ScheduledJobStaggerSettings(LiteLLMPydanticObjectBase):
+    """
+    Spreads the proxy's scheduled background jobs across a window instead of firing them
+    all on one instant, on every replica, forever.
+    """
+
+    model_config = ConfigDict(frozen=True, extra="forbid", protected_namespaces=())
+
+    enabled: bool = Field(default=True, description="apply deterministic phase offsets to scheduled background jobs")
+    window_seconds: int = Field(
+        default=DEFAULT_STAGGER_WINDOW_SECONDS,
+        ge=0,
+        description=(
+            "width of the window jobs are spread over. An interval job is never offset by "
+            "more than one of its own periods, so it is not delayed past the wait it already has"
+        ),
+    )
+    identity: str | None = Field(
+        default=None,
+        description=(
+            "replaces the POD_NAME/HOSTNAME-derived component of the offset hash. Set this "
+            "when replicas share a hostname and would otherwise land on the same offset"
+        ),
+    )
+    offsets: Mapping[str, int] = Field(
+        default_factory=dict,
+        description=(
+            "explicit offset in seconds per scheduler job id, overriding the derived value. "
+            "0 pins a job to its unshifted schedule"
+        ),
+    )
+
+
 class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
     """
     Documents all the fields supported by `general_settings` in config.yaml
@@ -2437,6 +2491,14 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
         None,
         description="By default, the user calling /team/new is automatically added to the new team as a team admin. If True, proxy admins are no longer auto-added; members explicitly listed in members_with_roles are unaffected. Default is False.",
     )
+    scheduled_job_stagger: ScheduledJobStaggerSettings | None = Field(
+        None,
+        description=(
+            "Spreads the proxy's scheduled background jobs (spend flushes, budget resets, "
+            "config reloads, exports) across a window instead of firing them together on "
+            "every replica. On by default; set to tune the window, pin a job, or turn it off."
+        ),
+    )
     maximum_spend_logs_retention_period: str | None = Field(
         None,
         description="Maximum retention period for spend logs (e.g., '7d' for 7 days). Logs older than this will be deleted.",

litellm/proxy/auth/model_checks.py

@@ -52,20 +52,20 @@ def _get_models_from_access_groups(
     model_access_groups: dict[str, list[str]],
     all_models: list[str],
     include_model_access_groups: bool | None = False,
+    proxy_model_list: Sequence[str] | None = None,
 ) -> list[str]:
-    idx_to_remove: Final = []
-    new_models: Final = []
-    for idx, model in enumerate(all_models):
-        if model in model_access_groups:
-            if not include_model_access_groups:  # remove access group, unless requested - e.g. when creating a key
-                idx_to_remove.append(idx)
-            new_models.extend(model_access_groups[model])
-
-    for idx in sorted(idx_to_remove, reverse=True):
-        all_models.pop(idx)
-
-    all_models.extend(new_models)
-    return all_models
+    # a grant naming both a deployed model and an access group means both at runtime
+    # (_check_model_access_helper unions them), so listings must keep the literal too
+    deployed_model_names: Final = frozenset(proxy_model_list or ())
+    kept_models: Final = [
+        model
+        for model in all_models
+        if model not in model_access_groups or include_model_access_groups or model in deployed_model_names
+    ]
+    member_models: Final = [
+        member for model in all_models if model in model_access_groups for member in model_access_groups[model]
+    ]
+    return kept_models + member_models
 
 
 async def get_mcp_server_ids(
@@ -128,6 +128,7 @@ def get_key_models(
         model_access_groups=model_access_groups,
         all_models=all_models,
         include_model_access_groups=include_model_access_groups,
+        proxy_model_list=proxy_model_list,
     )
 
     # deduplicate while preserving order
@@ -169,6 +170,7 @@ def get_team_models(
         model_access_groups=model_access_groups,
         all_models=list(all_models_set),
         include_model_access_groups=include_model_access_groups,
+        proxy_model_list=proxy_model_list,
     )
 
     # deduplicate while preserving order

litellm/proxy/auth/user_api_key_auth.py

@@ -1060,6 +1060,31 @@ async def _read_request_body_deferring_parse_failure(
     return populate_request_with_path_params(request_data=parsed_body, request=request), None
 
 
+async def _record_unparsable_body_failure(
+    user_api_key_dict: UserAPIKeyAuth,
+    body_parse_exception: ProxyException,
+    route: str,
+) -> None:
+    """Record the 400 an unparsable body earns as a failed request log.
+
+    The endpoint never runs for these, so no downstream failure hook writes the
+    spend log row the Admin UI reads. Logging must not change what the caller
+    sees, so a failure here is swallowed and the 400 is raised either way.
+    """
+    from litellm.proxy.proxy_server import proxy_logging_obj
+
+    try:
+        await proxy_logging_obj.post_call_failure_hook(  # pyright: ignore[reportUnknownMemberType]  # bare dict in sig
+            request_data={},  # mutable-ok: the failure hook seeds the call id and metadata onto this dict
+            original_exception=body_parse_exception,
+            user_api_key_dict=user_api_key_dict,
+            error_type=ProxyErrorTypes.bad_request_error,
+            route=route,
+        )
+    except Exception as e:  # noqa: BLE001  # any logging failure must leave the caller's 400 untouched
+        verbose_proxy_logger.exception("Failed to log the request rejected for an unparsable body: %s", e)
+
+
 async def _user_api_key_auth_builder(
     request: Request,
     api_key: str,
@@ -2673,6 +2698,11 @@ async def user_api_key_auth(
     user_api_key_auth_obj.request_route = normalize_request_route(route)
 
     if body_parse_exception is not None:
+        await _record_unparsable_body_failure(
+            user_api_key_dict=user_api_key_auth_obj,
+            body_parse_exception=body_parse_exception,
+            route=route,
+        )
         raise body_parse_exception
 
     # Resolve caller identity once, here at the seam, into a single per-request

litellm/proxy/common_utils/callback_utils.py

@@ -6,7 +6,11 @@ from typing import TYPE_CHECKING, Any, Final, Optional
 import litellm
 from litellm import get_secret
 from litellm._logging import verbose_proxy_logger
-from litellm.constants import PRE_CALL_EXECUTED_GUARDRAILS_KEY, SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY
+from litellm.constants import (
+    CONSUMED_REQUEST_TAGS_METADATA_KEY,
+    PRE_CALL_EXECUTED_GUARDRAILS_KEY,
+    SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
+)
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.core_helpers import (
     get_metadata_variable_name_from_kwargs,
@@ -426,6 +430,7 @@ LITELLM_PROXY_INTERNAL_METADATA_KEYS: Final = frozenset(
         "_pipeline_managed_guardrails",
         PRE_CALL_EXECUTED_GUARDRAILS_KEY,
         SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
+        CONSUMED_REQUEST_TAGS_METADATA_KEY,
         "disable_global_guardrails",
         "disable_global_guardrail",
         "opted_out_global_guardrails",

litellm/proxy/common_utils/scheduled_job_stagger.py

@@ -0,0 +1,347 @@
+"""
+Deterministic phase offsets for the proxy's scheduled background jobs.
+
+APScheduler anchors an ``interval`` job at ``now + interval``, so every job registered in
+the same startup shares one firing instant for the life of the process, and every replica
+brought up by the same rollout shares it too. The result is a burst: each tick, every job
+on every replica queries Postgres at the same moment, competing with the request path for
+the connection pool. The product's own daily/monthly crons are worse still, since they name
+a wall-clock instant that is identical on every replica by construction.
+
+The fix is a phase offset derived from ``sha256(job_id, identity)``, where ``identity``
+covers the pod and the worker process. Different jobs get different offsets, different
+replicas get different offsets for the same job, and nothing collapses back onto a shared
+instant after a restart. Hashing rather than randomising keeps a given process's schedule
+stable for its whole life and lets the applied offsets be logged once and reasoned about
+later.
+
+The offset lives in the trigger rather than in a one-off ``next_run_time`` because a cron
+trigger recomputes each fire from the wall clock and would otherwise snap straight back
+onto the shared instant after its first shifted run.
+
+Only schedules LiteLLM itself chose are shifted. Interval jobs are always eligible; cron
+jobs only when their id is one of the product's own defaults, so an operator-supplied
+crontab keeps the exact instant it asks for. A job whose call site passed an explicit
+``next_run_time`` already anchors itself and is left alone.
+"""
+
+# apscheduler ships no type information, so its imports have no stubs. The Protocols below
+# narrow everything it hands back, which is why this is the only diagnostic left to silence.
+# pyright: reportMissingTypeStubs=false
+
+import hashlib
+import os
+import socket
+from collections.abc import Callable, Mapping, Sequence
+from datetime import datetime, timedelta
+from types import MappingProxyType
+from typing import Final, Protocol
+
+from apscheduler.events import EVENT_JOB_SUBMITTED
+from apscheduler.triggers.base import BaseTrigger
+from apscheduler.triggers.interval import IntervalTrigger
+from pydantic import ValidationError
+
+from litellm._logging import verbose_proxy_logger
+from litellm._uuid import uuid
+from litellm.constants import (
+    MONTHLY_SPEND_REPORT_JOB_ID,
+    PROMETHEUS_FALLBACK_STATS_JOB_ID,
+    PTU_ROLLUP_JOB_ID,
+    PTU_ROLLUP_LOCK_TTL_SECONDS,
+)
+from litellm.proxy._types import ScheduledJobStaggerSettings
+
+GENERAL_SETTINGS_KEY: Final = "scheduled_job_stagger"
+
+#: Cron schedules LiteLLM picks on the operator's behalf, so shifting them changes nothing the
+#: operator asked for. Every other cron trigger is an operator-supplied crontab, preserved exactly.
+#:
+#: The value is the span over which a second firing would redo work the first already did, which
+#: is how long each job's leader-election lock stays held. Two replicas further apart than that
+#: both find the key free and both run, which for the spend report means the customer gets it
+#: twice. Offsets for these jobs are bounded by it, so widening the window cannot resurrect the
+#: duplicate-work failure this feature exists to avoid.
+DEFAULT_CRON_DEDUPE_SECONDS: Final = MappingProxyType(
+    {
+        MONTHLY_SPEND_REPORT_JOB_ID: 3600,
+        PROMETHEUS_FALLBACK_STATS_JOB_ID: 3600,
+        PTU_ROLLUP_JOB_ID: PTU_ROLLUP_LOCK_TTL_SECONDS,
+    }
+)
+
+
+class Trigger(Protocol):
+    """The one method APScheduler asks a trigger for"""
+
+    def get_next_fire_time(self, previous_fire_time: datetime | None, now: datetime) -> datetime | None: ...
+
+
+class ScheduledJob(Protocol):
+    @property
+    def id(self) -> str: ...
+
+    @property
+    def trigger(self) -> Trigger: ...
+
+
+class JobScheduler(Protocol):
+    """The slice of ``AsyncIOScheduler`` this module uses, which ships no type information"""
+
+    @property
+    def running(self) -> bool: ...
+
+    def get_jobs(self) -> Sequence[ScheduledJob]: ...
+
+    def modify_job(self, job_id: str, *, trigger: Trigger) -> object: ...
+
+    def add_listener(self, callback: Callable[["JobSubmission"], None], mask: int = ...) -> None: ...
+
+
+class JobSubmission(Protocol):
+    """An ``EVENT_JOB_SUBMITTED`` event"""
+
+    @property
+    def job_id(self) -> str: ...
+
+    @property
+    def scheduled_run_times(self) -> Sequence[datetime]: ...
+
+
+class _OffsetTrigger:
+    """
+    Delegates to ``base`` on a clock rolled back by ``offset``, then rolls the answer
+    forward again, so every fire lands exactly ``offset`` later than it otherwise would
+    while the underlying schedule keeps its own semantics.
+
+    Composed rather than derived from ``BaseTrigger``: APScheduler only ever asks a trigger
+    for its next fire time, and it accepts this by virtual registration below.
+    """
+
+    __slots__ = ("base", "offset")
+
+    def __init__(self, base: Trigger, offset: timedelta) -> None:
+        self.base = base
+        self.offset = offset
+
+    def get_next_fire_time(self, previous_fire_time: datetime | None, now: datetime) -> datetime | None:
+        shifted_previous: Final = None if previous_fire_time is None else previous_fire_time - self.offset
+        next_fire_time: Final = self.base.get_next_fire_time(shifted_previous, now - self.offset)
+        return None if next_fire_time is None else next_fire_time + self.offset
+
+    def __str__(self) -> str:
+        return f"{self.base}[+{int(self.offset.total_seconds())}s]"
+
+
+# APScheduler type-checks assigned triggers with isinstance, so it has to accept this one
+BaseTrigger.register(_OffsetTrigger)
+
+
+def parse_stagger_settings(general_settings: Mapping[str, object]) -> ScheduledJobStaggerSettings:
+    raw: Final = general_settings.get(GENERAL_SETTINGS_KEY)
+    if raw is None:
+        return ScheduledJobStaggerSettings()
+    try:
+        return ScheduledJobStaggerSettings.model_validate(raw)
+    except ValidationError as exc:
+        verbose_proxy_logger.warning(
+            "Ignoring invalid general_settings.%s, falling back to defaults: %s",
+            GENERAL_SETTINGS_KEY,
+            exc,
+        )
+        return ScheduledJobStaggerSettings()
+
+
+def resolve_stagger_identity(configured: str | None) -> str:
+    """
+    The value hashed alongside a job id to place this process in the stagger window.
+
+    The process id is part of it because a pod runs one scheduler per uvicorn worker, and
+    workers sharing a hostname would otherwise all land on the same offset. That makes the
+    offsets change across restarts, which is what stops a simultaneous rollout from
+    reconverging; the applied values are logged so a given run stays explainable.
+    """
+    host: Final = configured or os.getenv("POD_NAME") or os.getenv("HOSTNAME") or _hostname()
+    return f"{host}:{os.getpid()}"
+
+
+def _hostname() -> str:
+    try:
+        return socket.gethostname()
+    except OSError:
+        return str(uuid.uuid4())
+
+
+def offset_seconds(*, job_id: str, identity: str, window_seconds: int) -> int:
+    """A stable point in ``[0, window_seconds)`` for this job on this process"""
+    if window_seconds <= 0:
+        return 0
+    digest: Final = hashlib.sha256(f"{job_id}\x00{identity}".encode()).digest()
+    return int.from_bytes(digest[:8], "big") % window_seconds
+
+
+def _interval_seconds(job: ScheduledJob) -> int | None:
+    if not isinstance(job.trigger, IntervalTrigger):
+        return None
+    interval: Final = getattr(job.trigger, "interval", None)
+    return int(interval.total_seconds()) if isinstance(interval, timedelta) else None
+
+
+def _is_staggerable(job: ScheduledJob) -> bool:
+    if hasattr(job, "next_run_time"):
+        # the call site anchored the first fire itself
+        return False
+    if _interval_seconds(job) is not None:
+        return True
+    return job.id in DEFAULT_CRON_DEDUPE_SECONDS
+
+
+def _window_for(*, job_id: str, period_seconds: int | None, settings: ScheduledJobStaggerSettings) -> int:
+    """
+    Exclusive upper bound on this job's offset. An interval job is never offset by more than
+    one of its own periods, so it is not delayed past the wait it already had, and a
+    leader-elected cron is never offset past the span in which a second replica would redo
+    its work.
+    """
+    limits: Final = (settings.window_seconds, period_seconds, DEFAULT_CRON_DEDUPE_SECONDS.get(job_id))
+    return min(limit for limit in limits if limit is not None)
+
+
+def _clamped_override(*, job_id: str, requested: int) -> int:
+    horizon: Final = DEFAULT_CRON_DEDUPE_SECONDS.get(job_id)
+    if horizon is None or requested < horizon:
+        return requested
+    verbose_proxy_logger.warning(
+        "general_settings.%s.offsets[%s]=%ss would place replicas more than %ss apart, "
+        "which is long enough for a second replica to redo the run; using %ss instead",
+        GENERAL_SETTINGS_KEY,
+        job_id,
+        requested,
+        horizon,
+        horizon - 1,
+    )
+    return horizon - 1
+
+
+def _offset_for(
+    *,
+    job_id: str,
+    period_seconds: int | None,
+    staggerable: bool,
+    settings: ScheduledJobStaggerSettings,
+    identity: str,
+) -> int:
+    override: Final = settings.offsets.get(job_id)
+    if override is not None:
+        return _clamped_override(job_id=job_id, requested=max(0, override))
+    if not staggerable:
+        return 0
+    return offset_seconds(
+        job_id=job_id,
+        identity=identity,
+        window_seconds=_window_for(job_id=job_id, period_seconds=period_seconds, settings=settings),
+    )
+
+
+def stagger_trigger(
+    *,
+    job_id: str,
+    trigger: Trigger,
+    period_seconds: int | None,
+    settings: ScheduledJobStaggerSettings,
+    identity: str | None = None,
+) -> Trigger:
+    """
+    The trigger a job should carry, shifted by its own share of the window.
+
+    For a job registered against an already-running scheduler, which the startup sweep cannot
+    reach: every job carries a ``next_run_time`` by then, so re-running the sweep would treat
+    them all as self-anchored and change nothing.
+    """
+    offset: Final = _offset_for(
+        job_id=job_id,
+        period_seconds=period_seconds,
+        staggerable=True,
+        settings=settings,
+        identity=identity or resolve_stagger_identity(settings.identity),
+    )
+    return trigger if offset == 0 else _OffsetTrigger(trigger, timedelta(seconds=offset))
+
+
+def apply_scheduled_job_stagger(
+    *,
+    scheduler: JobScheduler,
+    settings: ScheduledJobStaggerSettings,
+    identity: str | None = None,
+) -> Mapping[str, int]:
+    """
+    Shift each eligible job's schedule by its own offset. Call this once, after every job is
+    registered and before the scheduler starts, so the offset is folded into the first fire
+    rather than applied to a schedule already running.
+
+    ``identity`` is resolved from the environment when the caller does not supply one.
+
+    Returns the offset applied to every registered job, including the zeroes, so the caller
+    and the logs describe the same thing.
+    """
+    resolved_identity: Final = identity or resolve_stagger_identity(settings.identity)
+    if scheduler.running:
+        # every job already carries a next_run_time by now, so the sweep would skip all of
+        # them and report success while changing nothing
+        verbose_proxy_logger.warning(
+            "Scheduled job stagger skipped: the scheduler is already running, so offsets must be "
+            "applied before it starts"
+        )
+        return MappingProxyType({job.id: 0 for job in scheduler.get_jobs()})
+    if not settings.enabled:
+        verbose_proxy_logger.info(
+            "Scheduled job stagger disabled via general_settings.%s; all jobs keep their unshifted schedule",
+            GENERAL_SETTINGS_KEY,
+        )
+        return MappingProxyType({job.id: 0 for job in scheduler.get_jobs()})
+
+    offsets: Final = MappingProxyType(
+        {
+            job.id: _offset_for(
+                job_id=job.id,
+                period_seconds=_interval_seconds(job),
+                staggerable=_is_staggerable(job),
+                settings=settings,
+                identity=resolved_identity,
+            )
+            for job in scheduler.get_jobs()
+        }
+    )
+    for job in scheduler.get_jobs():
+        if offsets[job.id] > 0:
+            scheduler.modify_job(
+                job.id,
+                trigger=_OffsetTrigger(job.trigger, timedelta(seconds=offsets[job.id])),
+            )
+
+    verbose_proxy_logger.info(
+        "Scheduled job stagger applied (identity=%s, window=%ss): %s",
+        resolved_identity,
+        settings.window_seconds,
+        ", ".join(f"{job_id}=+{seconds}s" for job_id, seconds in sorted(offsets.items())),
+    )
+    return offsets
+
+
+def attach_job_timing_logger(scheduler: JobScheduler) -> None:
+    """Log each fire's scheduled instant against the instant it actually started"""
+    scheduler.add_listener(_log_job_submitted, EVENT_JOB_SUBMITTED)
+
+
+def _log_job_submitted(event: JobSubmission) -> None:
+    if not event.scheduled_run_times:
+        return
+    scheduled: Final = event.scheduled_run_times[0]
+    started: Final = datetime.now(scheduled.tzinfo)
+    verbose_proxy_logger.debug(
+        "Scheduled job %s started: scheduled_run_time=%s actual_start_time=%s delay=%.3fs",
+        event.job_id,
+        scheduled.isoformat(),
+        started.isoformat(),
+        (started - scheduled).total_seconds(),
+    )

litellm/proxy/common_utils/sse_keepalive.py

@@ -21,6 +21,17 @@ def _coerce_interval(ping_interval_seconds: float | str | None) -> float | None:
     return interval
 
 
+def keepalive_ping_has_fired(elapsed_seconds: float, ping_interval_seconds: float | str | None) -> bool:
+    """Whether a keepalive ping has already gone out, which flushes the response headers.
+
+    A caller that discovers a failure after that point cannot raise its way to the client, since
+    the status line is already on the wire. With pings disabled nothing flushes early, so a raise
+    still carries its real status.
+    """
+    interval: Final = _coerce_interval(ping_interval_seconds)
+    return interval is not None and elapsed_seconds >= interval
+
+
 def wrap_sse_stream_with_keepalive_pings(
     stream: AsyncGenerator[str, None],
     ping_interval_seconds: float | str | None,

litellm/proxy/guardrails/anthropic_sse.py

@@ -0,0 +1,125 @@
+"""Anthropic SSE <-> ModelResponse conversion for guardrail streaming hooks.
+
+`/v1/messages` streams reach a guardrail's `async_post_call_streaming_iterator_hook` as raw SSE
+frames rather than chunk objects, which `stream_chunk_builder` cannot assemble. These helpers let a
+hook scan such a stream, and re-emit it when the guardrail rewrote the response.
+"""
+
+from __future__ import annotations
+
+import json
+from collections.abc import Mapping, Sequence
+from typing import Final
+
+from litellm.types.utils import Choices, ModelResponse
+
+
+def is_raw_sse_stream(all_chunks: Sequence[object]) -> bool:
+    return any(isinstance(chunk, (str, bytes)) for chunk in all_chunks)
+
+
+def _joined_sse_stream(all_chunks: Sequence[object]) -> str | None:
+    raw: Final = b"".join(
+        chunk if isinstance(chunk, bytes) else chunk.encode("utf-8")
+        for chunk in all_chunks
+        if isinstance(chunk, (str, bytes))
+    )
+    try:
+        return raw.decode("utf-8")
+    except UnicodeDecodeError:
+        return None
+
+
+def _anthropic_message_start(sse_stream: str) -> Mapping[str, object] | None:
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
+        AnthropicPassthroughLoggingHandler,
+    )
+
+    return next(
+        (
+            message
+            for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream)  # pyright: ignore[reportPrivateUsage]  # same parser the assembler uses
+            if (event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None  # pyright: ignore[reportPrivateUsage]  # same parser the assembler uses; a private import beats forking SSE parsing
+            and event_data.get("type") == "message_start"
+            and isinstance(message := event_data.get("message"), dict)
+        ),
+        None,
+    )
+
+
+def assemble_anthropic_sse_stream(
+    all_chunks: Sequence[object], *, restore_identity: bool = False
+) -> ModelResponse | None:
+    """Assemble raw Anthropic SSE frames into a ModelResponse.
+
+    ``restore_identity`` stamps the upstream message id and model onto the result, which the
+    assembler does not carry through. It is off by default so callers that re-emit the assembled
+    response keep the wire shape they had before this helper was shared. The writes land on a
+    freshly built object that is unreachable from caller state until returned.
+    """
+    from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
+        AnthropicPassthroughLoggingHandler,
+    )
+
+    sse_stream: Final = _joined_sse_stream(all_chunks)
+    if sse_stream is None:
+        return None
+    message_start: Final = _anthropic_message_start(sse_stream)
+    if message_start is None:
+        return None
+    model: Final = message_start.get("model") if restore_identity else None
+    try:
+        assembled: Final = AnthropicPassthroughLoggingHandler._build_complete_streaming_response(  # pyright: ignore[reportPrivateUsage]  # the only SSE-to-ModelResponse assembler; reimplementing it here would fork the parser
+            all_chunks=(sse_stream,),
+            litellm_logging_obj=None,  # pyright: ignore[reportArgumentType]  # only forwarded to stream_chunk_builder, which accepts None
+            model=model if isinstance(model, str) else "",
+        )
+    except Exception:  # noqa: BLE001  # stream_chunk_builder re-raises every assembly failure as litellm.APIError
+        return None
+    if not isinstance(assembled, ModelResponse):
+        return None
+    if not restore_identity:
+        return assembled
+    message_id: Final = message_start.get("id")
+    if isinstance(message_id, str):
+        assembled.id = message_id
+    if isinstance(model, str) and model:
+        assembled.model = model
+    return assembled
+
+
+def model_response_text(response: ModelResponse) -> str:
+    """Assistant text of a response, used to detect whether a guardrail rewrote it."""
+    return "".join(
+        choice.message.content
+        for choice in response.choices
+        if isinstance(choice, Choices)  # pyright: ignore[reportUnnecessaryIsInstance]  # runtime choices can be StreamingChoices
+        and isinstance(choice.message.content, str)
+    )
+
+
+def anthropic_sse_error_frames(message: str) -> tuple[bytes, ...]:
+    """Anthropic error event, for a failure discovered after the response headers were flushed.
+
+    Once a keepalive ping has been sent a raise cannot reach the client, so the failure has to
+    travel as a frame.
+    """
+    body: Final = json.dumps(message)
+    return (
+        f'event: error\ndata: {{"type": "error", "error": {{"type": "guardrail_error", '
+        f'"message": {body}}}}}\n\n'.encode(),
+    )
+
+
+def anthropic_sse_chunks_from_response(assembled: ModelResponse) -> tuple[bytes, ...]:
+    from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
+        LiteLLMAnthropicMessagesAdapter,
+    )
+    from litellm.llms.anthropic.experimental_pass_through.messages.fake_stream_iterator import (
+        FakeAnthropicMessagesStreamIterator,
+    )
+
+    anthropic_response: Final = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
+        response=assembled
+    )
+    return tuple(FakeAnthropicMessagesStreamIterator(response=anthropic_response).chunks)

litellm/proxy/guardrails/guardrail_hooks/bedrock_guardrails.py

@@ -14,6 +14,7 @@ import copy
 import json
 import re
 import sys
+import time
 from collections.abc import AsyncGenerator, Mapping, Sequence
 from datetime import datetime, timezone
 from itertools import accumulate, groupby
@@ -30,6 +31,7 @@ from litellm.constants import BEDROCK_APPLY_GUARDRAIL_CHUNK_BUDGET_CHARS
 from litellm.exceptions import ModifyResponseException
 from litellm.integrations.custom_guardrail import CustomGuardrail
 from litellm.litellm_core_utils.core_helpers import redact_nested_match_and_regex_keys
+from litellm.llms.anthropic.chat.guardrail_translation.handler import AnthropicMessagesHandler
 from litellm.llms.base_llm.guardrail_translation.utils import (
     effective_scan_only_tool_results_for_guardrail,
 )
@@ -39,6 +41,15 @@ from litellm.llms.custom_httpx.http_handler import (
     httpxSpecialProvider,
 )
 from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy.common_request_processing import _serialize_http_exception_detail
+from litellm.proxy.common_utils.sse_keepalive import keepalive_ping_has_fired
+from litellm.proxy.guardrails.anthropic_sse import (
+    anthropic_sse_chunks_from_response,
+    anthropic_sse_error_frames,
+    assemble_anthropic_sse_stream,
+    is_raw_sse_stream,
+    model_response_text,
+)
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.guardrails import BedrockChecksConfigModel, GuardrailEventHooks
 from litellm.types.llms.openai import AllMessageValues, ChatCompletionUserMessage
@@ -2578,14 +2589,21 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
         from litellm.types.utils import TextCompletionResponse
 
         # Collect all chunks to process them together
+        started_at: Final = time.monotonic()
         all_chunks: Final[list[ModelResponseStream]] = []
         async for chunk in response:
             all_chunks.append(chunk)
 
-        assembled_model_response: ModelResponse | TextCompletionResponse | None = stream_chunk_builder(
-            chunks=all_chunks,
+        # /v1/messages arrives as SSE frames, which stream_chunk_builder cannot assemble
+        raw_sse: Final = is_raw_sse_stream(all_chunks)
+        assembled_model_response: ModelResponse | TextCompletionResponse | None = (
+            assemble_anthropic_sse_stream(all_chunks, restore_identity=True)
+            if raw_sse
+            else stream_chunk_builder(chunks=all_chunks)
         )
         if isinstance(assembled_model_response, ModelResponse):
+            pre_guardrail_text: Final = model_response_text(assembled_model_response)
+            _pre_block_response: Final = assembled_model_response
             ####################################################################
             ########## 1. Make Bedrock Apply Guardrail API request ##########
             #
@@ -2609,7 +2627,32 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
                     request_data=request_data,
                     logging_event_type=GuardrailEventHooks.post_call,
                 )
+            except HTTPException as block_exc:
+                block_detail: Final = block_exc.detail
+                # A policy block is the only 400 carrying a structured detail; a service failure
+                # either details a plain string or reports a non-400 status. Re-raising a service
+                # failure keeps its real status, but only while the headers are unflushed: past the
+                # first keepalive ping the raise reaches nobody, so it has to travel as a frame too
+                is_block: Final = raw_sse and block_exc.status_code == 400 and isinstance(block_detail, Mapping)
+                headers_flushed: Final = keepalive_ping_has_fired(
+                    time.monotonic() - started_at, litellm.anthropic_sse_ping_interval_seconds
+                )
+                if not raw_sse or (not is_block and not headers_flushed):
+                    raise
+                block_message, _ = _serialize_http_exception_detail(block_detail)
+                for error_frame in anthropic_sse_error_frames(
+                    block_message if is_block else f"{block_exc.status_code}: {block_message}"
+                ):
+                    yield error_frame
+                return
             except ModifyResponseException as e:
+                if raw_sse:
+                    e.model = _pre_block_response.model or e.model  # rebind-ok: exc.model defaults to the guardrail
+                    if e.original_response is None:
+                        e.original_response = _pre_block_response  # rebind-ok: the block builder reads usage off this
+                    for block_chunk in AnthropicMessagesHandler().build_block_sse_chunks(e, stream_started=False):
+                        yield block_chunk
+                    return
                 # Preserve upstream usage from the LLM call we already
                 # consumed. Non-streaming blocks carry it via
                 # ModifyResponseException.original_response +
@@ -2642,11 +2685,29 @@ class BedrockGuardrail(CustomGuardrail, BaseAWSLLM):
             #########################################################################
             ########## 3. Return the (potentially masked) chunks ##########
             #########################################################################
+            if raw_sse:
+                for sse_chunk in (
+                    anthropic_sse_chunks_from_response(assembled_model_response)
+                    if model_response_text(assembled_model_response) != pre_guardrail_text
+                    else all_chunks
+                ):
+                    yield sse_chunk
+                return
+
             mock_response: Final = MockResponseIterator(model_response=assembled_model_response)
 
             # Return the reconstructed stream
             async for chunk in mock_response:
                 yield chunk
+        elif raw_sse:
+            # Forwarding an unscannable stream would silently disable the guardrail, so fail closed.
+            # A raise cannot reach the client once a keepalive ping has flushed the headers, so the
+            # refusal travels as a frame, matching how a block is delivered above
+            for error_frame in anthropic_sse_error_frames(
+                f"{self.guardrail_name}: streamed response could not be assembled for scanning, blocking it"
+            ):
+                yield error_frame
+            return
         else:
             for chunk in all_chunks:
                 yield chunk

litellm/proxy/guardrails/guardrail_hooks/tool_permission.py

@@ -17,6 +17,11 @@ from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.common_utils.callback_utils import (
     add_guardrail_to_applied_guardrails_header,
 )
+from litellm.proxy.guardrails.anthropic_sse import (
+    anthropic_sse_chunks_from_response,
+    assemble_anthropic_sse_stream,
+    is_raw_sse_stream,
+)
 from litellm.types.guardrails import GuardrailEventHooks, LitellmParams
 from litellm.types.proxy.guardrails.guardrail_hooks.tool_permission import (
     PermissionError,
@@ -870,7 +875,7 @@ class ToolPermissionGuardrail(CustomGuardrail):
             all_chunks.append(chunk)
 
         assembled_model_response: Final[ModelResponse | TextCompletionResponse | None] = (
-            stream_chunk_builder(chunks=all_chunks) if not self._is_raw_sse_stream(all_chunks) else None
+            stream_chunk_builder(chunks=all_chunks) if not is_raw_sse_stream(all_chunks) else None
         )
         if isinstance(assembled_model_response, ModelResponse):
             denied_tools = self._check_assembled_stream(assembled_model_response)
@@ -883,9 +888,9 @@ class ToolPermissionGuardrail(CustomGuardrail):
                 yield chunk
             return
 
-        anthropic_response: Final = self._assemble_anthropic_stream(all_chunks)
+        anthropic_response: Final = assemble_anthropic_sse_stream(all_chunks)
         if anthropic_response is None:
-            if self._is_raw_sse_stream(all_chunks):
+            if is_raw_sse_stream(all_chunks):
                 raise GuardrailRaisedException(
                     guardrail_name=self.guardrail_name,
                     message=(
@@ -904,13 +909,9 @@ class ToolPermissionGuardrail(CustomGuardrail):
             return
 
         self._modify_response_with_permission_errors(anthropic_response, anthropic_denials)
-        for sse_chunk in self._rewritten_anthropic_sse_chunks(anthropic_response):
+        for sse_chunk in anthropic_sse_chunks_from_response(anthropic_response):
             yield sse_chunk
 
-    @staticmethod
-    def _is_raw_sse_stream(all_chunks: Sequence[Any]) -> bool:
-        return any(isinstance(chunk, (str, bytes)) for chunk in all_chunks)
-
     def _check_assembled_stream(
         self, assembled: ModelResponse
     ) -> tuple[tuple[ChatCompletionMessageToolCall, PermissionError], ...]:
@@ -924,60 +925,3 @@ class ToolPermissionGuardrail(CustomGuardrail):
         if not denied_tools:
             verbose_proxy_logger.debug("Tool Permission Guardrail Post-Call Hook: All tools allowed")
         return denied_tools
-
-    @staticmethod
-    def _joined_sse_stream(all_chunks: Sequence[Any]) -> str | None:
-        raw: Final = b"".join(
-            chunk if isinstance(chunk, bytes) else chunk.encode("utf-8")
-            for chunk in all_chunks
-            if isinstance(chunk, (str, bytes))
-        )
-        try:
-            return raw.decode("utf-8")
-        except UnicodeDecodeError:
-            return None
-
-    @staticmethod
-    def _has_anthropic_message_start(sse_stream: str) -> bool:
-        from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
-            AnthropicPassthroughLoggingHandler,
-        )
-
-        return any(
-            (event_data := AnthropicPassthroughLoggingHandler._extract_sse_data(event)) is not None  # pyright: ignore[reportPrivateUsage]  # same parser the assembler uses; a private import beats forking SSE parsing
-            and event_data.get("type") == "message_start"
-            for event in AnthropicPassthroughLoggingHandler._split_sse_chunk_into_events(sse_stream)  # pyright: ignore[reportPrivateUsage]  # same parser the assembler uses
-        )
-
-    @staticmethod
-    def _assemble_anthropic_stream(all_chunks: Sequence[Any]) -> ModelResponse | None:
-        from litellm.proxy.pass_through_endpoints.llm_provider_handlers.anthropic_passthrough_logging_handler import (
-            AnthropicPassthroughLoggingHandler,
-        )
-
-        sse_stream: Final = ToolPermissionGuardrail._joined_sse_stream(all_chunks)
-        if sse_stream is None or not ToolPermissionGuardrail._has_anthropic_message_start(sse_stream):
-            return None
-        try:
-            assembled = AnthropicPassthroughLoggingHandler._build_complete_streaming_response(  # pyright: ignore[reportPrivateUsage]  # the only SSE-to-ModelResponse assembler; reimplementing it here would fork the parser
-                all_chunks=(sse_stream,),
-                litellm_logging_obj=None,  # pyright: ignore[reportArgumentType]  # only forwarded to stream_chunk_builder, which accepts None
-                model="",
-            )
-        except (AttributeError, TypeError, ValueError, json.JSONDecodeError):
-            return None
-        return assembled if isinstance(assembled, ModelResponse) else None
-
-    @staticmethod
-    def _rewritten_anthropic_sse_chunks(assembled: ModelResponse) -> tuple[bytes, ...]:
-        from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
-            LiteLLMAnthropicMessagesAdapter,
-        )
-        from litellm.llms.anthropic.experimental_pass_through.messages.fake_stream_iterator import (
-            FakeAnthropicMessagesStreamIterator,
-        )
-
-        anthropic_response: Final = LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
-            response=assembled
-        )
-        return tuple(FakeAnthropicMessagesStreamIterator(response=anthropic_response).chunks)

litellm/proxy/health_endpoints/_health_endpoints.py

@@ -50,6 +50,7 @@ from litellm.router_utils.clientside_credential_handler import (
     _ADMIN_CONFIG_FIELDS_TO_CLEAR_ON_BASE_OVERRIDE,  # pyright: ignore[reportPrivateUsage]  # one canonical list, shared with the router path
     clientside_credential_keys,
 )
+from litellm.secret_managers.main import get_secret_bool
 
 #### Health ENDPOINTS ####
 
@@ -1447,6 +1448,31 @@ def callback_name(callback):
             return str(callback)
 
 
+DISABLE_NO_REDIS_WARNING_ENV_VAR: Final = "LITELLM_DISABLE_NO_REDIS_WARNING"
+
+
+def _show_no_redis_warning() -> bool:
+    """
+    Whether the UI should warn that no Redis is configured.
+
+    Redis is what makes rate limits, budgets, router state, and cache
+    invalidation consistent across workers, so a proxy running without it is
+    only safe as a single worker. Both places a Redis can land count: the
+    coordination cache (from a Redis response cache, general_settings.
+    coordination_redis, or the REDIS_* env fallback) and the router's own
+    Redis (router_settings.redis_host), which backs cooldowns and usage-based
+    routing on its own. Operators who know they run one worker can silence the
+    warning with LITELLM_DISABLE_NO_REDIS_WARNING=true.
+    """
+    from litellm.proxy.proxy_server import llm_router, redis_usage_cache
+
+    if redis_usage_cache is not None:
+        return False
+    if llm_router is not None and llm_router.cache.redis_cache is not None:
+        return False
+    return get_secret_bool(DISABLE_NO_REDIS_WARNING_ENV_VAR, False) is not True
+
+
 async def _get_health_readiness_details(
     response: Response | None = None,
 ) -> dict[str, Any]:
@@ -1487,6 +1513,7 @@ async def _get_health_readiness_details(
         # check log level
         log_level_name: Final = logging.getLevelName(verbose_logger.getEffectiveLevel())
         is_detailed_debug: Final = verbose_logger.isEnabledFor(logging.DEBUG)
+        show_no_redis_warning: Final = _show_no_redis_warning()
 
         # check DB
         if prisma_client is not None:  # if db passed in, check if it's connected
@@ -1506,6 +1533,7 @@ async def _get_health_readiness_details(
                 "use_aiohttp_transport": AsyncHTTPHandler._should_use_aiohttp_transport(),
                 "log_level": log_level_name,
                 "is_detailed_debug": is_detailed_debug,
+                "show_no_redis_warning": show_no_redis_warning,
             }
         else:
             return {
@@ -1517,6 +1545,7 @@ async def _get_health_readiness_details(
                 "use_aiohttp_transport": AsyncHTTPHandler._should_use_aiohttp_transport(),
                 "log_level": log_level_name,
                 "is_detailed_debug": is_detailed_debug,
+                "show_no_redis_warning": show_no_redis_warning,
             }
     except Exception as e:
         raise HTTPException(status_code=503, detail=f"Service Unhealthy ({e})")

litellm/proxy/hooks/proxy_track_cost_callback.py

@@ -39,11 +39,10 @@ _UNATTRIBUTED_TRACKABLE_CALL_TYPES: Final[frozenset[str]] = frozenset(
         CallTypes.pass_through.value,
         CallTypes.llm_passthrough_route.value,
         CallTypes.allm_passthrough_route.value,
-        # CheckBatchCost's synthetic logging_obj for a completed managed batch only ever
-        # carries user_api_key_user_id (from LiteLLM_ManagedObjectTable.created_by) and
-        # user_api_key_team_id (from .team_id) -- both are None for batches created with
-        # the master key or a team-less key, since the table never stores the raw key
-        # hash. The batch already incurred real provider cost, so track it regardless.
+        # CheckBatchCost's synthetic logging_obj for a completed managed batch carries
+        # whatever LiteLLM_ManagedObjectTable stored at create time, and all of it is
+        # None for a batch created before those columns were persisted, or by the master
+        # key. The batch already incurred real provider cost, so track it regardless.
         CallTypes.aretrieve_batch.value,
     }
 )

litellm/proxy/litellm_pre_call_utils.py

@@ -16,6 +16,7 @@ import litellm
 from litellm._logging import verbose_logger, verbose_proxy_logger
 from litellm._service_logger import ServiceLogging
 from litellm.constants import (
+    CONSUMED_REQUEST_TAGS_METADATA_KEY,
     INTERNAL_CALL_ORIGIN_METADATA_KEY,
     LITELLM_PROXY_MASTER_KEY_ALIAS,
     PRE_CALL_EXECUTED_GUARDRAILS_KEY,
@@ -261,6 +262,7 @@ _UNTRUSTED_METADATA_CONTROL_FIELDS: Final = (
     "policy_sources",
     "routing_decision",
     SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
+    CONSUMED_REQUEST_TAGS_METADATA_KEY,
     INTERNAL_CALL_ORIGIN_METADATA_KEY,
     "standard_logging_object",
     "proxy_server_request",

litellm/proxy/management_endpoints/cost_tracking_settings.py

@@ -10,6 +10,8 @@ PATCH /config/cost_margin_config - Update cost margin configuration
 POST /cost/estimate - Estimate cost for a given model and token counts
 """
 
+from collections.abc import Mapping
+from dataclasses import dataclass
 from typing import Final
 
 from fastapi import APIRouter, Depends, HTTPException
@@ -24,29 +26,65 @@ from litellm.proxy._types import (
     UserAPIKeyAuth,
 )
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
-from litellm.types.utils import LlmProvidersSet
+from litellm.types.utils import CostPerToken, LlmProvidersSet, ModelInfo
 
 router: Final = APIRouter()
 
 
-def _resolve_model_for_cost_lookup(model: str) -> tuple[str, str | None]:
+@dataclass(frozen=True, slots=True)
+class ResolvedCostModel:
+    model: str
+    provider: str | None
+    custom_cost_per_token: CostPerToken | None
+
+
+def _configured_price(key: str, sources: tuple[Mapping[str, object], ...]) -> float | None:
+    values: Final = (source.get(key) for source in sources)
+    numeric: Final = (float(value) for value in values if isinstance(value, (int, float)))
+    return next(numeric, None)
+
+
+def _extract_custom_pricing(
+    litellm_params: Mapping[str, object], model_info: Mapping[str, object]
+) -> CostPerToken | None:
+    """
+    Pull per-token pricing configured on a deployment so on-prem / self-hosted
+    models (absent from the public cost map) still estimate a real cost.
+    Pricing may live on ``litellm_params`` or ``model_info``; ``litellm_params``
+    wins, matching the router's cost-map registration precedence.
+    """
+    sources: Final = (litellm_params, model_info)
+    input_price: Final = _configured_price("input_cost_per_token", sources)
+    output_price: Final = _configured_price("output_cost_per_token", sources)
+
+    if input_price is None and output_price is None:
+        return None
+
+    return CostPerToken(
+        input_cost_per_token=input_price or 0.0,
+        output_cost_per_token=output_price or 0.0,
+    )
+
+
+def _lookup_model_info(model: str) -> ModelInfo | None:
+    try:
+        return litellm.get_model_info(model=model)
+    except Exception:
+        return None
+
+
+def _resolve_model_for_cost_lookup(model: str) -> ResolvedCostModel:
     """
     Resolve a model name (which may be a router alias/model_group) to the
-    underlying litellm model name for cost lookup.
+    underlying litellm model name, provider, and any deployment-configured
+    pricing used for cost lookup.
 
     Args:
         model: The model name from the request (could be a router alias like 'e-model-router'
                or an actual model name like 'azure_ai/gpt-4')
-
-    Returns:
-        Tuple of (resolved_model_name, custom_llm_provider)
-        - resolved_model_name: The actual model name to use for cost lookup
-        - custom_llm_provider: The provider if resolved from router, None otherwise
     """
     from litellm.proxy.proxy_server import llm_router
 
-    custom_llm_provider: str | None = None
-
     # Try to resolve from router if available
     if llm_router is not None:
         try:
@@ -57,31 +95,25 @@ def _resolve_model_for_cost_lookup(model: str) -> tuple[str, str | None]:
                 first_deployment: Final = deployments[0]
                 litellm_params: Final = first_deployment.get("litellm_params", {})
                 model_info: Final = first_deployment.get("model_info", {})
+                custom_llm_provider: Final = litellm_params.get("custom_llm_provider")
+                provider: Final = str(custom_llm_provider) if custom_llm_provider is not None else None
+                custom_cost_per_token: Final = _extract_custom_pricing(litellm_params, model_info)
 
                 # Check base_model first (needed for Azure custom deployment names)
                 base_model: Final = model_info.get("base_model") or litellm_params.get("base_model")
                 if base_model:
                     verbose_proxy_logger.debug("Resolved model '%s' to base_model '%s' from router", model, base_model)
-                    custom_llm_provider = litellm_params.get("custom_llm_provider")
-                    return (
-                        str(base_model),
-                        (str(custom_llm_provider) if custom_llm_provider is not None else None),
-                    )
+                    return ResolvedCostModel(str(base_model), provider, custom_cost_per_token)
 
                 resolved_model: Final = litellm_params.get("model")
-
                 if resolved_model:
                     verbose_proxy_logger.debug("Resolved model '%s' to '%s' from router", model, resolved_model)
-                    custom_llm_provider = litellm_params.get("custom_llm_provider")
-                    return (
-                        str(resolved_model),
-                        (str(custom_llm_provider) if custom_llm_provider is not None else None),
-                    )
+                    return ResolvedCostModel(str(resolved_model), provider, custom_cost_per_token)
         except Exception as e:
             verbose_proxy_logger.debug("Could not resolve model '%s' from router: %s", model, e)
 
     # Return original model if not resolved
-    return model, custom_llm_provider
+    return ResolvedCostModel(model, None, None)
 
 
 def _calculate_period_costs(num_requests, cost_per_request, input_cost, output_cost, margin_cost):
@@ -450,7 +482,9 @@ async def estimate_cost(
     from litellm.types.utils import ModelResponse, Usage
 
     # Resolve model name (handles router aliases like 'e-model-router' -> 'azure_ai/gpt-4')
-    resolved_model, resolved_provider = _resolve_model_for_cost_lookup(request.model)
+    resolved: Final = _resolve_model_for_cost_lookup(request.model)
+    resolved_model: Final = resolved.model
+    resolved_provider: Final = resolved.provider
 
     verbose_proxy_logger.debug("Cost estimate: request.model='%s' resolved to '%s'", request.model, resolved_model)
 
@@ -480,6 +514,8 @@ async def estimate_cost(
         cost_per_request: Final = completion_cost(
             completion_response=mock_response,
             model=resolved_model,
+            custom_llm_provider=resolved_provider,
+            custom_cost_per_token=resolved.custom_cost_per_token,
             litellm_logging_obj=litellm_logging_obj,
         )
     except Exception as e:
@@ -497,20 +533,22 @@ async def estimate_cost(
     output_cost: Final = cost_breakdown.get("output_cost", 0.0) if cost_breakdown else 0.0
     margin_cost: Final = cost_breakdown.get("margin_total_amount", 0.0) if cost_breakdown else 0.0
 
-    # Get model info for per-token pricing display
-    try:
-        model_info: Final = litellm.get_model_info(model=resolved_model)
-        input_cost_per_token = model_info.get("input_cost_per_token")
-        output_cost_per_token = model_info.get("output_cost_per_token")
-        custom_llm_provider = model_info.get("litellm_provider")
-    except Exception:
-        input_cost_per_token = None
-        output_cost_per_token = None
-        custom_llm_provider = None
+    model_info: Final = _lookup_model_info(resolved_model)
+    mapped_input_price: Final = model_info.get("input_cost_per_token") if model_info is not None else None
+    mapped_output_price: Final = model_info.get("output_cost_per_token") if model_info is not None else None
+    mapped_provider: Final = model_info.get("litellm_provider") if model_info is not None else None
 
-    # Use provider from router resolution if not found in model_info
-    if custom_llm_provider is None and resolved_provider is not None:
-        custom_llm_provider = resolved_provider
+    input_cost_per_token: Final = (
+        resolved.custom_cost_per_token["input_cost_per_token"]
+        if resolved.custom_cost_per_token is not None
+        else mapped_input_price
+    )
+    output_cost_per_token: Final = (
+        resolved.custom_cost_per_token["output_cost_per_token"]
+        if resolved.custom_cost_per_token is not None
+        else mapped_output_price
+    )
+    custom_llm_provider: Final = mapped_provider if mapped_provider is not None else resolved_provider
 
     # Calculate daily and monthly costs
     (

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -35,6 +35,7 @@ from litellm.proxy._types import (
     PrismaCompatibleUpdateDBModel,
     ProxyErrorTypes,
     ProxyException,
+    ReconcileOutcome,
     TeamModelAddRequest,
     TeamModelDeleteRequest,
     UserAPIKeyAuth,
@@ -534,7 +535,7 @@ async def patch_model(
 
         # Clear cache and reload models (uses config setting or defaults to preserving config models for DB updates)
         live_before_reload: Final = live_model_ids_snapshot()
-        still_desired_ids: Final = await clear_cache()
+        reload_outcome: Final = await clear_cache()
 
         ## CREATE AUDIT LOG ##
         asyncio.create_task(
@@ -554,7 +555,8 @@ async def patch_model(
             before=live_before_reload,
             written_models=[(model_id, getattr(updated_model, "model_info", None))],
             action="update",
-            still_desired=still_desired_ids,
+            still_desired=reload_outcome.still_desired,
+            live_after=reload_outcome.live_after,
         )
 
         return updated_model
@@ -640,7 +642,7 @@ async def _set_model_blocked_status(
         )
 
         live_before_reload: Final = live_model_ids_snapshot()
-        still_desired_ids: Final = await clear_cache()
+        reload_outcome: Final = await clear_cache()
 
         asyncio.create_task(
             create_object_audit_log(
@@ -661,7 +663,8 @@ async def _set_model_blocked_status(
             before=live_before_reload,
             written_models=[(data.model_id, getattr(updated_model, "model_info", None))],
             action=action,
-            still_desired=still_desired_ids,
+            still_desired=reload_outcome.still_desired,
+            live_after=reload_outcome.live_after,
         )
 
         return updated_model
@@ -1033,9 +1036,15 @@ async def delete_team_models(
     if deleted_model_ids:
         await publish_config_change(redis_cache=coordination_redis_cache(), object_type="litellm_proxymodeltable")
 
+    # Under MODEL_RECONCILE_LOCK, for the same reason as delete_model: the rows are
+    # gone, but a reconcile holding a pre-delete snapshot would upsert these ids back
+    # onto this pod. The lock orders the eviction after any in-flight reconcile.
     if llm_router is not None:
-        for model_id in deleted_model_ids:
-            llm_router.delete_deployment(id=model_id)
+        from litellm.proxy.proxy_server import MODEL_RECONCILE_LOCK
+
+        async with MODEL_RECONCILE_LOCK:
+            for model_id in deleted_model_ids:
+                llm_router.delete_deployment(id=model_id)
 
     return deleted_model_ids
 
@@ -1355,6 +1364,7 @@ async def delete_model(
         """
 
         from litellm.proxy.proxy_server import (
+            MODEL_RECONCILE_LOCK,
             llm_router,
             premium_user,
             prisma_client,
@@ -1403,8 +1413,15 @@ async def delete_model(
                 )
 
             ## DELETE FROM ROUTER ##
+            # Under MODEL_RECONCILE_LOCK. The db row is already gone, but a reconcile
+            # that snapshotted the db BEFORE that delete still lists this id as desired,
+            # and its _add_deployment upserts the deployment straight back -- leaving
+            # this pod serving a model the database no longer has, until the next
+            # reconcile. Taking the lock orders this eviction after any such in-flight
+            # reconcile's re-add, so the eviction is the last word.
             if llm_router is not None:
-                llm_router.delete_deployment(id=model_info.id)
+                async with MODEL_RECONCILE_LOCK:
+                    llm_router.delete_deployment(id=model_info.id)
 
             # Runs after the row delete so the sibling check sees post-delete state.
             if model_params.model_info.team_id is not None:
@@ -1579,7 +1596,7 @@ async def add_new_model(
             """
 
             live_before_reload: Final = live_model_ids_snapshot()
-            still_desired_ids: frozenset[str] | None = None
+            reload_outcome: ReconcileOutcome = ReconcileOutcome(still_desired=None, live_after=None)
             try:
                 _original_litellm_model_name: Final = model_params.model_name
                 if model_params.model_info.team_id is None:
@@ -1594,7 +1611,7 @@ async def add_new_model(
                         user_api_key_dict=user_api_key_dict,
                         prisma_client=prisma_client,
                     )
-                still_desired_ids = await proxy_config.add_deployment(
+                reload_outcome = await proxy_config.add_deployment(
                     prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj
                 )
                 # don't let failed slack alert block the /model/new response
@@ -1641,7 +1658,8 @@ async def add_new_model(
             before=live_before_reload,
             written_models=[(model_response.model_id, getattr(model_response, "model_info", None))],
             action="create",
-            still_desired=still_desired_ids,
+            still_desired=reload_outcome.still_desired,
+            live_after=reload_outcome.live_after,
         )
 
         return model_response
@@ -1768,7 +1786,7 @@ async def update_model(
 
             # Clear cache and reload models (uses config setting or defaults to preserving config models for DB updates)
             live_before_reload: Final = live_model_ids_snapshot()
-            still_desired_ids: Final = await clear_cache()
+            reload_outcome: Final = await clear_cache()
             ## CREATE AUDIT LOG ##
             asyncio.create_task(
                 create_object_audit_log(
@@ -1795,7 +1813,8 @@ async def update_model(
                 before=live_before_reload,
                 written_models=[(_model_id, getattr(model_response, "model_info", None))],
                 action="update",
-                still_desired=still_desired_ids,
+                still_desired=reload_outcome.still_desired,
+                live_after=reload_outcome.live_after,
             )
 
             return model_response
@@ -2100,6 +2119,7 @@ def reload_serving_verdict(
     written_models: Sequence[tuple[str, object]],
     written_must_serve: bool,
     still_desired: frozenset[str] | None = None,
+    live_after: frozenset[str] | None = None,
 ) -> tuple[tuple[str, ...], tuple[str, ...]]:
     """Judge a write-triggered reload by diffing the router's serving state instead of
     trusting any layer of the reload stack to report its own failure.
@@ -2121,9 +2141,16 @@ def reload_serving_verdict(
     yet polled, so the reload dropping it is the reconcile working rather than damage.
     Without it (no reconcile ran) every drop is reported, which is the safe direction.
 
+    ``live_after`` is the router's serving state captured by the reload itself, while it
+    still held MODEL_RECONCILE_LOCK. Pass it whenever the caller has it: re-reading the
+    router here instead means sampling it after the lock was released, where the NEXT
+    reconcile's leading wipe (clear_cache un-serves every db model before reloading
+    them) shows up as this reload having dropped them. Falling back to a fresh read is
+    only correct when no reconcile ran and there is nothing to be concurrent with.
+
     Returns (written ids violating their obligation, collateral ids no longer served).
     """
-    now: Final = live_model_ids_snapshot()
+    now: Final = live_model_ids_snapshot() if live_after is None else live_after
     written_ids: Final = frozenset(model_id for model_id, _ in written_models)
     if written_must_serve:
         missing = tuple(
@@ -2143,16 +2170,23 @@ def raise_if_reload_degraded_serving(
     written_models: Sequence[tuple[str, object]],
     action: str,
     still_desired: frozenset[str] | None = None,
+    live_after: frozenset[str] | None = None,
 ) -> None:
     """The caller-visible error this pod's model-write endpoints owe their caller when
     the model they wrote is not being served after the reload they triggered. The DB
     write is durable either way and every other pod reloads on its own interval; this
-    speaks only for the handling pod."""
+    speaks only for the handling pod.
+
+    Callers hold a ReconcileOutcome from the reload; pass BOTH of its fields. Supplying
+    still_desired without live_after mixes a snapshot taken under the reconcile lock
+    with one taken after it was released, which is what makes a concurrent model write
+    look like collateral damage."""
     missing, collateral = reload_serving_verdict(
         before=before,
         written_models=written_models,
         written_must_serve=True,
         still_desired=still_desired,
+        live_after=live_after,
     )
     if not missing and not collateral:
         return
@@ -2179,14 +2213,20 @@ def raise_if_reload_degraded_serving(
     )
 
 
-async def clear_cache() -> frozenset[str] | None:
+async def clear_cache() -> ReconcileOutcome:
     """
     Clear router caches and reload models.
 
-    Returns the db + config id set the reload reconciled against, or None when no
-    reload ran, so callers can pass it to raise_if_reload_degraded_serving.
+    Returns what the reload saw (see ReconcileOutcome) so callers can pass it to
+    raise_if_reload_degraded_serving.
+
+    Runs under MODEL_RECONCILE_LOCK for its whole extent, not just the reload at the
+    end, so the auto-router reset and the reload that rebuilds those routers are atomic
+    to any other reconcile. The inner call is _add_deployment_locked because
+    add_deployment would re-acquire the same non-reentrant lock and deadlock.
     """
     from litellm.proxy.proxy_server import (
+        MODEL_RECONCILE_LOCK,
         llm_router,
         prisma_client,
         proxy_config,
@@ -2196,61 +2236,88 @@ async def clear_cache() -> frozenset[str] | None:
 
     if llm_router is None or prisma_client is None:
         verbose_proxy_logger.debug("llm_router or prisma_client is None, skipping cache clear")
-        return None
+        return ReconcileOutcome(still_desired=None, live_after=None)
 
-    try:
-        # Only clear DB models, preserve config models
-        verbose_proxy_logger.debug("Clearing only DB models, preserving config models")
+    async with MODEL_RECONCILE_LOCK:
+        try:
+            # Only clear DB models, preserve config models
+            verbose_proxy_logger.debug("Clearing only DB models, preserving config models")
 
-        # Get current models and filter out DB models
-        current_models: Final = llm_router.model_list.copy()
-        config_models: Final = []
-        db_model_ids: Final = []
+            # Get current models and filter out DB models
+            current_models: Final = llm_router.model_list.copy()
+            config_models: Final = []
+            db_model_ids: Final = []
 
-        for model in current_models:
-            model_info = model.get("model_info", {})
-            if model_info.get("db_model", False):
-                # This is a DB model, mark for deletion
-                db_model_ids.append(model_info.get("id"))
-            else:
-                # This is a config model, preserve it
-                config_models.append(model)
+            db_router_names: Final = set()
 
-        # Clear only DB models
-        for model_id in db_model_ids:
-            llm_router.delete_deployment(id=model_id)
+            for model in current_models:
+                model_info = model.get("model_info", {})
+                if model_info.get("db_model", False):
+                    db_model_ids.append(model_info.get("id"))
+                    # Auto-router deployments (and only those) are wiped here, in the
+                    # same pass, so the reload rebuilds them -- see the comment below.
+                    model_name = model.get("model_name")
+                    if model_name is not None and str(model.get("litellm_params", {}).get("model", "")).startswith(
+                        "auto_router/"
+                    ):
+                        db_router_names.add(model_name)
+                        router_model_id = model_info.get("id")
+                        if router_model_id is not None:
+                            llm_router.delete_deployment(id=router_model_id)
+                else:
+                    # This is a config model, preserved by the reconcile below
+                    config_models.append(model)
 
-        # Clear only DB-backed auto-router-family entries, keyed by model_name, so the
-        # reload below rebuilds them fresh. A blanket .clear() would also drop config-defined
-        # routers, which are never re-added below (add_deployment only reloads DB models),
-        # leaving them permanently unroutable until a full proxy restart for every tenant.
-        # Restrict to deployments whose model is actually an auto_router/* so a config
-        # router that merely shares a model_name with a regular DB model isn't evicted. The
-        # auto_router/ prefix also covers quality_router/ and adaptive_router/, so pop the
-        # name from every router registry (no-op where absent); missing quality/adaptive
-        # entries would otherwise make init raise "already exists" on reload and abort it.
-        db_router_names: Final = {
-            model.get("model_name")
-            for model in current_models
-            if model.get("model_name") is not None
-            and model.get("model_info", {}).get("db_model", False)
-            and str(model.get("litellm_params", {}).get("model", "")).startswith("auto_router/")
-        }
-        for model_name in db_router_names:
-            llm_router.auto_routers.pop(model_name, None)
-            llm_router.complexity_routers.pop(model_name, None)
-            llm_router.adaptive_routers.pop(model_name, None)
-            llm_router.quality_routers.pop(model_name, None)
+            # ORDINARY db deployments are deliberately NOT wiped. This used to
+            # delete_deployment() every db model before the reload put them back, which
+            # left the router serving ZERO db models for the whole width of the reload
+            # -- a real data-plane hole that every inference request landing in it fell
+            # into. It was also redundant for them: the reload's _delete_deployment
+            # evicts exactly the ids the db no longer lists, and upsert_deployment
+            # pops-and-re-adds a deployment whose params changed while no-opping one
+            # that did not, so the reconcile converges on its own. Every mutation is
+            # visible to that comparison -- `blocked` and (for premium) `updated_at`
+            # are written into model_info.
+            #
+            # AUTO-ROUTER db deployments are the exception and ARE wiped -- in the
+            # classification pass above, together with the strategy entries popped
+            # just below. Their strategy registries are keyed
+            # by model_name, which no deployment-id reconcile touches, so they have to
+            # be popped and rebuilt here. But the rebuild only happens on the ADD path:
+            # Router.upsert_deployment returns early when a deployment is unchanged and
+            # never reaches add_deployment -> _add_deployment ->
+            # init_auto_router_deployment, which is what repopulates the registries.
+            # Popping without deleting would therefore strip every db-backed auto,
+            # complexity, adaptive and quality router on this pod and never put it back,
+            # so ANY unrelated model write would leave them unroutable until a restart.
+            # Deleting the deployment forces upsert down the add path, which rebuilds
+            # both the deployment and its strategy entry.
+            #
+            # That pass restricts the wipe to deployments whose model is actually an
+            # auto_router/* so a config router that merely shares a model_name with a
+            # regular db model isn't evicted -- config routers are never re-added by the
+            # reload (it only reloads db models) and would be permanently unroutable.
+            # The auto_router/ prefix also covers quality_router/ and adaptive_router/,
+            # so pop the name from every registry (no-op where absent); a missing
+            # quality/adaptive entry would otherwise make init raise "already exists"
+            # on reload and abort it.
+            for model_name in db_router_names:
+                llm_router.auto_routers.pop(model_name, None)
+                llm_router.complexity_routers.pop(model_name, None)
+                llm_router.adaptive_routers.pop(model_name, None)
+                llm_router.quality_routers.pop(model_name, None)
 
-        # Reload only DB models
-        still_desired_ids: Final = await proxy_config.add_deployment(
-            prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj
-        )
+            # Reload only DB models. _add_deployment_locked, not add_deployment: this
+            # coroutine already holds MODEL_RECONCILE_LOCK and asyncio.Lock is not
+            # reentrant, so the public wrapper would deadlock against itself.
+            outcome: Final = await proxy_config._add_deployment_locked(
+                prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj
+            )
 
-        verbose_proxy_logger.debug(
-            "Cleared %s DB models, preserved %s config models", len(db_model_ids), len(config_models)
-        )
-        return still_desired_ids
-    except Exception as e:
-        verbose_proxy_logger.exception("Failed to clear cache and reload models. Due to error - %s", e)
-        return None
+            verbose_proxy_logger.debug(
+                "Reconciled %s DB models, preserved %s config models", len(db_model_ids), len(config_models)
+            )
+            return outcome
+        except Exception as e:
+            verbose_proxy_logger.exception("Failed to clear cache and reload models. Due to error - %s", e)
+            return ReconcileOutcome(still_desired=None, live_after=None)

litellm/proxy/pass_through_endpoints/llm_provider_handlers/anthropic_passthrough_logging_handler.py

@@ -1,5 +1,6 @@
+import asyncio
 import json
-from collections.abc import Sequence
+from collections.abc import Mapping, Sequence
 from datetime import datetime
 from typing import TYPE_CHECKING, Any, Final, cast
 
@@ -7,6 +8,7 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_proxy_logger
+from litellm.constants import ANTHROPIC_BATCHES_ROUTE
 from litellm.litellm_core_utils.core_helpers import map_finish_reason
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.litellm_logging import use_custom_pricing_for_model
@@ -20,6 +22,12 @@ from litellm.llms.anthropic.chat.handler import (
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 from litellm.proxy._types import PassThroughEndpointLoggingTypedDict
 from litellm.proxy.auth.auth_utils import get_end_user_id_from_request_body
+from litellm.proxy.pass_through_endpoints.llm_provider_handlers.batch_attribution import (
+    is_collection_route,
+    log_batch_registration_result,
+    optional_str,
+    request_tags_from_metadata,
+)
 from litellm.types.passthrough_endpoints.pass_through_endpoints import (
     PassthroughStandardLoggingPayload,
 )
@@ -74,6 +82,9 @@ class AnthropicPassthroughLoggingHandler:
             )
 
         model: Final = response_body.get("model", "")
+        speed: Final = AnthropicPassthroughLoggingHandler._cost_relevant_speed(
+            request_body or kwargs.get("request_body")
+        )
         anthropic_config: Final = get_anthropic_config(url_route)
         litellm_model_response: Final[ModelResponse] = anthropic_config().transform_response(
             raw_response=httpx_response,
@@ -81,7 +92,7 @@ class AnthropicPassthroughLoggingHandler:
             model=model,
             messages=[],
             logging_obj=logging_obj,
-            optional_params={},
+            optional_params={"speed": speed} if speed else {},
             api_key="",
             request_data={},
             encoding=litellm.encoding,
@@ -103,6 +114,15 @@ class AnthropicPassthroughLoggingHandler:
             "kwargs": kwargs,
         }
 
+    @staticmethod
+    def _cost_relevant_speed(request_body: Mapping[str, object] | None) -> str | None:
+        """
+        Anthropic's ``speed=fast`` multiplies non-cache token cost, and only the request
+        carries it, so it has to reach the usage-building paths for spend to be right.
+        """
+        speed: Final = (request_body or {}).get("speed")
+        return speed if isinstance(speed, str) else None
+
     @staticmethod
     def _get_user_from_metadata(
         passthrough_logging_payload: PassthroughStandardLoggingPayload,
@@ -316,6 +336,7 @@ class AnthropicPassthroughLoggingHandler:
         - Logs in litellm callbacks
         """
 
+        speed: Final = AnthropicPassthroughLoggingHandler._cost_relevant_speed(request_body)
         model = request_body.get("model", "")
         # Check if it's available in the logging object
         if (
@@ -335,6 +356,7 @@ class AnthropicPassthroughLoggingHandler:
                 all_chunks=all_chunks,
                 litellm_logging_obj=litellm_logging_obj,
                 model=model,
+                speed=speed,
             )
         except Exception as e:
             # stream_chunk_builder re-raises assembly failures (as litellm.APIError)
@@ -356,6 +378,7 @@ class AnthropicPassthroughLoggingHandler:
                 complete_streaming_response = AnthropicPassthroughLoggingHandler._build_usage_only_response_from_chunks(
                     all_chunks=all_chunks,
                     model=model,
+                    speed=speed,
                 )
             except Exception as e:
                 verbose_proxy_logger.warning(
@@ -420,6 +443,7 @@ class AnthropicPassthroughLoggingHandler:
         all_chunks: Sequence[str | bytes],
         litellm_logging_obj: LiteLLMLoggingObj,
         model: str,
+        speed: str | None = None,
     ) -> ModelResponse | TextCompletionResponse | None:
         """
         Builds complete response from raw Anthropic chunks.
@@ -444,11 +468,13 @@ class AnthropicPassthroughLoggingHandler:
                 all_chunks=collapsed,
                 litellm_logging_obj=litellm_logging_obj,
                 model=model,
+                speed=speed,
             )
         return AnthropicPassthroughLoggingHandler._build_complete_streaming_response_legacy(
             all_chunks=all_chunks,
             litellm_logging_obj=litellm_logging_obj,
             model=model,
+            speed=speed,
         )
 
     # Anthropic SSE block/delta types that the fast path is NOT allowed to
@@ -576,6 +602,7 @@ class AnthropicPassthroughLoggingHandler:
         all_chunks: Sequence[str | bytes],
         litellm_logging_obj: LiteLLMLoggingObj,
         model: str,
+        speed: str | None = None,
     ) -> ModelResponse | TextCompletionResponse | None:
         """
         Original reconstruction: convert every SSE event to a generic chunk
@@ -591,6 +618,7 @@ class AnthropicPassthroughLoggingHandler:
         anthropic_model_response_iterator: Final = AnthropicModelResponseIterator(
             streaming_response=None,
             sync_stream=False,
+            speed=speed,
         )
         all_openai_chunks: Final = []
 
@@ -650,6 +678,7 @@ class AnthropicPassthroughLoggingHandler:
     def _build_usage_only_response_from_chunks(
         all_chunks: Sequence[str | bytes],
         model: str,
+        speed: str | None = None,
     ) -> ModelResponse | None:
         """
         Build a usage-bearing ModelResponse from Anthropic SSE token-usage events, for
@@ -743,7 +772,9 @@ class AnthropicPassthroughLoggingHandler:
             usage_object["server_tool_use"] = _server_tool_use
         if inference_geo is not None:
             usage_object["inference_geo"] = inference_geo
-        usage_obj: Final = AnthropicConfig().calculate_usage(usage_object=usage_object, reasoning_content=None)
+        usage_obj: Final = AnthropicConfig().calculate_usage(
+            usage_object=usage_object, reasoning_content=None, speed=speed
+        )
         return ModelResponse(
             model=resolved_model,
             choices=[
@@ -833,13 +864,14 @@ class AnthropicPassthroughLoggingHandler:
 
                 # Store the managed object for cost tracking
                 # This will be picked up by check_batch_cost polling mechanism
-                AnthropicPassthroughLoggingHandler._store_batch_managed_object(
-                    unified_object_id=unified_object_id,
-                    batch_object=litellm_batch_response,
-                    model_object_id=batch_id,
-                    logging_obj=logging_obj,
-                    **kwargs,
-                )
+                if is_collection_route(url_route, ANTHROPIC_BATCHES_ROUTE):
+                    AnthropicPassthroughLoggingHandler._store_batch_managed_object(
+                        unified_object_id=unified_object_id,
+                        batch_object=litellm_batch_response,
+                        model_object_id=batch_id,
+                        logging_obj=logging_obj,
+                        **kwargs,
+                    )
 
                 # Create a batch job response for logging
                 litellm_model_response = ModelResponse()
@@ -964,8 +996,12 @@ class AnthropicPassthroughLoggingHandler:
         **kwargs,
     ) -> None:
         """
-        Store batch managed object for cost tracking.
+        Register a newly created batch for cost tracking.
         This will be picked up by the check_batch_cost polling mechanism.
+
+        Only the create reaches here, so the row records the creating key and its tags.
+        An id-scoped route cannot rebuild the unified object id anyway: the model comes
+        from the create's request body, which a retrieve does not have.
         """
         try:
             # Get the managed files hook from the logging object
@@ -981,7 +1017,7 @@ class AnthropicPassthroughLoggingHandler:
 
                 user_api_key_dict: Final = UserAPIKeyAuth(
                     user_id=_request_metadata.get("user_api_key_user_id", "default-user"),
-                    api_key="",
+                    api_key=optional_str(_request_metadata.get("user_api_key")),
                     team_id=_request_metadata.get("user_api_key_team_id"),
                     team_alias=None,
                     user_role=LitellmUserRoles.CUSTOMER,  # Use proper enum value
@@ -1003,9 +1039,7 @@ class AnthropicPassthroughLoggingHandler:
                 )
 
                 # Store the unified object for batch cost tracking
-                import asyncio
-
-                asyncio.create_task(
+                task: Final = asyncio.create_task(
                     managed_files_hook.store_unified_object_id(
                         unified_object_id=unified_object_id,
                         file_object=batch_object,
@@ -1013,13 +1047,14 @@ class AnthropicPassthroughLoggingHandler:
                         model_object_id=model_object_id,
                         file_purpose="batch",
                         user_api_key_dict=user_api_key_dict,
+                        request_tags=request_tags_from_metadata(_request_metadata),
+                        persist_attribution=True,
                     )
                 )
-
-                verbose_proxy_logger.info(
-                    "Stored Anthropic batch managed object with unified_object_id=%s, batch_id=%s",
-                    unified_object_id,
-                    model_object_id,
+                task.add_done_callback(
+                    lambda finished: log_batch_registration_result(
+                        finished, "Anthropic", unified_object_id, model_object_id, is_batch_create=True
+                    )
                 )
             else:
                 verbose_proxy_logger.warning(

litellm/proxy/pass_through_endpoints/llm_provider_handlers/batch_attribution.py

@@ -0,0 +1,79 @@
+"""Spend attribution for batches created through a passthrough endpoint.
+
+The creating key and its tags are read off the passthrough request's metadata and
+persisted on the managed object row, because the batch cost lands hours later in a
+background poll that has no request to read them from.
+"""
+
+import asyncio
+from collections.abc import Mapping, Sequence
+from typing import Final
+
+from litellm._logging import verbose_proxy_logger
+from litellm.litellm_core_utils.safe_json_dumps import strip_null_bytes
+
+
+def optional_str(value: object) -> str | None:
+    return value if isinstance(value, str) else None
+
+
+def _sanitized_str_tuple(value: object) -> tuple[str, ...] | None:
+    if not isinstance(value, list):
+        return None
+    items: Final[Sequence[object]] = value
+    return tuple(strip_null_bytes(tag) for tag in items if isinstance(tag, str))
+
+
+def is_collection_route(url_route: str, collection_suffix: str) -> bool:
+    """Whether the route addresses the batch collection itself rather than one batch.
+    A POST to the collection is the create; every id-scoped route is a retrieve,
+    results or cancel.
+    """
+    return url_route.split("?")[0].rstrip("/").endswith(collection_suffix)
+
+
+def request_tags_from_metadata(request_metadata: Mapping[str, object]) -> tuple[str, ...] | None:
+    """Tags for the batch-cost spend row: the request's own tags when it sent any,
+    otherwise the key's tags, which auth exposes as user_api_key_auth_metadata (a
+    tagged key does not put its tags in the top-level metadata "tags" on the
+    passthrough path)
+    """
+    tags: Final = _sanitized_str_tuple(request_metadata.get("tags"))
+    if tags:
+        return tags
+    key_auth_metadata: Final = request_metadata.get("user_api_key_auth_metadata")
+    if isinstance(key_auth_metadata, dict):
+        return _sanitized_str_tuple(key_auth_metadata.get("tags"))
+    return None
+
+
+def log_batch_registration_result(
+    finished: asyncio.Task[None],
+    provider: str,
+    unified_object_id: str,
+    model_object_id: str,
+    is_batch_create: bool,
+) -> None:
+    """Report the outcome of the fire-and-forget managed object write. A create that
+    fails is not retried by a later poll, so its cost is never tracked at all.
+    """
+    error: Final = finished.exception() if not finished.cancelled() else None
+    if finished.cancelled() or error is not None:
+        consequence: Final = (
+            "its cost will not be tracked" if is_batch_create else "its status and output file may be stale"
+        )
+        verbose_proxy_logger.error(
+            "Failed to store %s batch managed object with unified_object_id=%s, batch_id=%s; %s: %s",
+            provider,
+            unified_object_id,
+            model_object_id,
+            consequence,
+            error,
+        )
+        return
+    verbose_proxy_logger.info(
+        "Stored %s batch managed object with unified_object_id=%s, batch_id=%s",
+        provider,
+        unified_object_id,
+        model_object_id,
+    )

litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_passthrough_logging_handler.py

@@ -464,7 +464,7 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
 
     def _build_complete_streaming_response(
         self,
-        all_chunks: list,
+        all_chunks: list[str],
         litellm_logging_obj: LiteLLMLoggingObj,
         model: str,
     ) -> ModelResponse | TextCompletionResponse | None:
@@ -536,13 +536,19 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
             # Extract model from request body
             model: Final = request_body.get("model", "gpt-4o")
 
+            is_responses: Final = OpenAIPassthroughLoggingHandler.is_openai_responses_route(url_route)
+
             # Build complete response from chunks using our streaming handler
             handler: Final = OpenAIPassthroughLoggingHandler()
             handler_instance: Final = handler
-            complete_response: Final = handler._build_complete_streaming_response(
-                all_chunks=all_chunks,
-                litellm_logging_obj=litellm_logging_obj,
-                model=model,
+            complete_response: Final = (
+                OpenAIResponsesAPIConfig.parse_terminal_response_from_stream_chunks(all_chunks=all_chunks)
+                if is_responses
+                else handler._build_complete_streaming_response(
+                    all_chunks=all_chunks,
+                    litellm_logging_obj=litellm_logging_obj,
+                    model=model,
+                )
             )
 
             if complete_response is None:
@@ -554,10 +560,19 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
 
             custom_llm_provider: Final = litellm_logging_obj.model_call_details.get("custom_llm_provider", "openai")
             # Calculate cost using LiteLLM's cost calculator
-            response_cost: Final = litellm.completion_cost(
-                completion_response=complete_response,
-                model=model,
-                custom_llm_provider=custom_llm_provider,
+            response_cost: Final = (
+                litellm.completion_cost(
+                    completion_response=complete_response,
+                    model=model,
+                    custom_llm_provider=custom_llm_provider,
+                    call_type="responses",
+                )
+                if is_responses
+                else litellm.completion_cost(
+                    completion_response=complete_response,
+                    model=model,
+                    custom_llm_provider=custom_llm_provider,
+                )
             )
 
             # Preserve existing litellm_params to maintain metadata tags
@@ -568,6 +583,8 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                 "response_cost": response_cost,
                 "model": model,
                 "custom_llm_provider": custom_llm_provider,
+                "call_type": litellm_logging_obj.call_type,
+                "messages": litellm_logging_obj.model_call_details.get("messages"),
                 "litellm_params": existing_litellm_params.copy(),
             }
 
@@ -584,8 +601,11 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                         user
                     )
 
-            # Create standard logging object
-            get_standard_logging_object_payload(
+            # Attach the payload to kwargs so the success handler adopts it;
+            # its later rebuild runs on a copy whose Responses usage was
+            # coerced to chat shape and serializes as total_tokens only,
+            # zeroing the prompt/completion split in spend logs.
+            standard_logging_object: Final = get_standard_logging_object_payload(
                 kwargs=kwargs,
                 init_response_obj=complete_response,
                 start_time=start_time,
@@ -593,6 +613,8 @@ class OpenAIPassthroughLoggingHandler(BasePassthroughLoggingHandler):
                 logging_obj=litellm_logging_obj,
                 status="success",
             )
+            if standard_logging_object is not None:
+                kwargs["standard_logging_object"] = standard_logging_object
 
             # Update logging object with cost information
             litellm_logging_obj.model_call_details["model"] = model

litellm/proxy/pass_through_endpoints/llm_provider_handlers/vertex_passthrough_logging_handler.py

@@ -1,6 +1,5 @@
 import asyncio
 import re
-from collections.abc import Mapping
 from datetime import datetime
 from typing import TYPE_CHECKING, Any, Final, cast
 from urllib.parse import urlparse
@@ -9,6 +8,7 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_proxy_logger
+from litellm.constants import VERTEX_BATCH_PREDICTION_JOBS_ROUTE
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
     ModelResponseIterator as VertexModelResponseIterator,
@@ -18,6 +18,12 @@ from litellm.llms.vertex_ai.vector_stores.search_api.transformation import (
 )
 from litellm.llms.vertex_ai.videos.transformation import VertexAIVideoConfig
 from litellm.proxy._types import PassThroughEndpointLoggingTypedDict
+from litellm.proxy.pass_through_endpoints.llm_provider_handlers.batch_attribution import (
+    is_collection_route,
+    log_batch_registration_result,
+    optional_str,
+    request_tags_from_metadata,
+)
 from litellm.types.utils import (
     Choices,
     EmbeddingResponse,
@@ -41,32 +47,6 @@ else:
 EndpointType = Any
 
 
-def _optional_str(value: object) -> str | None:
-    return value if isinstance(value, str) else None
-
-
-def _optional_str_tuple(value: object) -> tuple[str, ...] | None:
-    if not isinstance(value, list):
-        return None
-    items: Final = cast(list[object], value)  # cast-ok: isinstance-narrowed; element type unknown
-    return tuple(tag for tag in items if isinstance(tag, str))
-
-
-def _request_tags(request_metadata: Mapping[str, object]) -> tuple[str, ...] | None:
-    """Tags for the batch-cost spend row: the request's own tags when it sent any,
-    otherwise the key's tags, which auth exposes as user_api_key_auth_metadata (a
-    tagged key does not put its tags in the top-level metadata "tags" on the
-    passthrough path)
-    """
-    tags: Final = _optional_str_tuple(request_metadata.get("tags"))
-    if tags:
-        return tags
-    key_auth_metadata: Final = request_metadata.get("user_api_key_auth_metadata")
-    if isinstance(key_auth_metadata, dict):
-        return _optional_str_tuple(key_auth_metadata.get("tags"))
-    return None
-
-
 class VertexPassthroughLoggingHandler:
     @staticmethod
     def vertex_passthrough_handler(
@@ -685,7 +665,7 @@ class VertexPassthroughLoggingHandler:
 
                 # Store the managed object for cost tracking
                 # This will be picked up by check_batch_cost polling mechanism
-                is_batch_create: Final = url_route.split("?")[0].rstrip("/").endswith("batchPredictionJobs")
+                is_batch_create: Final = is_collection_route(url_route, VERTEX_BATCH_PREDICTION_JOBS_ROUTE)
                 VertexPassthroughLoggingHandler._store_batch_managed_object(
                     unified_object_id=unified_object_id,
                     batch_object=litellm_batch_response,
@@ -809,29 +789,6 @@ class VertexPassthroughLoggingHandler:
                 "kwargs": kwargs,
             }
 
-    @staticmethod
-    def _log_batch_registration_result(
-        finished: asyncio.Task, unified_object_id: str, model_object_id: str, is_batch_create: bool
-    ) -> None:
-        error: Final = finished.exception() if not finished.cancelled() else None
-        if finished.cancelled() or error is not None:
-            consequence: Final = (
-                "its cost will not be tracked" if is_batch_create else "its status and output file may be stale"
-            )
-            verbose_proxy_logger.error(
-                "Failed to store batch managed object with unified_object_id=%s, batch_id=%s; %s: %s",
-                unified_object_id,
-                model_object_id,
-                consequence,
-                error,
-            )
-            return
-        verbose_proxy_logger.info(
-            "Stored batch managed object with unified_object_id=%s, batch_id=%s",
-            unified_object_id,
-            model_object_id,
-        )
-
     @staticmethod
     def _store_batch_managed_object(
         unified_object_id: str,
@@ -863,7 +820,7 @@ class VertexPassthroughLoggingHandler:
 
                 user_api_key_dict: Final = UserAPIKeyAuth(
                     user_id=_request_metadata.get("user_api_key_user_id", "default-user"),
-                    api_key=_optional_str(_request_metadata.get("user_api_key")),
+                    api_key=optional_str(_request_metadata.get("user_api_key")),
                     team_id=_request_metadata.get("user_api_key_team_id"),
                     team_alias=None,
                     user_role=LitellmUserRoles.CUSTOMER,  # Use proper enum value
@@ -893,14 +850,14 @@ class VertexPassthroughLoggingHandler:
                         model_object_id=model_object_id,
                         file_purpose="batch",
                         user_api_key_dict=user_api_key_dict,
-                        request_tags=_request_tags(_request_metadata),
+                        request_tags=request_tags_from_metadata(_request_metadata),
                         persist_attribution=is_batch_create,
                         create_if_missing=is_batch_create,
                     )
                 )
                 task.add_done_callback(
-                    lambda finished: VertexPassthroughLoggingHandler._log_batch_registration_result(
-                        finished, unified_object_id, model_object_id, is_batch_create
+                    lambda finished: log_batch_registration_result(
+                        finished, "Vertex AI", unified_object_id, model_object_id, is_batch_create
                     )
                 )
             else:

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -233,15 +233,7 @@ async def chat_completion_pass_through_endpoint(
         # skip router if user passed their key
         if "api_key" in data:
             llm_response = asyncio.create_task(litellm.aadapter_completion(**data))
-        elif llm_router is not None and data["model"] in router_model_names:  # model in router model list
-            llm_response = asyncio.create_task(llm_router.aadapter_completion(**data))
-        elif (
-            llm_router is not None
-            and llm_router.model_group_alias is not None
-            and data["model"] in llm_router.model_group_alias
-        ):  # model set in model_group_alias
-            llm_response = asyncio.create_task(llm_router.aadapter_completion(**data))
-        elif llm_router is not None and llm_router.has_model_id(data["model"]):  # model in router model list
+        elif llm_router is not None and llm_router.is_recognized_model(data["model"]):
             llm_response = asyncio.create_task(llm_router.aadapter_completion(**data))
         elif (
             llm_router is not None
@@ -565,6 +557,7 @@ class HttpPassThroughEndpointHelpers(BasePassthroughUtils):
         # real parent span.
         _metadata["user_api_key"] = user_api_key_dict.api_key
         _metadata["litellm_parent_otel_span"] = user_api_key_dict.parent_otel_span
+        _metadata["user_api_key_budget_reservation"] = user_api_key_dict.budget_reservation
         _metadata.update(
             LiteLLMProxyRequestSetup.get_sanitized_user_information_from_key(user_api_key_dict=user_api_key_dict)
         )

litellm/proxy/proxy_server.py

@@ -171,6 +171,7 @@ try:
     import orjson
     import yaml
     from apscheduler.schedulers.asyncio import AsyncIOScheduler
+    from apscheduler.triggers.interval import IntervalTrigger
 except ImportError as e:
     raise ImportError(f"Missing dependency {e}. Run `pip install 'litellm[proxy]'`")
 
@@ -344,6 +345,12 @@ from litellm.proxy.common_utils.periodic_reload_schedule import (
 )
 from litellm.proxy.common_utils.proxy_state import ProxyState
 from litellm.proxy.common_utils.reset_budget_job import ResetBudgetJob
+from litellm.proxy.common_utils.scheduled_job_stagger import (
+    apply_scheduled_job_stagger,
+    attach_job_timing_logger,
+    parse_stagger_settings,
+    stagger_trigger,
+)
 from litellm.proxy.common_utils.swagger_utils import ERROR_RESPONSES
 from litellm.proxy.common_utils.timezone_utils import (
     get_budget_reset_settings,
@@ -460,6 +467,7 @@ from litellm.proxy.management_endpoints.model_management_endpoints import (
     _add_model_to_db,
     _add_team_model_to_db,
     _deduplicate_litellm_router_models,
+    live_model_ids_snapshot,
 )
 from litellm.proxy.management_endpoints.model_management_endpoints import (
     router as model_management_router,
@@ -837,6 +845,22 @@ def cleanup_router_config_variables():
     prisma_client = None
 
 
+async def _flush_spend_logs_queue_on_shutdown() -> None:
+    if prisma_client is None:
+        return
+
+    try:
+        from litellm.proxy.utils import drain_spend_logs_queue
+
+        await drain_spend_logs_queue(
+            prisma_client=prisma_client,
+            db_writer_client=db_writer_client,
+            proxy_logging_obj=proxy_logging_obj,
+        )
+    except Exception as e:  # noqa: BLE001  # shutdown must continue even if the drain fails
+        verbose_proxy_logger.exception("Error flushing spend logs queue on shutdown: %s", e)
+
+
 async def proxy_shutdown_event():
     global prisma_client, master_key, user_custom_auth, user_custom_key_generate, user_custom_key_update
     verbose_proxy_logger.info("Shutting down LiteLLM Proxy Server")
@@ -1247,6 +1271,8 @@ async def proxy_startup_event(app: FastAPI):
         except Exception as e:
             verbose_proxy_logger.error("Error stopping DB health watchdog task: %s", e)
 
+    await _flush_spend_logs_queue_on_shutdown()
+
     await proxy_config.stop_config_sync_subscriber()
 
     await proxy_config.stop_auth_cache_invalidation_subscriber()
@@ -2159,6 +2185,15 @@ experimental = False
 #### GLOBAL VARIABLES ####
 llm_router: Router | None = None
 llm_model_list: list | None = None
+# Serializes every model reconcile (ProxyConfig.add_deployment and clear_cache) so the
+# read-modify-write of llm_router above is atomic. Without it, two concurrent model
+# writes each reconcile the router against their OWN db snapshot, and the one holding
+# the older snapshot evicts the deployment the newer one just added -- the db keeps the
+# row, this pod stops serving it. Control-plane only (model create/update/delete and
+# the config-sync tick), never on a completion path, so the serialization is free.
+# Module-level rather than per-ProxyConfig because llm_router is a module global and a
+# second ProxyConfig instance must not get its own independent lock over it.
+MODEL_RECONCILE_LOCK: Final = asyncio.Lock()
 general_settings: dict = {}
 config_passthrough_endpoints: list[dict[str, Any]] | None = None
 log_file: Final = "api_log.json"
@@ -6142,10 +6177,17 @@ class ProxyConfig:
                 retention_interval: Final = general_settings.get("maximum_spend_logs_retention_interval", "1d")
                 try:
                     interval_seconds: Final = duration_in_seconds(retention_interval)
+                    # this runs against a started scheduler, which the startup stagger sweep
+                    # cannot reach, so the offset is applied here or the job reconverges across
+                    # replicas the first time an admin edits the retention settings
                     scheduler.add_job(
                         spend_log_cleanup.cleanup_old_spend_logs,
-                        "interval",
-                        seconds=interval_seconds + random.randint(0, 60),
+                        stagger_trigger(
+                            job_id="spend_log_cleanup_job",
+                            trigger=IntervalTrigger(seconds=interval_seconds),
+                            period_seconds=interval_seconds,
+                            settings=parse_stagger_settings(general_settings),
+                        ),
                         args=[prisma_client],
                         id="spend_log_cleanup_job",
                         replace_existing=True,
@@ -6442,16 +6484,37 @@ class ProxyConfig:
         self,
         prisma_client: PrismaClient,
         proxy_logging_obj: ProxyLogging,
-    ) -> frozenset[str] | None:
+    ) -> ReconcileOutcome:
         """
         - Check db for new models
         - Check if model id's in router already
         - If not, add to router
 
-        Returns the ids the db + config say should be served after the reconcile, or
-        None when no reconcile ran. Callers that judge their own reload need it to tell
-        a deliberate eviction from a deployment that went missing.
+        Serialized against every other model reconcile by MODEL_RECONCILE_LOCK, because
+        the work below is a read-modify-write of the shared ``llm_router`` global: it
+        reads the db into a snapshot and then makes the router match that snapshot. Two
+        of those interleaving is not a lost update but an eviction -- the request whose
+        snapshot predates the other's commit reconciles the newer model *out* of the
+        router, since _delete_deployment removes every live deployment absent from the
+        snapshot it was handed. The model stays in the db and this pod stops serving it
+        until some later reload puts it back.
+
+        Returns what the reconcile saw, captured before the lock is released so a
+        caller's verdict cannot be corrupted by the next reconcile's own in-flight
+        window. See ReconcileOutcome.
         """
+        async with MODEL_RECONCILE_LOCK:
+            return await self._add_deployment_locked(prisma_client=prisma_client, proxy_logging_obj=proxy_logging_obj)
+
+    async def _add_deployment_locked(
+        self,
+        prisma_client: PrismaClient,
+        proxy_logging_obj: ProxyLogging,
+    ) -> ReconcileOutcome:
+        """add_deployment's body, minus the locking. MODEL_RECONCILE_LOCK MUST already
+        be held. Split out for the one caller that has to hold the lock across more than
+        this reconcile -- clear_cache, which un-serves every db model before calling it
+        and would deadlock on a re-acquire."""
         global llm_router, llm_model_list, master_key, general_settings
 
         still_desired_ids: frozenset[str] | None = None
@@ -6494,7 +6557,12 @@ class ProxyConfig:
         except Exception as e:
             verbose_proxy_logger.exception("litellm.proxy.proxy_server.py::ProxyConfig:add_deployment - %s", e)
 
-        return still_desired_ids
+        # Read while the lock is still held: once it is released the next reconcile can
+        # begin, and clear_cache's leading wipe would make this look like a mass drop.
+        return ReconcileOutcome(
+            still_desired=still_desired_ids,
+            live_after=None if still_desired_ids is None else live_model_ids_snapshot(),
+        )
 
     def start_config_sync_subscriber(
         self,
@@ -7865,8 +7933,17 @@ def _resolve_keepalive_seconds(request_data: Mapping[str, Any], response: object
     # keepalive_seconds is operator-only unless the deployment explicitly opts in:
     # a client can't unilaterally enable heartbeats (and the LB-idle-timeout
     # evasion that comes with them) for a deployment that never configured this.
+    # When neither the request nor the deployment sets a value, the operator's
+    # global `litellm_settings.sse_keepalive_ping_interval_seconds` applies; a
+    # deployment's explicit `keepalive_seconds: 0` above still hard-disables it.
     client_supplied: Final = request_data.get("keepalive_seconds") if allow_client_override else None
-    raw: Final = client_supplied if client_supplied is not None else deployment_raw
+    raw: Final = (
+        client_supplied
+        if client_supplied is not None
+        else deployment_raw
+        if deployment_raw is not None
+        else litellm.sse_keepalive_ping_interval_seconds
+    )
     try:
         value: Final = float(raw) if isinstance(raw, (int, float, str)) else 0.0
     except ValueError:
@@ -7977,18 +8054,19 @@ async def async_data_generator(
 
         # A stream can start on a deployment with keepalive off and fall back
         # mid-stream to one that enables it: only skip wrapping altogether when
-        # there's no router to ever fall back through in the first place (in
-        # which case _resolve_keepalive_seconds can never return non-zero for
-        # any chunk of this stream), not merely because the first chunk's
-        # deployment happens to start with it off.
+        # there's no router to ever fall back through AND the resolved interval
+        # (including the global sse_keepalive_ping_interval_seconds fallback)
+        # starts disabled, not merely because the first chunk's deployment
+        # happens to start with it off.
         resolve_keepalive_seconds: Final = _make_keepalive_resolver(request_data)
+        initial_keepalive_seconds: Final = resolve_keepalive_seconds(response)
         stream_source: Final = (
             _iter_with_keepalive(
                 stream_iterator.__aiter__(),
                 resolve_keepalive_seconds,
-                resolve_keepalive_seconds(response),
+                initial_keepalive_seconds,
             )
-            if llm_router is not None
+            if llm_router is not None or initial_keepalive_seconds > 0
             else stream_iterator
         )
 
@@ -8671,14 +8749,14 @@ class ProxyStartupEvent:
         if general_settings.get("disable_spend_logs", False) is False:
             from litellm.proxy.utils import _monitor_spend_logs_queue
 
-            # Start background task to monitor spend logs queue size
-            asyncio.create_task(
+            monitor_task: Final = asyncio.create_task(
                 _monitor_spend_logs_queue(
                     prisma_client=prisma_client,
                     db_writer_client=db_writer_client,
                     proxy_logging_obj=proxy_logging_obj,
                 )
             )
+            prisma_client.spend_logs_queue_monitor_task = monitor_task  # rebind-ok: the client owns its monitor handle
 
         ### ADD NEW MODELS ###
         store_model_in_db = get_secret_bool("STORE_MODEL_IN_DB", store_model_in_db) or store_model_in_db
@@ -8931,6 +9009,14 @@ class ProxyStartupEvent:
         # Do NOT reset job times to "now" as this can trigger the memory leak
         # The misfire_grace_time and coalesce settings will handle any missed runs properly
 
+        # Every job above anchors on this process's start instant, so without a phase offset
+        # they all fire together, on every replica the rollout brought up at the same time
+        attach_job_timing_logger(scheduler)
+        apply_scheduled_job_stagger(
+            scheduler=scheduler,
+            settings=parse_stagger_settings(general_settings),
+        )
+
         # Start the scheduler immediately without processing backlogs
         scheduler.start(paused=False)
         verbose_proxy_logger.info(
@@ -11858,6 +11944,8 @@ def _add_team_models_to_all_models(
     Add team models to all models
     """
     team_models: Final[dict[str, set[str]]] = {}
+    proxy_model_list: Final = llm_router.get_model_names()
+    model_access_groups: Final = llm_router.get_model_access_groups()
 
     for team_object in team_db_objects_typed:
         if (
@@ -11879,7 +11967,12 @@ def _add_team_models_to_all_models(
                     if can_add_model:
                         team_models.setdefault(model_id, set()).add(team_object.team_id)
         else:
-            for model_name in team_object.models:
+            resolved_model_names = get_team_models(
+                team_models=team_object.models,
+                proxy_model_list=proxy_model_list,
+                model_access_groups=model_access_groups,
+            )
+            for model_name in resolved_model_names:
                 _models = llm_router.get_model_list(model_name=model_name, team_id=team_object.team_id)
                 if _models is not None:
                     for model in _models:

litellm/proxy/response_api_endpoints/endpoints.py

@@ -121,7 +121,7 @@ def _parse_cursor_model_variant(model: str) -> _CursorModelVariant:
 def _router_can_serve(model: str, llm_router: "Router | None") -> bool:
     if llm_router is None:
         return False
-    if model in llm_router.model_names or model in llm_router.model_group_alias:
+    if llm_router.is_recognized_model(model):
         return True
     if model in llm_router.team_public_model_names:
         return True

litellm/proxy/route_llm_request.py

@@ -587,16 +587,10 @@ async def route_request(
             return getattr(llm_router, f"{route_type}")(**data)
 
         elif (
-            (
-                is_proxy_admin_without_team
-                and data["model"] not in router_model_names
-                and data["model"] in llm_router.team_public_model_names
-            )
-            or data["model"] in router_model_names
-            or llm_router.has_model_id(data["model"])
-            or llm_router.model_group_alias is not None
-            and data["model"] in llm_router.model_group_alias
-        ):
+            is_proxy_admin_without_team
+            and data["model"] not in router_model_names
+            and data["model"] in llm_router.team_public_model_names
+        ) or llm_router.is_recognized_model(data["model"]):
             return getattr(llm_router, f"{route_type}")(**data)
 
         elif data["model"] not in router_model_names:

litellm/proxy/utils.py

@@ -1,4 +1,5 @@
 import asyncio
+import contextlib
 import copy
 import hashlib
 import inspect
@@ -3006,6 +3007,7 @@ async def prefetch_config_params(prisma_client: "PrismaClient | None", param_nam
 class PrismaClient:
     spend_log_transactions: list = []
     _spend_log_transactions_lock = asyncio.Lock()
+    spend_logs_queue_monitor_task: "asyncio.Task[None] | None" = None
     tool_usage_transactions: list["ToolUsageTransaction"] = []
     _tool_usage_transactions_lock = asyncio.Lock()
     autorouter_turn_transactions: ClassVar[
@@ -5722,13 +5724,22 @@ async def update_spend_logs_job(
         logs_to_process: Final = prisma_client.spend_log_transactions[:MAX_LOGS_PER_INTERVAL]
         prisma_client.spend_log_transactions = prisma_client.spend_log_transactions[len(logs_to_process) :]
 
-    await ProxyUpdateSpend.update_spend_logs(
-        n_retry_times=n_retry_times,
-        prisma_client=prisma_client,
-        proxy_logging_obj=proxy_logging_obj,
-        db_writer_client=db_writer_client,
-        logs_to_process=logs_to_process,
-    )
+    try:
+        await ProxyUpdateSpend.update_spend_logs(
+            n_retry_times=n_retry_times,
+            prisma_client=prisma_client,
+            proxy_logging_obj=proxy_logging_obj,
+            db_writer_client=db_writer_client,
+            logs_to_process=logs_to_process,
+        )
+    except asyncio.CancelledError:
+        async with prisma_client._spend_log_transactions_lock:
+            prisma_client.spend_log_transactions[:0] = logs_to_process
+        verbose_proxy_logger.warning(
+            "Spend tracking - spend log write cancelled, requeued %d rows for the next flush",
+            len(logs_to_process),
+        )
+        raise
 
     # Guardrail/policy usage tracking (same batch, outside spend-logs update)
     try:
@@ -5787,6 +5798,39 @@ async def update_spend_logs_job(
         )
 
 
+MAX_SPEND_LOG_DRAIN_ITERATIONS: Final = 20
+
+
+async def drain_spend_logs_queue(
+    prisma_client: PrismaClient,
+    db_writer_client: "AsyncHTTPHandler | None",
+    proxy_logging_obj: ProxyLogging,
+) -> None:
+    monitor_task: Final = prisma_client.spend_logs_queue_monitor_task
+    if monitor_task is not None:
+        monitor_task.cancel()
+        with contextlib.suppress(asyncio.CancelledError):
+            await monitor_task
+        prisma_client.spend_logs_queue_monitor_task = None  # rebind-ok: the client owns its monitor handle
+
+    for _ in range(MAX_SPEND_LOG_DRAIN_ITERATIONS):
+        if await _total_queued_spend_transactions(prisma_client) == 0:
+            return
+        await update_spend_logs_job(
+            prisma_client=prisma_client,
+            db_writer_client=db_writer_client,
+            proxy_logging_obj=proxy_logging_obj,
+        )
+
+    remaining: Final = await _total_queued_spend_transactions(prisma_client)
+    if remaining > 0:
+        spend_log_error(
+            "Spend tracking - %d spend log rows still queued after %d drain passes",
+            remaining,
+            MAX_SPEND_LOG_DRAIN_ITERATIONS,
+        )
+
+
 async def _monitor_spend_logs_queue(
     prisma_client: PrismaClient,
     db_writer_client: AsyncHTTPHandler | None,

litellm/responses/litellm_completion_transformation/custom_tools.py

@@ -111,7 +111,7 @@ class _CustomToolFormat(BaseModel):
 _ALLOWED_CALLERS_ADAPTER: Final = TypeAdapter(list[str] | None)
 
 
-def _validated_allowed_callers(value: object) -> list[str] | None:
+def validated_allowed_callers(value: object) -> list[str] | None:
     try:
         return _ALLOWED_CALLERS_ADAPTER.validate_python(value, strict=True)
     except ValidationError as exc:
@@ -143,7 +143,7 @@ def convert_custom_tool_to_function_tool(tool: Mapping[str, object]) -> ChatComp
     name: Final = raw_name if isinstance(raw_name, str) else ""
     raw_description: Final = tool.get("description")
     description = (raw_description if isinstance(raw_description, str) else "") + _grammar_suffix(tool.get("format"))
-    allowed_callers: Final = _validated_allowed_callers(tool.get("allowed_callers"))
+    allowed_callers: Final = validated_allowed_callers(tool.get("allowed_callers"))
     function_chunk: Final = ChatCompletionToolParamFunctionChunk(
         name=name,
         description=description,

litellm/responses/litellm_completion_transformation/streaming_iterator.py

@@ -82,6 +82,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
         self.sent_output_item_done_event: bool = False
         self.sent_annotation_events: bool = False
         self.litellm_model_response: ModelResponse | TextCompletionResponse | None = None
+        self.completed_response: Any = None
         self.final_text: str = ""
         self._cached_item_id: str | None = None
         self._cached_response_id: str | None = None
@@ -105,6 +106,9 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
         self._accumulated_reasoning_content_parts: list[str] = []
         self._accumulated_provider_specific_fields: dict[str, Any] = {}
         self._custom_tool_names: set[str] = extract_custom_tool_names(self.responses_api_request.get("tools"))
+        self._namespace_tool_names = LiteLLMCompletionResponsesConfig.namespace_tool_name_map(
+            self.responses_api_request.get("tools")
+        )
 
     def _get_or_assign_tool_output_index(self, call_id: str) -> int:
         existing: Final = self._tool_output_index_by_call_id.get(call_id)
@@ -124,6 +128,13 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
         except (TypeError, ValueError):
             return None
 
+    def _responses_namespace_tool_call_fields(self, fn_name: str) -> tuple[str, str | None]:
+        mapped: Final = self._namespace_tool_names.get(fn_name)
+        if mapped:
+            namespace, tool_name = mapped
+            return tool_name, namespace
+        return fn_name, None
+
     def _is_reasoning_end(self, chunk):
         delta: Final = chunk.choices[0].delta
 
@@ -182,13 +193,17 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
             else:
                 fn_name = str(getattr(fn, "name", "") or "")
                 fn_args_delta = str(getattr(fn, "arguments", "") or "")
+            tool_name, tool_namespace = self._responses_namespace_tool_call_fields(fn_name)
 
             output_index = self._get_or_assign_tool_output_index(call_id)
 
             if call_id not in self._tool_args_by_call_id:
                 self._tool_args_by_call_id[call_id] = ""
                 self._sequence_number += 1
-                item_kwargs = build_tool_call_item_kwargs(call_id, fn_name, "", "in_progress", self._custom_tool_names)
+                names = self._custom_tool_names
+                item_kwargs = build_tool_call_item_kwargs(call_id, tool_name, "", "in_progress", names)
+                if tool_namespace:
+                    item_kwargs["namespace"] = tool_namespace
                 event = OutputItemAddedEvent(
                     type=ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED,
                     output_index=output_index,
@@ -249,6 +264,7 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
             else:
                 fn_name = str(getattr(fn, "name", "") or "")
                 fn_args = str(getattr(fn, "arguments", "") or "")
+            tool_name, tool_namespace = self._responses_namespace_tool_call_fields(fn_name)
 
             # Track if this is a new tool call that wasn't streamed
             is_new_tool_call = call_id not in self._tool_args_by_call_id
@@ -257,7 +273,10 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
             if is_new_tool_call:
                 self._tool_args_by_call_id[call_id] = ""
                 self._sequence_number += 1
-                item_kwargs = build_tool_call_item_kwargs(call_id, fn_name, "", "in_progress", self._custom_tool_names)
+                names = self._custom_tool_names
+                item_kwargs = build_tool_call_item_kwargs(call_id, tool_name, "", "in_progress", names)
+                if tool_namespace:
+                    item_kwargs["namespace"] = tool_namespace
                 event = OutputItemAddedEvent(
                     type=ResponsesAPIStreamEvents.OUTPUT_ITEM_ADDED,
                     output_index=output_index,
@@ -299,9 +318,10 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
             self._pending_tool_events.append(done_event)
 
             self._sequence_number += 1
-            item_kwargs = build_tool_call_item_kwargs(
-                call_id, fn_name, final_args, "completed", self._custom_tool_names
-            )
+            names = self._custom_tool_names
+            item_kwargs = build_tool_call_item_kwargs(call_id, tool_name, final_args, "completed", names)
+            if tool_namespace:
+                item_kwargs["namespace"] = tool_namespace
             item_done_event = OutputItemDoneEvent(
                 type=ResponsesAPIStreamEvents.OUTPUT_ITEM_DONE,
                 output_index=output_index,

litellm/responses/litellm_completion_transformation/transformation.py

@@ -5,7 +5,17 @@ Handles transforming from Responses API -> LiteLLM completion  (Chat Completion
 import json
 import re
 from collections.abc import Iterator, Mapping, Sequence
-from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, cast, runtime_checkable
+from types import MappingProxyType
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Final,
+    Literal,
+    Protocol,
+    TypeAlias,
+    cast,
+    runtime_checkable,
+)
 
 from openai.types.chat.chat_completion_named_tool_choice_param import (
     ChatCompletionNamedToolChoiceParam,
@@ -38,9 +48,11 @@ from litellm.types.llms.openai import (
     ChatCompletionToolCallFunctionChunk,
     ChatCompletionToolMessage,
     ChatCompletionToolParam,
+    ChatCompletionToolParamFunctionChunk,
     ChatCompletionUserMessage,
     GenericChatCompletionMessage,
     InputTokensDetails,
+    OpenAIChatCompletionTextObject,
     OpenAIMcpServerTool,
     OpenAIWebSearchOptions,
     OpenAIWebSearchUserLocation,
@@ -77,8 +89,13 @@ from .custom_tools import (
     extract_custom_tool_names,
     is_custom_tool_call,
     unwrap_custom_tool_arguments,
+    validated_allowed_callers,
 )
 
+NamespaceNameMap: TypeAlias = Mapping[str, tuple[str, str]]
+NamespaceTool: TypeAlias = Mapping[str, object]
+ResponseTools: TypeAlias = Sequence[Mapping[str, object]] | None
+
 if TYPE_CHECKING:
     from openai.types.responses.response_apply_patch_tool_call import (
         ResponseApplyPatchToolCall,
@@ -528,9 +545,52 @@ class LiteLLMCompletionResponsesConfig:
                     messages.extend(deduped_in_place)
                     continue
 
+                merged_assistant = LiteLLMCompletionResponsesConfig._merged_trailing_assistant_message(
+                    messages=messages,
+                    chat_completion_messages=chat_completion_messages,
+                )
+                if merged_assistant is not None:
+                    messages[-1] = merged_assistant
+                    continue
+
                 messages.extend(chat_completion_messages)
         return messages
 
+    @staticmethod
+    def _merged_trailing_assistant_message(
+        messages: Sequence[
+            AllMessageValues
+            | GenericChatCompletionMessage
+            | ChatCompletionMessageToolCall
+            | ChatCompletionResponseMessage
+        ],
+        chat_completion_messages: Sequence[
+            AllMessageValues | GenericChatCompletionMessage | ChatCompletionResponseMessage
+        ],
+    ) -> ChatCompletionResponseMessage | None:
+        """Fold an assistant content message into a directly preceding assistant
+        tool_calls message. Providers like DeepSeek and Anthropic require tool
+        results immediately after the tool_calls message, so an assistant message
+        between them is rejected."""
+        if not messages or len(chat_completion_messages) != 1:
+            return None
+        last_message = messages[-1]
+        new_message = chat_completion_messages[0]
+        if not isinstance(last_message, dict):
+            return None
+        if last_message.get("role") != "assistant" or new_message.get("role") != "assistant":
+            return None
+        if not last_message.get("tool_calls") or last_message.get("content") or new_message.get("tool_calls"):
+            return None
+        new_content = new_message.get("content")
+        if new_content is None:
+            return None
+        merged: Final = {  # mutable-ok: json.dumps rejects MappingProxyType in outbound chat messages
+            **last_message,
+            "content": new_content,
+        }
+        return cast(ChatCompletionResponseMessage, merged)  # cast-ok: TypedDict spread widens to dict[str, object]
+
     @staticmethod
     def _deduplicate_tool_call_output_messages(
         tool_call_output_messages: list[
@@ -1163,11 +1223,14 @@ class LiteLLMCompletionResponsesConfig:
         if not raw_arguments and function_call.get("type") == "custom_tool_call":
             raw_input: Final = function_call.get("input") or ""
             raw_arguments = json.dumps({"content": raw_input}) if raw_input else ""
+        raw_name: Final = function_call.get("name") or ""
+        namespace: Final = function_call.get("namespace") or ""
+        qualify: Final = bool(namespace) and function_call.get("type") != "custom_tool_call"
         tool_call: Final = ChatCompletionToolCallChunk(
             id=function_call.get("call_id") or function_call.get("id") or "",
             type="function",
             function=ChatCompletionToolCallFunctionChunk(
-                name=function_call.get("name") or "",
+                name=f"{namespace}__{raw_name}" if qualify else raw_name,
                 arguments=str(raw_arguments or ""),
             ),
             index=0,
@@ -1260,6 +1323,12 @@ class LiteLLMCompletionResponsesConfig:
                         if "cache_control" in item:
                             image_block["cache_control"] = item["cache_control"]
                         content_list.append(image_block)
+                    elif item.get("type") == "encrypted_content":
+                        encrypted_content = item.get("encrypted_content")
+                        if encrypted_content is not None:
+                            content_list.append(
+                                OpenAIChatCompletionTextObject(type="text", text=str(encrypted_content))
+                            )
                     else:
                         # Skip text blocks with None text to avoid downstream errors
                         text_value = item.get("text")
@@ -1320,6 +1389,92 @@ class LiteLLMCompletionResponsesConfig:
         """
         return ChatCompletionSystemMessage(role="system", content=instructions or "")
 
+    @staticmethod
+    def _build_ns_chat_tool(
+        namespace: str,
+        namespace_description: str,
+        namespace_tool: NamespaceTool,
+        nested: bool,
+    ) -> ChatCompletionToolParam | None:
+        if nested and namespace_tool.get("type") != "function":
+            return None
+
+        raw_parameters: Final = namespace_tool.get("parameters")
+        parameters: Final = (
+            MappingProxyType(raw_parameters) if isinstance(raw_parameters, Mapping) else MappingProxyType({})
+        )
+        normalized_parameters: Final = (
+            parameters if parameters and "type" in parameters else MappingProxyType({**parameters, "type": "object"})
+        )
+        tool_name: Final = str(namespace_tool.get("name") or "")
+        raw_description: Final = str(namespace_tool.get("description") or "")
+        description: Final = (
+            f"{namespace_description}\n\n{raw_description}"
+            if nested and namespace_description and raw_description
+            else namespace_description
+            if nested and namespace_description
+            else raw_description
+        )
+        chat_tool_name: Final = f"{namespace}__{tool_name}" if nested else tool_name
+        function: Final = ChatCompletionToolParamFunctionChunk(
+            name=chat_tool_name,
+            description=description,
+            parameters=dict(  # mutable-ok: json.dumps rejects MappingProxyType in the outbound payload
+                normalized_parameters
+            ),
+            strict=bool(namespace_tool.get("strict", False)),
+        )
+        allowed_callers: Final = validated_allowed_callers(namespace_tool.get("allowed_callers"))
+        if allowed_callers is None:
+            return ChatCompletionToolParam(type="function", function=function)
+        return ChatCompletionToolParam(type="function", function=function, allowed_callers=allowed_callers)
+
+    @staticmethod
+    def _namespace_chat_tools(tool: NamespaceTool) -> tuple[ChatCompletionToolParam, ...]:
+        namespace: Final = str(tool.get("name") or "")
+        namespace_description: Final = str(tool.get("description") or "")
+        namespace_tools: Final = tool.get("tools")
+        if isinstance(namespace_tools, Sequence) and not isinstance(namespace_tools, (str, bytes)):
+            return tuple(
+                chat_tool
+                for raw_tool in namespace_tools
+                if isinstance(raw_tool, Mapping)
+                if (
+                    chat_tool := LiteLLMCompletionResponsesConfig._build_ns_chat_tool(
+                        namespace,
+                        namespace_description,
+                        raw_tool,
+                        True,
+                    )
+                )
+                is not None
+            )
+        flat_tool: Final = LiteLLMCompletionResponsesConfig._build_ns_chat_tool(
+            namespace, namespace_description, tool, False
+        )
+        return (flat_tool,) if flat_tool is not None else ()
+
+    @staticmethod
+    def _validate_namespace_name_collisions(tools: ResponseTools) -> None:
+        top_level_function_names: Final = frozenset(
+            str(tool.get("name") or "") for tool in tools or () if tool.get("type") == "function"
+        )
+        flattened_namespace_names: Final = frozenset(
+            f"{(tool.get('name') or '')!s}__{(namespace_tool.get('name') or '')!s}"
+            for tool in tools or ()
+            if tool.get("type") == "namespace"
+            for namespace_tools in (tool.get("tools"),)
+            if isinstance(namespace_tools, Sequence) and not isinstance(namespace_tools, (str, bytes))
+            for namespace_tool in namespace_tools
+            if isinstance(namespace_tool, Mapping) and namespace_tool.get("type") == "function"
+        )
+        conflicting_tool_names: Final = top_level_function_names & flattened_namespace_names
+        if conflicting_tool_names:
+            raise ValueError(
+                "Top-level function names conflict with flattened namespace tools: "
+                + ", ".join(sorted(conflicting_tool_names))
+            )
+
     @staticmethod
     def transform_responses_api_tools_to_chat_completion_tools(
         tools: list[FunctionToolParam | OpenAIMcpServerTool] | None,
@@ -1332,6 +1487,7 @@ class LiteLLMCompletionResponsesConfig:
         """
         if tools is None:
             return [], None
+        LiteLLMCompletionResponsesConfig._validate_namespace_name_collisions(tools)
         chat_completion_tools: Final[list[ChatCompletionToolParam | OpenAIMcpServerTool]] = []
         web_search_options: OpenAIWebSearchOptions | None = None
         for tool in tools:
@@ -1373,13 +1529,15 @@ class LiteLLMCompletionResponsesConfig:
                 if tool.get("input_examples"):
                     chat_completion_tool["input_examples"] = tool.get("input_examples")
                 chat_completion_tools.append(cast(ChatCompletionToolParam, chat_completion_tool))
+            elif tool.get("type") == "namespace":
+                chat_completion_tools.extend(LiteLLMCompletionResponsesConfig._namespace_chat_tools(tool))
             elif tool.get("type") == "custom":
                 converted = convert_custom_tool_to_function_tool(tool)
                 if converted is not None:
                     chat_completion_tools.append(converted)
             else:
                 _tool_type = tool.get("type")
-                if _tool_type in ("computer_use", "image_generation", "namespace", "shell"):
+                if _tool_type in ("computer_use", "image_generation", "shell"):
                     # Drop unsupported Responses-API-only tool types that have no
                     # Chat Completions equivalent. Passing them through verbatim
                     # causes providers to reject the request with "'function' is a
@@ -1435,6 +1593,44 @@ class LiteLLMCompletionResponsesConfig:
                 result.append(dict(tool))
         return result
 
+    @staticmethod
+    def namespace_tool_name_map(tools: ResponseTools) -> NamespaceNameMap:
+        namespace_entries: Final = tuple(
+            (str(tool.get("name") or ""), str(namespace_tool.get("name") or ""))
+            for tool in tools or ()
+            if tool.get("type") == "namespace"
+            for namespace_tools in (tool.get("tools"),)
+            if isinstance(namespace_tools, Sequence) and not isinstance(namespace_tools, (str, bytes))
+            for namespace_tool in namespace_tools
+            if isinstance(namespace_tool, Mapping) and namespace_tool.get("type") == "function"
+        )
+        top_level_function_names: Final = frozenset(
+            str(tool.get("name") or "") for tool in tools or () if tool.get("type") == "function"
+        )
+        unqualified_counts: Final = MappingProxyType(
+            {
+                tool_name: sum(1 for _, candidate_name in namespace_entries if candidate_name == tool_name)
+                for tool_name in frozenset(tool_name for _, tool_name in namespace_entries)
+            }
+        )
+        unambiguous_entries: Final = tuple(
+            (tool_name, (namespace, tool_name))
+            for namespace, tool_name in namespace_entries
+            if tool_name not in top_level_function_names and unqualified_counts[tool_name] == 1
+        )
+        qualified_entries: Final = tuple(
+            (f"{namespace}__{tool_name}", (namespace, tool_name)) for namespace, tool_name in namespace_entries
+        )
+        return MappingProxyType(dict(qualified_entries + unambiguous_entries))
+
+    @staticmethod
+    def _restore_namespace_tool_name(tool_name: str, names: NamespaceNameMap) -> tuple[str, str | None]:
+        mapped = names.get(tool_name)
+        if mapped is None:
+            return tool_name, None
+        namespace, restored_tool_name = mapped
+        return restored_tool_name, namespace
+
     @staticmethod
     def transform_chat_completion_tools_to_responses_tools(
         chat_completion_response: ModelResponse,
@@ -1458,10 +1654,9 @@ class LiteLLMCompletionResponsesConfig:
                             value=tool_call,
                         )
 
-        # Extract custom tool names from the original request
-        custom_tool_names: set[str] = set()
-        if responses_api_request and "tools" in responses_api_request:
-            custom_tool_names = extract_custom_tool_names(responses_api_request["tools"])
+        request_tools: Final = responses_api_request.get("tools") if responses_api_request is not None else None
+        custom_tool_names: Final = extract_custom_tool_names(request_tools)
+        namespace_tool_names: Final = LiteLLMCompletionResponsesConfig.namespace_tool_name_map(request_tools)
 
         responses_tools: Final[list[ResponseFunctionToolCall | CustomToolCallOutputItem]] = []
         for tool in all_chat_completion_tools:
@@ -1486,6 +1681,9 @@ class LiteLLMCompletionResponsesConfig:
                     responses_tools.append(custom_item)
                 else:
                     # Build regular function_call output item
+                    restore_name = LiteLLMCompletionResponsesConfig._restore_namespace_tool_name
+                    tool_name, namespace = restore_name(tool_name, namespace_tool_names)
+
                     provider_specific_fields: dict | None = None
                     if hasattr(tool, "provider_specific_fields") and getattr(tool, "provider_specific_fields", None):
                         provider_specific_fields = getattr(tool, "provider_specific_fields")
@@ -1510,6 +1708,8 @@ class LiteLLMCompletionResponsesConfig:
                         type="function_call",
                         status=function_definition.get("status") or "completed",
                     )
+                    if namespace:
+                        output_tool_call.namespace = namespace
 
                     # Pass through provider_specific_fields as-is if present
                     if provider_specific_fields:

litellm/responses/utils.py

@@ -1033,13 +1033,18 @@ class ResponseAPILoggingUtils:
 
     @staticmethod
     def _transform_response_api_usage_to_chat_usage(
-        usage_input: dict | ResponseAPIUsage | None,
+        usage_input: Mapping[str, object] | ResponseAPIUsage | Usage | None,
     ) -> Usage:
         """
         Transforms ResponseAPIUsage or ImageUsage to a Usage object.
 
         Both have the same spec with input_tokens, output_tokens, and
         input_tokens_details (text_tokens, image_tokens).
+
+        Usage inputs are returned as-is so re-running this helper never drops
+        fields. Non-standard provider fields (e.g. xAI's
+        server_side_tool_usage_details) are carried onto the returned Usage so
+        provider cost calculators can read them after normalization.
         """
         if usage_input is None:
             return Usage(
@@ -1047,6 +1052,10 @@ class ResponseAPILoggingUtils:
                 completion_tokens=0,
                 total_tokens=0,
             )
+        if isinstance(usage_input, Usage):
+            return usage_input
+        if isinstance(usage_input, dict) and not ResponseAPILoggingUtils._is_response_api_usage(usage_input):
+            return Usage(**usage_input)
         response_api_usage: ResponseAPIUsage
         if isinstance(usage_input, dict):
             usage_input = dict(usage_input)  # shallow copy; avoid mutating caller
@@ -1055,13 +1064,11 @@ class ResponseAPILoggingUtils:
                 usage_input["input_tokens_details"] = usage_input["input_token_details"]
             if usage_input.get("output_tokens_details") is None and "output_token_details" in usage_input:
                 usage_input["output_tokens_details"] = usage_input["output_token_details"]
-            total_tokens = usage_input.get("total_tokens")
-            if total_tokens is None:
+            if usage_input.get("total_tokens") is None:
                 input_tokens: Final = usage_input.get("input_tokens")
                 output_tokens: Final = usage_input.get("output_tokens")
-                if input_tokens is not None and output_tokens is not None:
-                    total_tokens = input_tokens + output_tokens
-                    usage_input["total_tokens"] = total_tokens
+                if isinstance(input_tokens, int) and isinstance(output_tokens, int):
+                    usage_input["total_tokens"] = input_tokens + output_tokens
             response_api_usage = ResponseAPIUsage(**usage_input)
         else:
             response_api_usage = usage_input
@@ -1089,12 +1096,27 @@ class ResponseAPILoggingUtils:
                 audio_tokens=getattr(output_tokens_details, "audio_tokens", None),
             )
 
+        extra_usage_fields: Final = {
+            key: value
+            for key, value in (response_api_usage.model_extra or {}).items()
+            if key
+            not in (
+                "input_token_details",
+                "output_token_details",
+                "prompt_tokens",
+                "completion_tokens",
+                "total_tokens",
+                "prompt_tokens_details",
+                "completion_tokens_details",
+            )
+        }
         chat_usage: Final = Usage(
             prompt_tokens=prompt_tokens,
             completion_tokens=completion_tokens,
             total_tokens=prompt_tokens + completion_tokens,
             prompt_tokens_details=prompt_tokens_details,
             completion_tokens_details=completion_tokens_details,
+            **extra_usage_fields,
         )
 
         # Preserve cost attribute if it exists on ResponseAPIUsage

litellm/router.py

@@ -43,6 +43,7 @@ from litellm.caching.caching import (
     RedisClusterCache,
 )
 from litellm.constants import (
+    CONSUMED_REQUEST_TAGS_METADATA_KEY,
     DEFAULT_AUTO_ROUTER_MAX_INPUT_CHARS,
     DEFAULT_HEALTH_CHECK_INTERVAL,
     DEFAULT_HEALTH_CHECK_STALENESS_MULTIPLIER,
@@ -54,6 +55,7 @@ from litellm.litellm_core_utils.asyncify import run_async_function
 from litellm.litellm_core_utils.core_helpers import (
     _get_parent_otel_span_from_kwargs,
     coerce_token_limit,
+    get_litellm_metadata_from_kwargs,
     get_metadata_variable_name_from_kwargs,
     get_or_create_metadata_bucket,
 )
@@ -94,6 +96,7 @@ from litellm.router_utils.add_retry_fallback_headers import (
     response_in_flight_token_count,
 )
 from litellm.router_utils.auto_router_model_naming import (
+    AUTO_ROUTER_MODEL_PREFIX,
     classify_strategy_router_model,
 )
 from litellm.router_utils.batch_utils import (
@@ -170,6 +173,7 @@ from litellm.types.router import (
     AlertingConfig,
     AllowedFailsPolicy,
     AssistantsTypedDict,
+    ConsumedRequestTagsStamp,
     CredentialLiteLLMParams,
     CustomRoutingStrategyBase,
     Deployment,
@@ -315,6 +319,8 @@ def model_info_is_active_for_environment(model_info: Mapping[str, object] | None
 
 _PreRoutingStrategyT = TypeVar("_PreRoutingStrategyT")
 
+_ALIAS_PARAMS_NEVER_FORWARDED: Final = frozenset({"model", "api_base", "api_key", "api_version"})
+
 
 def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream]) -> bool:
     for chunk in chunks:
@@ -608,8 +614,10 @@ class Router:
         self.team_public_model_names: frozenset[str] = frozenset()
 
         # Initialize cache attributes that ``_invalidate_model_group_info_cache``
-        # touches *before* the first ``set_model_list`` below (which calls
-        # that invalidation as part of building the model index).
+        # and ``_invalidate_access_groups_cache`` touch *before* the first
+        # ``set_model_list`` below (which calls those invalidations as part of
+        # building the model index) and before ``_init_routing_groups(None)``
+        # (which calls them on every group rebuild).
         self._access_groups_cache: dict[str, list[str]] | None = None
         # Per-router cache for the proxy auth-layer "is this model explicitly
         # zero-cost?" check. Lives on the router so it is invalidated alongside
@@ -617,6 +625,8 @@ class Router:
         # ``id()``-reuse risk after GC). See
         # ``litellm.proxy.auth.auth_checks._is_model_cost_zero``.
         self._zero_cost_cache: dict[str, bool] = {}
+        self._routing_group_rows: tuple[DeploymentTypedDict, ...] | None = None
+        self._init_routing_groups(None)
 
         self.deployment_affinity_ttl_seconds = deployment_affinity_ttl_seconds
         self.model_group_affinity_config = model_group_affinity_config
@@ -1039,6 +1049,8 @@ class Router:
         self._routing_groups: dict[str, RoutingGroup] = {}
         self._model_to_group: dict[str, str] = {}
         self._group_selectors: dict[str, dict[str, RouterStrategySelector]] = {}
+        self._invalidate_model_group_info_cache()
+        self._invalidate_access_groups_cache()
 
         if not groups_input:
             return
@@ -1053,6 +1065,12 @@ class Router:
                 raise ValueError("routing_groups: group_name must be non-empty.")
             if group.group_name == "default":
                 raise ValueError("routing_groups: 'default' is reserved for the implicit fallback group.")
+            if group.group_name in known_model_names or group.group_name in (self.model_group_alias or {}):
+                verbose_router_logger.warning(
+                    "routing_groups: group_name '%s' is shadowed by an existing model_name or model_group_alias; "
+                    "the group's strategy still applies to its members, but the name is not callable until renamed.",
+                    group.group_name,
+                )
             if group.group_name in seen_group_names:
                 raise ValueError(
                     f"routing_groups: group names must be unique, duplicate group_name '{group.group_name}'."
@@ -1089,6 +1107,82 @@ class Router:
                 {strategy_value: group_selector} if group_selector is not None else {}
             )
 
+    def get_routing_group(self, model_name: str) -> RoutingGroup | None:
+        """
+        The routing group callable as `model_name`, or None. A real deployment
+        `model_name` added after init shadows a same-named group (mirroring
+        `_try_early_resolve_deployments_for_model_not_in_names`, where concrete
+        models win over indirection); config-time collisions are rejected by
+        `_init_routing_groups`.
+        """
+        if not self._routing_groups:
+            return None
+        group: Final = self._routing_groups.get(model_name)
+        if (
+            group is None
+            or model_name in self.model_name_to_deployment_indices
+            or model_name in (self.model_group_alias or {})
+        ):
+            return None
+        return group
+
+    def _get_routing_group_deployments(
+        self, model: str, team_id: str | None = None
+    ) -> list[DeploymentTypedDict] | None:  # mutable-ok: list matches _get_all_deployments' contract for callers
+        """
+        The union of member deployments for a routing group called as `model`,
+        or None when `model` is not a callable group. The requested name stays
+        the group name so strategy selectors key their state by it.
+
+        `_common_checks_available_deployment` consults this BEFORE its
+        early-resolve step so a wildcard `default_deployment` or pattern route
+        cannot hijack a group call. Overall resolution precedence there:
+        specific deployment > model id > model_group_alias > routing group >
+        model_name > team/pattern/default fallbacks.
+        """
+        if not self._routing_groups:
+            return None
+        routing_group: Final = self.get_routing_group(model)
+        if routing_group is None:
+            return None
+        return [  # mutable-ok: matches _get_all_deployments' list contract expected by downstream filters
+            deployment
+            for member in routing_group.models
+            for deployment in self._get_all_deployments(model_name=member, team_id=team_id)
+        ]
+
+    def is_recognized_model(self, model: str) -> bool:
+        """
+        Whether `model` names something this router serves directly: a
+        deployment model_name, a deployment id, a `model_group_alias`, or a
+        callable routing group. Proxy request gates share this predicate so a
+        new virtual-model kind cannot be forgotten at one of them; wildcard,
+        default-deployment, and deployment-name fallbacks stay caller policy.
+        """
+        return (
+            model in self.model_names
+            or self.has_model_id(model)
+            or (self.model_group_alias is not None and model in self.model_group_alias)
+            or self.get_routing_group(model) is not None
+        )
+
+    def routing_group_has_alternatives(self, model_group: str | None) -> bool:
+        """
+        True when `model_group` names a callable routing group whose member
+        union spans more than one deployment. Cooldown handling passes the
+        FAILING REQUEST's model group here: a 429 on a group call cools the
+        member down so selection moves to the group's alternatives, while a
+        direct call to a single-deployment member keeps the
+        single-deployment-model-group cooldown exemption.
+        """
+        if model_group is None:
+            return False
+        resolved: Final = self._get_model_from_alias(model=model_group) or model_group
+        group: Final = self.get_routing_group(resolved)
+        if group is None:
+            return False
+        return sum(len(self.model_name_to_deployment_indices.get(member) or ()) for member in group.models) > 1
+
     _OVERRIDABLE_ROUTING_STRATEGIES: frozenset[str] = frozenset({"simple-shuffle", *_DEFAULT_SELECTOR_ATTR_BY_STRATEGY})
 
     def _get_request_routing_strategy_override(self, request_kwargs: dict | None) -> str | None:
@@ -1149,8 +1243,10 @@ class Router:
         the most specific expression of caller intent.
 
         Otherwise every model belongs to exactly one group: an explicit entry
-        from `routing_groups`, or the implicit `"default"` group driven by the
-        router's top-level `routing_strategy` / `routing_strategy_args`.
+        from `routing_groups` (either because `model` IS a callable group name,
+        or because it is a member of one), or the implicit `"default"` group
+        driven by the router's top-level `routing_strategy` /
+        `routing_strategy_args`.
 
         `self.routing_strategy` may be either a string or a `RoutingStrategy`
         enum member (the constructor accepts both), so it is normalized to a
@@ -1162,7 +1258,7 @@ class Router:
             verbose_router_logger.debug("routing_group=request-override model=%s strategy=%s", model, override)
             return override, self._get_override_strategy_selector(override)
 
-        group_name: Final = self._model_to_group.get(model)
+        group_name: Final = model if self.get_routing_group(model) is not None else self._model_to_group.get(model)
         if group_name is None:
             strategy = self._normalize_strategy(self.routing_strategy)
             attr: Final = self._DEFAULT_SELECTOR_ATTR_BY_STRATEGY.get(strategy or "")
@@ -7143,6 +7239,7 @@ class Router:
                     original_exception=exception,
                     deployment=deployment_id,
                     time_to_cooldown=_time_to_cooldown,
+                    requested_model_group=(get_litellm_metadata_from_kwargs(kwargs) or {}).get("model_group"),
                 )  # setting deployment_id in cooldown deployments
 
                 return result
@@ -8326,6 +8423,7 @@ class Router:
                 self.model_name_to_deployment_indices[model_name] = updated_indices
             else:
                 del self.model_name_to_deployment_indices[model_name]
+                self.model_names.discard(model_name)
 
         # Update team_model_to_deployment_indices
         for key, indices in list(self.team_model_to_deployment_indices.items()):
@@ -8517,7 +8615,18 @@ class Router:
         Nothing is recorded for replay: a refresh walks the live routers instead,
         so a deleted, repointed or never-added deployment, and a discarded router,
         drop out of the rebuild on their own.
+
+        A strategy-router alias is never the deployment actually called or
+        billed, so custom pricing configured on it must not become a cost-map
+        price: an explicit zero would let ``_is_cost_explicitly_configured``
+        treat the alias as a genuinely free model and waive budget checks for
+        requests that route to (and bill as) a real deployment.
         """
+        if classify_strategy_router_model(model) is not None:
+            model_info = {  # mutable-ok: filtered copy of the caller's entry, handed straight to register_model
+                k: v for k, v in model_info.items() if k not in CustomPricingLiteLLMParams.model_fields
+            }
+
         if model_id is not None:
             litellm.register_model(model_cost={model_id: model_info}, persist_across_reloads=False)
 
@@ -9981,6 +10090,52 @@ class Router:
 
         return returned_models
 
+    def get_model_list_from_routing_groups(self, model_name: str | None = None) -> Sequence[DeploymentTypedDict]:
+        """
+        Callable routing groups materialized as model-list rows, mirroring
+        `get_model_list_from_model_alias`: each member deployment is emitted
+        under the group's name (via `_get_all_deployments`' `model_alias`
+        rewrite), which is what surfaces groups in `get_model_names`,
+        `/v1/models` discovery, `get_model_group_usage`, and the
+        blocked/unhealthy hiding that all read `get_model_list`.
+        """
+        if model_name is not None:
+            group: Final = self.get_routing_group(model_name)
+            return self._materialize_routing_group_rows((group,)) if group is not None else ()
+        cached: Final = self._routing_group_rows
+        if cached is not None:
+            return cached
+        rows: Final = self._materialize_routing_group_rows(
+            tuple(
+                callable_group
+                for name in self._routing_groups
+                if (callable_group := self.get_routing_group(name)) is not None
+            )
+        )
+        self._routing_group_rows = rows
+        return rows
+
+    def _materialize_routing_group_rows(self, groups: tuple[RoutingGroup, ...]) -> tuple[DeploymentTypedDict, ...]:
+        return tuple(
+            self._as_routing_group_row(deployment)
+            for group in groups
+            for member in group.models
+            for deployment in self._get_all_deployments(model_name=member, model_alias=group.group_name)
+        )
+
+    @staticmethod
+    def _as_routing_group_row(deployment: DeploymentTypedDict) -> DeploymentTypedDict:
+        """
+        A member deployment re-emitted under its group's name must not carry
+        the member's `access_groups`: access groups grant member names, never
+        the group, so inheriting them here would let a key holding a member's
+        access group list and call the whole group.
+        """
+        model_info: Final = {  # mutable-ok: DeploymentTypedDict rows are plain dicts
+            k: v for k, v in (deployment.get("model_info") or {}).items() if k != "access_groups"
+        }
+        return {**deployment, "model_info": model_info}  # mutable-ok: DeploymentTypedDict rows are plain dicts
+
     def get_model_list(
         self, model_name: str | None = None, team_id: str | None = None
     ) -> list[DeploymentTypedDict] | None:
@@ -9997,6 +10152,7 @@ class Router:
             returned_models.extend(self._get_all_deployments(model_name=model_name, team_id=team_id))
 
         returned_models.extend(self.get_model_list_from_model_alias(model_name=model_name))
+        returned_models.extend(self.get_model_list_from_routing_groups(model_name=model_name))
 
         if len(returned_models) == 0:  # check if wildcard route
             potential_wildcard_models: Final = self.pattern_router.route(model_name) or []
@@ -10028,6 +10184,7 @@ class Router:
         """
         self._cached_get_model_group_info.cache_clear()
         self._zero_cost_cache.clear()
+        self._routing_group_rows = None
 
     def _invalidate_access_groups_cache(self) -> None:
         """Invalidate the cached access groups.
@@ -10558,6 +10715,14 @@ class Router:
 
         return None
 
+    @staticmethod
+    def _is_strategy_marker_deployment(deployment: Mapping[str, object]) -> bool:
+        litellm_params: Final = deployment.get("litellm_params")
+        if not isinstance(litellm_params, Mapping):
+            return False
+        deployment_model: Final = litellm_params.get("model")
+        return isinstance(deployment_model, str) and classify_strategy_router_model(deployment_model) is not None
+
     def _common_checks_available_deployment(
         self,
         model: str,
@@ -10598,17 +10763,23 @@ class Router:
         if _model_from_alias is not None:
             model = _model_from_alias
 
-        early: Final = self._try_early_resolve_deployments_for_model_not_in_names(
-            model=model,
-            request_team_id=request_team_id,
-            include_team_models=_is_proxy_admin_request(request_kwargs),
-        )
-        if early is not None:
-            return early
+        _routing_group_deployments: Final = self._get_routing_group_deployments(model=model, team_id=request_team_id)
+        if _routing_group_deployments is None:
+            early: Final = self._try_early_resolve_deployments_for_model_not_in_names(
+                model=model,
+                request_team_id=request_team_id,
+                include_team_models=_is_proxy_admin_request(request_kwargs),
+            )
+            if early is not None:
+                return early
 
         ## get healthy deployments
         ### get all deployments
-        healthy_deployments = self._get_all_deployments(model_name=model, team_id=request_team_id)
+        healthy_deployments = (
+            _routing_group_deployments
+            if _routing_group_deployments is not None
+            else self._get_all_deployments(model_name=model, team_id=request_team_id)
+        )
         _pre_model_access_group_filter_len: Final = len(healthy_deployments)
         healthy_deployments = self._filter_deployments_by_model_access_groups(
             model=model,
@@ -10679,7 +10850,12 @@ class Router:
                 model
             ]  # update the model to the actual value if an alias has been passed in
 
-        return model, healthy_deployments
+        marker_flags: Final = tuple(self._is_strategy_marker_deployment(d) for d in healthy_deployments)
+        if all(marker_flags) or not any(marker_flags):
+            return model, healthy_deployments
+        return model, [  # mutable-ok: matches this function's list contract expected by downstream filters
+            d for d, is_marker in zip(healthy_deployments, marker_flags, strict=True) if not is_marker
+        ]
 
     def _filter_deployments_by_model_access_groups(
         self,
@@ -11192,11 +11368,26 @@ class Router:
 
         return filtered
 
-    def _select_pre_routing_strategy(self, model: str, request_kwargs: dict) -> "PreRoutingStrategy | None":
+    def _model_name_has_plain_deployments(self, model: str) -> bool:
+        indices: Final = self.model_name_to_deployment_indices.get(model) or ()
+        return any(not self._is_strategy_marker_deployment(self.model_list[idx]) for idx in indices)
+
+    def _select_pre_routing_strategy(
+        self, model: str, request_kwargs: dict
+    ) -> "TaggedPreRoutingStrategy[PreRoutingStrategy] | None":
         """
         Resolve the pre-routing strategy for `model`, disambiguating deployments
         that share a `model_name` by matching the request's tags against each
         registered strategy's tags before falling back to the first registered.
+        Returns the tagged registry entry so the caller can tell whether the
+        request's tags were what selected it, and can locate the marker
+        deployment the strategy was registered from via its (model_name, tags)
+        pair.
+
+        With tag filtering enabled, strategies that all carry real tags matching
+        none of the request's do not capture it when the name also has plain
+        deployments: returning None hands the request to ordinary tag-aware
+        deployment selection.
         """
         candidates: Final[list[TaggedPreRoutingStrategy[PreRoutingStrategy]]] = [
             *self.auto_routers.get(model, []),
@@ -11206,8 +11397,6 @@ class Router:
         ]
         if not candidates:
             return None
-        if len(candidates) == 1:
-            return candidates[0].strategy
 
         request_tags: Final = _get_tags_from_request_kwargs(request_kwargs)
         if request_tags:
@@ -11215,11 +11404,17 @@ class Router:
                 if tagged.tags and is_valid_deployment_tag(
                     list(tagged.tags), request_tags, self.tag_filtering_match_any
                 ):
-                    return tagged.strategy
+                    return tagged
         for tagged in candidates:
             if "default" in tagged.tags:
-                return tagged.strategy
-        return candidates[0].strategy
+                return tagged
+        if (
+            self.enable_tag_filtering
+            and all(tagged.tags for tagged in candidates)
+            and self._model_name_has_plain_deployments(model)
+        ):
+            return None
+        return candidates[0]
 
     async def async_pre_routing_hook(
         self,
@@ -11243,15 +11438,18 @@ class Router:
         if self.routing_plugins:
             await self._run_routing_plugins(model=model, request_kwargs=request_kwargs, messages=messages)
 
-        router_strategy: Final = self._select_pre_routing_strategy(model=model, request_kwargs=request_kwargs)
-        if router_strategy is None:
+        selected_strategy: Final = self._select_pre_routing_strategy(model=model, request_kwargs=request_kwargs)
+        if selected_strategy is None:
             self._record_routing_decision(request_kwargs=request_kwargs, routing_decision=None)
             self._stamp_or_clear_metadata_key(
                 request_kwargs=request_kwargs, key=SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY, value=None
             )
+            self._stamp_or_clear_metadata_key(
+                request_kwargs=request_kwargs, key=CONSUMED_REQUEST_TAGS_METADATA_KEY, value=None
+            )
             return None
 
-        pre_routing_hook_response: Final = await router_strategy.async_pre_routing_hook(
+        pre_routing_hook_response: Final = await selected_strategy.strategy.async_pre_routing_hook(
             model=model,
             request_kwargs=request_kwargs,
             messages=messages,
@@ -11267,24 +11465,80 @@ class Router:
             key=SESSION_DEPLOYMENT_AFFINITY_TTL_METADATA_KEY,
             value=(pre_routing_hook_response.session_affinity_ttl_seconds if pre_routing_hook_response else None),
         )
+        self._stamp_or_clear_metadata_key(
+            request_kwargs=request_kwargs,
+            key=CONSUMED_REQUEST_TAGS_METADATA_KEY,
+            value=self._consumed_request_tags_stamp(
+                selected_strategy=selected_strategy,
+                pre_routing_hook_response=pre_routing_hook_response,
+                request_tags=_get_tags_from_request_kwargs(request_kwargs),
+            ),
+        )
 
         # `model` (the alias, e.g. "smart-router") is never the deployment actually
-        # called - apply the alias's own litellm_params (besides `model` itself,
-        # which is just the alias marker) to the request, since the tier/route
-        # deployment the hook selected won't have them. Router-only fields
-        # (tpm, rpm, weight, complexity_router_config, ...) are excluded from the
-        # actual outbound LLM call downstream by litellm.types.utils.all_litellm_params,
-        # not here.
+        # called - apply the router marker's own litellm_params to the request,
+        # since the tier/route deployment the hook selected won't have them. The
+        # marker entry is looked up by its `auto_router/` model prefix and the
+        # selected strategy's tags, never by list position: plain deployments may
+        # share the alias `model_name` and must not leak their params (`api_base`,
+        # `api_key`, ...) onto the routed call. Router-only fields (tpm, rpm,
+        # weight, complexity_router_config, ...) are excluded from the actual
+        # outbound LLM call downstream by litellm.types.utils.all_litellm_params,
+        # not here. Custom pricing fields ARE call params, so they must be
+        # excluded here: they price the alias, not the deployment the hook
+        # selected, and forwarding them re-registers the routed deployment at
+        # the alias's price (an explicit 0 makes every alias request bill $0).
         if pre_routing_hook_response is not None:
-            alias_index: Final = self.model_name_to_deployment_indices.get(model, [])
-            if alias_index:
-                alias_litellm_params: Final = self.model_list[alias_index[0]].get("litellm_params", {})
-                for key, value in alias_litellm_params.items():
-                    if key != "model" and value is not None:
-                        request_kwargs.setdefault(key, value)
+            for key, value in self._forwardable_alias_marker_params(model=model, strategy_tags=selected_strategy.tags):
+                request_kwargs.setdefault(key, value)
 
         return pre_routing_hook_response
 
+    def _forwardable_alias_marker_params(
+        self, model: str, strategy_tags: tuple[str, ...]
+    ) -> tuple[tuple[str, object], ...]:
+        marker_params: Final = tuple(
+            litellm_params
+            for idx in self.model_name_to_deployment_indices.get(model, ())
+            if isinstance(litellm_params := self.model_list[idx].get("litellm_params", {}), dict)
+            and str(litellm_params.get("model", "")).startswith(AUTO_ROUTER_MODEL_PREFIX)
+        )
+        tag_matched: Final = tuple(
+            params for params in marker_params if tuple(params.get("tags") or ()) == strategy_tags
+        )
+        selected: Final = tag_matched[0] if tag_matched else (marker_params[0] if marker_params else None)
+        if selected is None:
+            return ()
+        return tuple(
+            (key, value)
+            for key, value in selected.items()
+            if key not in _ALIAS_PARAMS_NEVER_FORWARDED
+            and key not in CustomPricingLiteLLMParams.model_fields
+            and value is not None
+        )
+
+    def _consumed_request_tags_stamp(
+        self,
+        selected_strategy: "TaggedPreRoutingStrategy[PreRoutingStrategy]",
+        pre_routing_hook_response: PreRoutingHookResponse | None,
+        request_tags: Sequence[str],
+    ) -> ConsumedRequestTagsStamp | None:
+        """Record which tags picked the router and which model group it rewrote to, or None.
+
+        A request whose tags matched the selected strategy's tags has already spent those
+        tags on picking the router; re-applying them to the routed tier's model group would
+        empty the pool unless every tier deployment repeats the marker's tag. Only the
+        strategy's own tags are spent: the request's other tags keep constraining
+        deployment selection inside the routed group, and key/team policy tags are
+        untouched because tag filtering separately re-applies whatever
+        `metadata.inherited_tags` carries for the stamped group.
+        """
+        if pre_routing_hook_response is None or not selected_strategy.tags or not request_tags:
+            return None
+        if not is_valid_deployment_tag(selected_strategy.tags, request_tags, self.tag_filtering_match_any):
+            return None
+        return ConsumedRequestTagsStamp(model_group=pre_routing_hook_response.model, tags=selected_strategy.tags)
+
     @staticmethod
     def _record_routing_decision(
         request_kwargs: dict,

litellm/router_strategy/tag_based_routing.py

@@ -13,7 +13,9 @@ from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal
 
 from litellm._logging import verbose_logger
-from litellm.types.router import RouterErrors
+from litellm.constants import CONSUMED_REQUEST_TAGS_METADATA_KEY
+from litellm.litellm_core_utils.core_helpers import get_metadata_variable_name_from_kwargs
+from litellm.types.router import ConsumedRequestTagsStamp, RouterErrors
 
 if TYPE_CHECKING:
     from litellm.router import Router as _Router
@@ -46,7 +48,9 @@ def _is_valid_deployment_tag_regex(
     return None
 
 
-def is_valid_deployment_tag(deployment_tags: list[str], request_tags: list[str], match_any: bool = True) -> bool:
+def is_valid_deployment_tag(
+    deployment_tags: Sequence[str], request_tags: Sequence[str], match_any: bool = True
+) -> bool:
     """
     Check if a tag is valid, the matching can be either any or all based on `match_any` flag
     """
@@ -389,6 +393,25 @@ def _tag_known_to_group(
     )
 
 
+def _request_tags_after_router_consumption(metadata: Mapping[Any, Any], model: str) -> Sequence[str] | None:
+    # The pre-routing hook stamps which tags selected the router it rewrote the request
+    # to: those tags already did their job and must not also constrain deployment choice
+    # inside the routed group. The request's other tags still apply there, on top of the
+    # inherited_tags snapshot that keeps key/team policy applying. Every other model
+    # group keeps the full list.
+    stamp: Final = metadata.get(CONSUMED_REQUEST_TAGS_METADATA_KEY)
+    if not isinstance(stamp, ConsumedRequestTagsStamp) or stamp.model_group != model:
+        return metadata.get("tags")
+    request_tags: Final = metadata.get("tags")
+    leftover: Final = tuple(
+        tag for tag in (request_tags if isinstance(request_tags, (list, tuple)) else ()) if tag not in stamp.tags
+    )
+    inherited_tags: Final = metadata.get("inherited_tags")
+    if not isinstance(inherited_tags, (list, tuple)):
+        return leftover or None
+    return tuple(dict.fromkeys((*leftover, *inherited_tags)))
+
+
 async def get_deployments_for_tag(
     llm_router_instance: LitellmRouter,
     model: str,  # used to raise the correct error
@@ -429,7 +452,7 @@ async def get_deployments_for_tag(
     verbose_logger.debug("request metadata: %s", request_kwargs.get(metadata_variable_name))
     if metadata_variable_name in request_kwargs:
         metadata: Final = request_kwargs[metadata_variable_name]
-        request_tags: Final = metadata.get("tags")
+        request_tags: Final = _request_tags_after_router_consumption(metadata, model)
         match_any: Final = llm_router_instance.tag_filtering_match_any
         routing_prefix: Final = llm_router_instance.tag_routing_prefix or ""
 
@@ -561,28 +584,49 @@ async def get_deployments_for_tag(
     return healthy_deployments
 
 
+def _tags_in_metadata(metadata: object) -> list[str]:
+    """
+    Tags out of a metadata bucket the caller controls the shape of.
+
+    A request can send its metadata (and its ``tags``) as anything the JSON body
+    allowed, an unparsed string or null included, so any shape that is not a list
+    of string tags carries no tags rather than raising.
+    """
+    if not isinstance(metadata, Mapping):
+        return []
+    typed_metadata: Final[Mapping[str, object]] = metadata
+    tags: Final = typed_metadata.get("tags")
+    if isinstance(tags, str) or not isinstance(tags, Sequence):
+        return []
+    typed_tags: Final[Sequence[object]] = tags
+    return [tag for tag in typed_tags if isinstance(tag, str)]
+
+
 def _get_tags_from_request_kwargs(
-    request_kwargs: dict[Any, Any] | None = None,
-    metadata_variable_name: Literal["metadata", "litellm_metadata"] = "metadata",
+    request_kwargs: Mapping[Any, Any] | None = None,
+    metadata_variable_name: Literal["metadata", "litellm_metadata"] | None = None,
 ) -> list[str]:
     """
     Helper to get tags from request kwargs
 
     Args:
         request_kwargs: The request kwargs to get tags from
+        metadata_variable_name: Which metadata dict holds proxy metadata; resolved
+            from the kwargs when not pinned, so /v1/messages-shaped requests
+            (``litellm_metadata``) read the same bucket the proxy wrote tags to
 
     Returns:
         List[str]: The tags from the request kwargs
     """
     if request_kwargs is None:
         return []
-    if metadata_variable_name in request_kwargs:
-        metadata: Final = request_kwargs[metadata_variable_name] or {}
-        tags = metadata.get("tags", [])
-        return tags if tags is not None else []
-    elif "litellm_params" in request_kwargs:
-        litellm_params: Final = request_kwargs["litellm_params"] or {}
-        _metadata: Final = litellm_params.get(metadata_variable_name, {}) or {}
-        tags = _metadata.get("tags", [])
-        return tags if tags is not None else []
+    resolved_variable_name: Final = metadata_variable_name or get_metadata_variable_name_from_kwargs(request_kwargs)
+    if resolved_variable_name in request_kwargs:
+        return _tags_in_metadata(request_kwargs[resolved_variable_name])
+    if "litellm_params" in request_kwargs:
+        litellm_params: Final = request_kwargs["litellm_params"]
+        if not isinstance(litellm_params, Mapping):
+            return []
+        typed_litellm_params: Final[Mapping[str, object]] = litellm_params
+        return _tags_in_metadata(typed_litellm_params.get(resolved_variable_name))
     return []

litellm/router_utils/cooldown_handlers.py

@@ -319,6 +319,7 @@ def _should_cooldown_deployment(
     deployment: str,
     exception_status: str | int,
     original_exception: Any,
+    requested_model_group: str | None = None,
 ) -> bool:
     """
     Helper that decides if a deployment should be put in cooldown
@@ -341,7 +342,9 @@ def _should_cooldown_deployment(
     model_group: Final = litellm_router_instance.get_model_group(id=deployment)
     is_single_deployment_model_group = False
     if model_group is not None and len(model_group) == 1:
-        is_single_deployment_model_group = True
+        is_single_deployment_model_group = not litellm_router_instance.routing_group_has_alternatives(
+            requested_model_group
+        )
 
     ## CHECK DEPLOYMENT-LEVEL POLICY FIRST (overrides router-level)
     dep_policy, dep_allowed_fails = _get_deployment_cooldown_policy(litellm_router_instance, deployment)
@@ -413,6 +416,7 @@ def _set_cooldown_deployments(
     exception_status: str | int,
     deployment: str | None = None,
     time_to_cooldown: float | None = None,
+    requested_model_group: str | None = None,
 ) -> bool:
     """
     Add a model to the list of models being cooled down for that minute, if it exceeds the allowed fails / minute
@@ -449,6 +453,7 @@ def _set_cooldown_deployments(
         deployment=deployment,
         exception_status=exception_status,
         original_exception=original_exception,
+        requested_model_group=requested_model_group,
     ):
         litellm_router_instance.cooldown_cache.add_deployment_to_cooldown(
             model_id=deployment,

litellm/types/integrations/langfuse_otel.py

@@ -16,12 +16,13 @@ class LangfuseOtelConfig(BaseModel):
 
 class LangfuseSpanAttributes(str, Enum):
     LANGFUSE_ENVIRONMENT = "langfuse.environment"
+    VERSION = "langfuse.version"
+    RELEASE = "langfuse.release"
 
     # ---- Generation-level metadata ----
     GENERATION_NAME = "langfuse.generation.name"
     GENERATION_ID = "langfuse.generation.id"
     PARENT_OBSERVATION_ID = "langfuse.generation.parent_observation_id"
-    GENERATION_VERSION = "langfuse.generation.version"
     MASK_INPUT = "langfuse.generation.mask_input"
     MASK_OUTPUT = "langfuse.generation.mask_output"
 
@@ -36,8 +37,6 @@ class LangfuseSpanAttributes(str, Enum):
     TRACE_NAME = "langfuse.trace.name"
     TRACE_ID = "langfuse.trace.id"
     TRACE_METADATA = "langfuse.trace.metadata"
-    TRACE_VERSION = "langfuse.trace.version"
-    TRACE_RELEASE = "langfuse.trace.release"
     EXISTING_TRACE_ID = "langfuse.trace.existing_id"
     UPDATE_TRACE_KEYS = "langfuse.trace.update_keys"
 

litellm/types/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -13,3 +13,5 @@ class UsagePerChunk(TypedDict):
     completion_tokens_details: CompletionTokensDetails | None
     prompt_tokens_details: PromptTokensDetailsWrapper | None
     cost: float | None
+    inference_geo: str | None
+    speed: str | None

litellm/types/router.py

@@ -902,6 +902,14 @@ class TaggedPreRoutingStrategy(Generic[_PreRoutingStrategyT_co]):
     strategy: _PreRoutingStrategyT_co
 
 
+@dataclass(frozen=True, slots=True)
+class ConsumedRequestTagsStamp:
+    """The model group a tagged router rewrote to, plus the request tags spent selecting it."""
+
+    model_group: str
+    tags: tuple[str, ...]
+
+
 @runtime_checkable
 class PreRoutingStrategy(Protocol):
     """Structural interface shared by the auto / complexity / adaptive / quality routers."""

model_prices_and_context_window.json

@@ -9707,6 +9707,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -9830,6 +9831,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -9922,6 +9924,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10007,6 +10010,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10410,6 +10414,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10624,6 +10629,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -10701,6 +10707,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -11262,6 +11269,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -12940,6 +12948,103 @@
         "supports_system_messages": true,
         "supports_tool_choice": false
     },
+    "dashscope/deepseek-v4-flash": {
+        "cache_read_input_token_cost": 4e-08,
+        "input_cost_per_token": 2e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4e-07,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/deepseek-v4-flash-0731": {
+        "cache_read_input_token_cost": 4e-08,
+        "input_cost_per_token": 2e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4e-07,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/deepseek-v4-pro": {
+        "cache_read_input_token_cost": 2e-07,
+        "input_cost_per_token": 2.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1000000,
+        "max_output_tokens": 393216,
+        "max_tokens": 393216,
+        "mode": "chat",
+        "output_cost_per_token": 4.8e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/glm-5.1": {
+        "cache_read_input_token_cost": 2.6e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 202745,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/glm-5.2": {
+        "cache_read_input_token_cost": 2.8e-07,
+        "input_cost_per_token": 1.4e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 4.4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true
+    },
+    "dashscope/kimi-k2.7-code": {
+        "cache_read_input_token_cost": 1.9e-07,
+        "input_cost_per_token": 9.5e-07,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 229376,
+        "max_output_tokens": 16384,
+        "max_tokens": 16384,
+        "mode": "chat",
+        "output_cost_per_token": 4e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "dashscope/qwen-coder": {
         "input_cost_per_token": 3e-07,
         "litellm_provider": "dashscope",
@@ -13733,6 +13838,23 @@
             }
         ]
     },
+    "dashscope/qwen3.8-max": {
+        "cache_read_input_token_cost": 2.5e-07,
+        "input_cost_per_token": 2e-06,
+        "litellm_provider": "dashscope",
+        "max_input_tokens": 991808,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
+        "mode": "chat",
+        "output_cost_per_token": 6e-06,
+        "source": "https://www.alibabacloud.com/help/en/model-studio/models",
+        "supports_function_calling": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "dashscope/qwq-plus": {
         "input_cost_per_token": 8e-07,
         "litellm_provider": "dashscope",
@@ -15430,6 +15552,17 @@
         "supports_tool_choice": true,
         "supports_function_calling": true
     },
+    "deepinfra/nvidia/NVIDIA-Nemotron-3.5-Lightning": {
+        "max_input_tokens": 262144,
+        "input_cost_per_token": 5e-08,
+        "output_cost_per_token": 2e-07,
+        "litellm_provider": "deepinfra",
+        "mode": "chat",
+        "source": "https://deepinfra.com/nvidia/NVIDIA-Nemotron-3.5-Lightning",
+        "supports_tool_choice": true,
+        "supports_function_calling": true,
+        "supports_reasoning": true
+    },
     "deepinfra/nvidia/NVIDIA-Nemotron-Nano-9B-v2": {
         "max_tokens": 131072,
         "max_input_tokens": 131072,
@@ -18888,6 +19021,60 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "vertex_ai/gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "vertex_ai",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "vertex_ai/gemini-3.1-pro-preview": {
         "cache_read_input_token_cost": 2e-07,
         "cache_read_input_token_cost_above_200k_tokens": 4e-07,
@@ -20563,6 +20750,63 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "gemini/gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "gemini",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "rpm": 2000,
+        "source": "https://ai.google.dev/pricing/gemini-3",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_output": false,
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "tpm": 800000,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "gemini/gemini-omni-flash-preview": {
         "input_cost_per_audio_token": 1.5e-06,
         "input_cost_per_token": 1.5e-06,
@@ -20898,6 +21142,61 @@
         },
         "web_search_billing_unit": "per_query"
     },
+    "gemini-3.7-flash": {
+        "cache_read_input_token_cost": 7.5e-08,
+        "cache_read_input_token_cost_flex": 3.75e-08,
+        "input_cost_per_token": 7.5e-07,
+        "input_cost_per_token_batches": 3.75e-07,
+        "input_cost_per_token_flex": 3.75e-07,
+        "litellm_provider": "vertex_ai-language-models",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
+        "mode": "chat",
+        "output_cost_per_reasoning_token": 3.75e-06,
+        "output_cost_per_token": 3.75e-06,
+        "output_cost_per_token_batches": 1.875e-06,
+        "output_cost_per_token_flex": 1.875e-06,
+        "source": "https://ai.google.dev/pricing/gemini-3",
+        "supported_endpoints": [
+            "/v1/chat/completions",
+            "/v1/completions",
+            "/v1/batch"
+        ],
+        "supported_modalities": [
+            "text",
+            "image",
+            "audio",
+            "video"
+        ],
+        "supported_output_modalities": [
+            "text"
+        ],
+        "supports_audio_output": false,
+        "supports_audio_input": true,
+        "supports_function_calling": true,
+        "supports_parallel_function_calling": true,
+        "supports_pdf_input": true,
+        "supports_prompt_caching": true,
+        "supports_reasoning": true,
+        "supports_response_schema": true,
+        "supports_system_messages": true,
+        "supports_tool_choice": true,
+        "supports_url_context": true,
+        "supports_video_input": true,
+        "supports_vision": true,
+        "supports_web_search": true,
+        "supports_native_streaming": true,
+        "input_cost_per_token_priority": 1.35e-06,
+        "output_cost_per_token_priority": 6.75e-06,
+        "cache_read_input_token_cost_priority": 1.35e-07,
+        "search_context_cost_per_query": {
+            "search_context_size_low": 0.014,
+            "search_context_size_medium": 0.014,
+            "search_context_size_high": 0.014
+        },
+        "web_search_billing_unit": "per_query"
+    },
     "gemini/gemini-2.5-pro-preview-tts": {
         "cache_read_input_token_cost": 1.25e-07,
         "cache_read_input_token_cost_above_200k_tokens": 2.5e-07,
@@ -25987,11 +26286,12 @@
         "supports_vision": true
     },
     "groq/llama-3.1-8b-instant": {
+        "deprecation_date": "2026-08-16",
         "input_cost_per_token": 5e-08,
         "litellm_provider": "groq",
-        "max_input_tokens": 128000,
-        "max_output_tokens": 8192,
-        "max_tokens": 8192,
+        "max_input_tokens": 131072,
+        "max_output_tokens": 131072,
+        "max_tokens": 131072,
         "mode": "chat",
         "output_cost_per_token": 8e-08,
         "supports_function_calling": true,
@@ -25999,9 +26299,10 @@
         "supports_tool_choice": true
     },
     "groq/llama-3.3-70b-versatile": {
+        "deprecation_date": "2026-08-16",
         "input_cost_per_token": 5.9e-07,
         "litellm_provider": "groq",
-        "max_input_tokens": 128000,
+        "max_input_tokens": 131072,
         "max_output_tokens": 32768,
         "max_tokens": 32768,
         "mode": "chat",
@@ -26022,7 +26323,28 @@
         "supports_response_schema": false,
         "supports_tool_choice": true
     },
+    "groq/meta-llama/llama-prompt-guard-2-22m": {
+        "input_cost_per_token": 3e-08,
+        "litellm_provider": "groq",
+        "max_input_tokens": 512,
+        "max_output_tokens": 512,
+        "max_tokens": 512,
+        "mode": "chat",
+        "output_cost_per_token": 3e-08,
+        "source": "https://console.groq.com/docs/models"
+    },
+    "groq/meta-llama/llama-prompt-guard-2-86m": {
+        "input_cost_per_token": 4e-08,
+        "litellm_provider": "groq",
+        "max_input_tokens": 512,
+        "max_output_tokens": 512,
+        "max_tokens": 512,
+        "mode": "chat",
+        "output_cost_per_token": 4e-08,
+        "source": "https://console.groq.com/docs/model/meta-llama/llama-prompt-guard-2-86m"
+    },
     "groq/meta-llama/llama-guard-4-12b": {
+        "deprecation_date": "2026-03-05",
         "input_cost_per_token": 2e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 8192,
@@ -26032,6 +26354,7 @@
         "output_cost_per_token": 2e-07
     },
     "groq/meta-llama/llama-4-maverick-17b-128e-instruct": {
+        "deprecation_date": "2026-03-09",
         "input_cost_per_token": 2e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
@@ -26045,6 +26368,7 @@
         "supports_vision": true
     },
     "groq/meta-llama/llama-4-scout-17b-16e-instruct": {
+        "deprecation_date": "2026-07-17",
         "input_cost_per_token": 1.1e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
@@ -26058,6 +26382,7 @@
         "supports_vision": true
     },
     "groq/moonshotai/kimi-k2-instruct-0905": {
+        "deprecation_date": "2026-04-15",
         "input_cost_per_token": 1e-06,
         "output_cost_per_token": 3e-06,
         "cache_read_input_token_cost": 5e-07,
@@ -26075,8 +26400,8 @@
         "input_cost_per_token": 1.5e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
-        "max_output_tokens": 32766,
-        "max_tokens": 32766,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
         "mode": "chat",
         "output_cost_per_token": 6e-07,
         "search_context_cost_per_query": {
@@ -26096,8 +26421,8 @@
         "input_cost_per_token": 7.5e-08,
         "litellm_provider": "groq",
         "max_input_tokens": 131072,
-        "max_output_tokens": 32768,
-        "max_tokens": 32768,
+        "max_output_tokens": 65536,
+        "max_tokens": 65536,
         "mode": "chat",
         "output_cost_per_token": 3e-07,
         "search_context_cost_per_query": {
@@ -26132,7 +26457,26 @@
         "supports_tool_choice": true,
         "supports_web_search": true
     },
+    "groq/canopylabs/orpheus-v1-english": {
+        "input_cost_per_character": 2.2e-05,
+        "litellm_provider": "groq",
+        "max_input_tokens": 4000,
+        "max_output_tokens": 50000,
+        "max_tokens": 50000,
+        "mode": "audio_speech",
+        "source": "https://console.groq.com/docs/model/canopylabs/orpheus-v1-english"
+    },
+    "groq/canopylabs/orpheus-arabic-saudi": {
+        "input_cost_per_character": 4e-05,
+        "litellm_provider": "groq",
+        "max_input_tokens": 4000,
+        "max_output_tokens": 50000,
+        "max_tokens": 50000,
+        "mode": "audio_speech",
+        "source": "https://console.groq.com/docs/models"
+    },
     "groq/playai-tts": {
+        "deprecation_date": "2025-12-31",
         "input_cost_per_character": 5e-05,
         "litellm_provider": "groq",
         "max_input_tokens": 10000,
@@ -26140,7 +26484,23 @@
         "max_tokens": 10000,
         "mode": "audio_speech"
     },
+    "groq/qwen/qwen3.6-27b": {
+        "input_cost_per_token": 6e-07,
+        "litellm_provider": "groq",
+        "max_input_tokens": 131072,
+        "max_output_tokens": 16384,
+        "max_tokens": 16384,
+        "mode": "chat",
+        "output_cost_per_token": 3e-06,
+        "source": "https://console.groq.com/docs/model/qwen/qwen3.6-27b",
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_response_schema": false,
+        "supports_tool_choice": true,
+        "supports_vision": true
+    },
     "groq/qwen/qwen3-32b": {
+        "deprecation_date": "2026-07-17",
         "input_cost_per_token": 2.9e-07,
         "litellm_provider": "groq",
         "max_input_tokens": 131000,
@@ -31632,6 +31992,17 @@
         "supports_video_input": true,
         "supports_vision": true
     },
+    "openrouter/nvidia/nemotron-3.5-lightning": {
+        "input_cost_per_token": 5e-08,
+        "litellm_provider": "openrouter",
+        "max_input_tokens": 262144,
+        "mode": "chat",
+        "output_cost_per_token": 2e-07,
+        "source": "https://openrouter.ai/nvidia/nemotron-3.5-lightning",
+        "supports_function_calling": true,
+        "supports_reasoning": true,
+        "supports_tool_choice": true
+    },
     "openrouter/openai/gpt-3.5-turbo": {
         "input_cost_per_token": 1.5e-06,
         "litellm_provider": "openrouter",
@@ -35769,6 +36140,7 @@
         "output_cost_per_token": 1.85e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true
     },
     "eu.deepseek.v3.2": {
@@ -35781,6 +36153,7 @@
         "output_cost_per_token": 2.22e-06,
         "supports_function_calling": true,
         "supports_reasoning": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true
     },
     "us.meta.llama3-1-405b-instruct-v1:0": {
@@ -45680,11 +46053,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-sol": {
         "input_cost_per_token": 5.5e-06,
+        "input_cost_per_token_above_272k_tokens": 1.1e-05,
         "cache_creation_input_token_cost": 6.875e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 1.375e-05,
         "cache_read_input_token_cost": 5.5e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 1.1e-06,
         "output_cost_per_token": 3.3e-05,
+        "output_cost_per_token_above_272k_tokens": 4.95e-05,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -45708,11 +46085,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-terra": {
         "input_cost_per_token": 2.2e-06,
+        "input_cost_per_token_above_272k_tokens": 4.4e-06,
         "cache_creation_input_token_cost": 2.75e-06,
+        "cache_creation_input_token_cost_above_272k_tokens": 5.5e-06,
         "cache_read_input_token_cost": 2.2e-07,
+        "cache_read_input_token_cost_above_272k_tokens": 4.4e-07,
         "output_cost_per_token": 1.32e-05,
+        "output_cost_per_token_above_272k_tokens": 1.98e-05,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -45736,11 +46117,15 @@
     },
     "bedrock_mantle/openai.gpt-5.6-luna": {
         "input_cost_per_token": 2.2e-07,
+        "input_cost_per_token_above_272k_tokens": 4.4e-07,
         "cache_creation_input_token_cost": 2.75e-07,
+        "cache_creation_input_token_cost_above_272k_tokens": 5.5e-07,
         "cache_read_input_token_cost": 2.2e-08,
+        "cache_read_input_token_cost_above_272k_tokens": 4.4e-08,
         "output_cost_per_token": 1.32e-06,
+        "output_cost_per_token_above_272k_tokens": 1.98e-06,
         "litellm_provider": "bedrock_mantle",
-        "max_input_tokens": 272000,
+        "max_input_tokens": 1000000,
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "responses",
@@ -46057,6 +46442,7 @@
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_system_messages": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },
@@ -46071,6 +46457,7 @@
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_system_messages": true,
+        "supports_native_structured_output": true,
         "supports_tool_choice": true,
         "source": "https://aws.amazon.com/bedrock/pricing/"
     },

scripts/check_type_discipline.py

@@ -29,8 +29,8 @@ LIT003  noqa suppression without rule codes or without a reason.
         Required shape: `# noqa: TID251  # <reason>`
 LIT004  pyright/mypy ignore without bracketed codes or without a reason.
         Required shape: `# pyright: ignore[reportArgumentType]  # <reason>`
-LIT005  A `# mutable-ok` / `# cast-ok` / `# guard-ok` / `# kwargs-ok`
-        suppression without a reason.
+LIT005  A `# mutable-ok` / `# cast-ok` / `# guard-ok` / `# kwargs-ok` /
+        `# rebind-ok` / `# writable-ok` suppression without a reason.
 LIT006  `cast(...)` call. typing.cast is an unchecked assertion (the moral equivalent
         of TypeScript's `as`); it lies to the type checker with zero runtime guarantee.
         Validate into a concrete frozen type at the boundary instead.
@@ -80,6 +80,15 @@ LIT011  Function-argument mutation: a parameter that is re-bound (`param = ...`,
         instance), not from re-binding. Method-call mutation (`param.append(x)`) is
         out of reach without type information; LIT001/LIT002 keep mutable collections
         off signatures instead. Suppress with `# rebind-ok: <reason>`.
+LIT012  TypedDict field without a `ReadOnly[...]` qualifier. A writable key lets any
+        holder of the payload rewrite it after construction; qualify every field with
+        `ReadOnly[...]` (PEP 705), which nests freely with Required/NotRequired/
+        Annotated in any order. Detection is name-based, like MUTABLE_COLLECTIONS:
+        a class is a TypedDict when `TypedDict` appears among its bases or when it
+        inherits, transitively within the same module, from a class that has it;
+        the functional form (`X = TypedDict("X", {...})`) is checked too. A base
+        imported from another module is out of reach without import resolution.
+        Suppress with `# writable-ok: <reason>`.
 
 LIT000  Setup failure: a target file could not be read, or contains a syntax error.
         Reported as a violation rather than crashing the run.
@@ -130,6 +139,11 @@ MUTABLE_CONSTRUCTORS = frozenset((
 QUALIFIED_CONSTRUCTORS = MUTABLE_CONSTRUCTORS - frozenset(("dict", "list", "set"))
 FREEZING_WRAPPERS = frozenset(("tuple", "frozenset", "MappingProxyType"))
 UNSAFE_GUARDS = frozenset(("TypeGuard", "TypeIs"))
+READONLY_QUALIFIER = "ReadOnly"
+# Qualifiers ReadOnly may nest under, in any order (PEP 705); for Annotated only the
+# first argument is type syntax, the rest is metadata and never qualifies the field.
+FIELD_QUALIFIER_WRAPPERS = frozenset(("Required", "NotRequired", "Annotated"))
+TYPEDDICT_BASE = "TypedDict"
 MIN_REASON_LEN = 3
  
 NOQA_RE = re.compile(
@@ -147,6 +161,7 @@ CAST_OK_RE = re.compile(r"#\s*cast-ok(?::\s*(?P<reason>.*))?")
 GUARD_OK_RE = re.compile(r"#\s*guard-ok(?::\s*(?P<reason>.*))?")
 KWARGS_OK_RE = re.compile(r"#\s*kwargs-ok(?::\s*(?P<reason>.*))?")
 REBIND_OK_RE = re.compile(r"#\s*rebind-ok(?::\s*(?P<reason>.*))?")
+WRITABLE_OK_RE = re.compile(r"#\s*writable-ok(?::\s*(?P<reason>.*))?")
 
 # Suppression tokens that must each carry a reason (LIT005).
 OK_SUPPRESSIONS: tuple[tuple[str, re.Pattern[str]], ...] = (
@@ -155,6 +170,7 @@ OK_SUPPRESSIONS: tuple[tuple[str, re.Pattern[str]], ...] = (
     ("guard-ok", GUARD_OK_RE),
     ("kwargs-ok", KWARGS_OK_RE),
     ("rebind-ok", REBIND_OK_RE),
+    ("writable-ok", WRITABLE_OK_RE),
 )
  
  
@@ -177,6 +193,7 @@ class Comments:
     guard_ok_lines: frozenset[int]
     kwargs_ok_lines: frozenset[int]
     rebind_ok_lines: frozenset[int]
+    writable_ok_lines: frozenset[int]
  
  
 # --------------------------------------------------------------------------- #
@@ -232,7 +249,7 @@ def scan_comments(path: Path, source: str) -> tuple[Comments, tuple[Violation, .
         # tokenize raises TokenError (EOF mid-construct) or a SyntaxError subclass
         # (IndentationError / TabError) on malformed source; defer to ast.parse below,
         # which re-raises and is reported as LIT000 rather than crashing the run.
-        return Comments(frozenset(), frozenset(), frozenset(), frozenset(), frozenset()), ()
+        return Comments(frozenset(), frozenset(), frozenset(), frozenset(), frozenset(), frozenset()), ()
 
     def _lines_with(regex: re.Pattern[str]) -> frozenset[int]:
         return frozenset(line for line, text in comment_toks if _valid_ok(regex, text))
@@ -244,6 +261,7 @@ def scan_comments(path: Path, source: str) -> tuple[Comments, tuple[Violation, .
             guard_ok_lines=_lines_with(GUARD_OK_RE),
             kwargs_ok_lines=_lines_with(KWARGS_OK_RE),
             rebind_ok_lines=_lines_with(REBIND_OK_RE),
+            writable_ok_lines=_lines_with(WRITABLE_OK_RE),
         ),
         tuple(v for line, text in comment_toks for v in _comment_violations(path, line, text)),
     )
@@ -828,6 +846,111 @@ def iter_param_violations(path: Path, tree: ast.AST, comments: Comments) -> Iter
             )
 
 
+# --------------------------------------------------------------------------- #
+# Writable TypedDict fields (LIT012)
+# --------------------------------------------------------------------------- #
+
+
+def _head_name(node: ast.expr) -> str | None:
+    if isinstance(node, ast.Name):
+        return node.id
+    if isinstance(node, ast.Attribute):
+        return node.attr
+    return None
+
+
+def _base_names(cls: ast.ClassDef) -> frozenset[str]:
+    """The names of a class's bases; a subscripted base (`Foo[int]`) counts as `Foo`."""
+    return frozenset(
+        name
+        for base in cls.bases
+        for name in (_head_name(base.value if isinstance(base, ast.Subscript) else base),)
+        if name is not None
+    )
+
+
+def _typeddict_classes(tree: ast.AST) -> tuple[ast.ClassDef, ...]:
+    """ClassDefs that are TypedDicts: `TypedDict` among the bases, or -- transitively,
+    within this module -- a base that is itself one of these classes. A base defined
+    in another module is invisible here; that subclass goes unchecked."""
+    classes = tuple(node for node in ast.walk(tree) if isinstance(node, ast.ClassDef))
+    bases_of = {cls.name: _base_names(cls) for cls in classes}
+
+    def expand(known: frozenset[str]) -> frozenset[str]:
+        grown = known | frozenset(name for name, bases in bases_of.items() if bases & known)
+        return grown if grown == known else expand(grown)
+
+    names = expand(frozenset((TYPEDDICT_BASE,)))
+    return tuple(cls for cls in classes if cls.name in names)
+
+
+def _has_readonly_qualifier(annotation: ast.expr) -> bool:
+    """True iff the annotation is `ReadOnly[...]`, possibly nested under
+    Required/NotRequired/Annotated (in any order) or a string forward reference."""
+    if isinstance(annotation, ast.Constant) and isinstance(annotation.value, str):
+        try:
+            inner = ast.parse(annotation.value, mode="eval").body
+        except SyntaxError:
+            return False
+        return _has_readonly_qualifier(inner)
+    if not isinstance(annotation, ast.Subscript):
+        return False
+    name = _head_name(annotation.value)
+    if name == READONLY_QUALIFIER:
+        return True
+    if name not in FIELD_QUALIFIER_WRAPPERS:
+        return False
+    if name == "Annotated":
+        if isinstance(annotation.slice, ast.Tuple) and annotation.slice.elts:
+            return _has_readonly_qualifier(annotation.slice.elts[0])
+        return False
+    return _has_readonly_qualifier(annotation.slice)
+
+
+class _Field(NamedTuple):
+    owner: str
+    name: str
+    annotation: ast.expr
+    line: int
+
+
+def _class_fields(cls: ast.ClassDef) -> Iterator[_Field]:
+    for stmt in cls.body:
+        if isinstance(stmt, ast.AnnAssign) and isinstance(stmt.target, ast.Name):
+            yield _Field(cls.name, stmt.target.id, stmt.annotation, stmt.lineno)
+
+
+def _functional_fields(tree: ast.AST) -> Iterator[_Field]:
+    """Fields of the functional form: `X = TypedDict("X", {"field": type, ...})`."""
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call) or _head_name(node.func) != TYPEDDICT_BASE:
+            continue
+        if len(node.args) < 2 or not isinstance(node.args[1], ast.Dict):
+            continue
+        first = node.args[0]
+        owner = first.value if isinstance(first, ast.Constant) and isinstance(first.value, str) else "<TypedDict>"
+        for key, value in zip(node.args[1].keys, node.args[1].values):
+            if isinstance(key, ast.Constant) and isinstance(key.value, str):
+                yield _Field(owner, key.value, value, value.lineno)
+
+
+def iter_typeddict_violations(path: Path, tree: ast.AST, comments: Comments) -> Iterator[Violation]:
+    fields = (
+        *(f for cls in _typeddict_classes(tree) for f in _class_fields(cls)),
+        *_functional_fields(tree),
+    )
+    for field in fields:
+        if _has_readonly_qualifier(field.annotation) or field.line in comments.writable_ok_lines:
+            continue
+        yield Violation(
+            path, field.line, "LIT012",
+            f"TypedDict field `{field.name}` of `{field.owner}` is writable: any holder "
+            f"of the payload can rewrite the key after construction. Qualify it as "
+            f"`ReadOnly[...]` (PEP 705; nests freely with Required/NotRequired/Annotated) "
+            f"(suppress: `# writable-ok: <reason>`)",
+        )
+
+
 # --------------------------------------------------------------------------- #
 # Driver
 # --------------------------------------------------------------------------- #
@@ -854,6 +977,7 @@ def check_file(path: Path) -> tuple[Violation, ...]:
         *iter_construction_violations(path, tree, comments),
         *iter_final_violations(path, tree, comments),
         *iter_param_violations(path, tree, comments),
+        *iter_typeddict_violations(path, tree, comments),
     )
  
  

scripts/type_discipline_gate.py

@@ -13,10 +13,12 @@ emits is gated: LIT001 (mutable collection in any annotation), LIT002
 without codes or reason), LIT006 (cast), LIT008 (`**kwargs`), LIT009 (inert
 `# type: ignore`, dead syntax while enableTypeIgnoreComments is false), LIT010
 (assignment without a Final declaration; suppress deliberate rebinding with
-`# rebind-ok: <reason>`), and LIT011 (parameter rebinding or in-place mutation)
-carry limits at or above their current count to ratchet down; LIT005 (`*-ok`
-suppression without a reason) is frozen at limit 0 so any net-new reasonless
-suppression trips the gate; and LIT007 (TypeGuard/TypeIs) is a hard zero.
+`# rebind-ok: <reason>`), LIT011 (parameter rebinding or in-place mutation), and
+LIT012 (TypedDict field without a `ReadOnly[...]` qualifier; suppress with
+`# writable-ok: <reason>`) carry limits at or above their current count to
+ratchet down; LIT005 (`*-ok` suppression without a reason) is frozen at limit 0
+so any net-new reasonless suppression trips the gate; and LIT007
+(TypeGuard/TypeIs) is a hard zero.
 LIT010 and LIT011 were seeded at 1.5x the count left after the sweep that
 annotated every never-rebound name with Final, so that headroom is the hard
 line new code cannot cross.
@@ -201,7 +203,8 @@ def cmd_check(base: str) -> None:
         "Remove the new violations, give each a reason (`# noqa: XXX  # <reason>`, "
         "`# pyright: ignore[rule]  # <reason>`, `# mutable-ok: <reason>`, "
         "`# cast-ok: <reason>`, `# guard-ok: <reason>`, `# kwargs-ok: <reason>`, "
-        "`# rebind-ok: <reason>`), or remove an equal number elsewhere; the ceiling "
+        "`# rebind-ok: <reason>`, `# writable-ok: <reason>`), or remove an equal "
+        "number elsewhere; the ceiling "
         "is the limit in type-discipline-budget.json."
     )
     raise SystemExit(1)

terraform/litellm/aws/README.md

@@ -2,9 +2,9 @@
 
 Deploys the componentized LiteLLM proxy on AWS:
 
-- **VPC** with public + private subnets across the AZs you pass in, one NAT gateway
-- **Aurora Postgres** cluster — one writer instance + one reader instance, **IAM database authentication enabled**
-- **ElastiCache Redis** (private, replication group with multi-AZ failover and at-rest + in-transit encryption) for caching + rate limiting
+- **VPC** with public + private subnets across the AZs you pass in, one NAT gateway (skipped when you pass an existing `vpc_id`)
+- **Aurora Postgres** cluster — one writer instance + one reader instance, **IAM database authentication enabled** (skipped when `create_database = false`)
+- **ElastiCache Redis** (private, replication group with multi-AZ failover and at-rest + in-transit encryption) for caching + rate limiting (skipped when `create_redis = false`)
 - **S3 bucket** (private, versioned, SSE-S3) — exposed to gateway + backend as `S3_BUCKET_NAME` / `S3_REGION_NAME` for cache backend, request log archival, and `/v1/files` storage
 - **Secrets Manager** entries for `LITELLM_MASTER_KEY` (auto-generated, `sk-…`) and the Aurora master password (bootstrap-only)
 - **ECS Fargate cluster** running three services — `gateway`, `backend`, `ui`
@@ -14,6 +14,58 @@ Deploys the componentized LiteLLM proxy on AWS:
   - Everything else (management API: `/key/*`, `/user/*`, …) → `backend`
 - **One-off migration task** (`litellm-migrations`) that runs `prisma migrate deploy` from the dedicated `ghcr.io/berriai/litellm-migrations` image
 
+## Bring your own networking, database, and Redis
+
+The three infrastructure pieces the stack would otherwise own are each
+optional, so it can slot into an account where networking and data stores are
+already provisioned (often by another team, in another Terraform state).
+
+**Networking.** Set `vpc_id` plus `public_subnet_ids` and `private_subnet_ids`
+and no VPC, subnet, route table, internet gateway, or NAT gateway is created.
+The ALB goes in the public subnets, the ECS tasks and any subnet group the
+stack still needs go in the private ones, and `vpc_cidr` / `azs` go unused.
+The private subnets need their own egress (NAT gateway, or VPC endpoints
+covering ECR, S3, CloudWatch Logs, and Secrets Manager) since tasks pull
+images, resolve secrets, and call LLM providers.
+
+Security groups stay module-owned in either mode: the ALB group, the tasks
+group, and the database/cache groups when it creates those. To let the tasks
+reach infrastructure the module doesn't manage, either allow inbound from the
+group named by the `task_security_group_id` output, or attach a group of your
+own with `additional_task_security_group_ids`.
+
+```hcl
+vpc_id             = "vpc-0123456789abcdef0"
+public_subnet_ids  = ["subnet-aaa", "subnet-bbb"]
+private_subnet_ids = ["subnet-ccc", "subnet-ddd"]
+```
+
+**Database and Redis.** `create_database` and `create_redis` default to `true`
+(today's behavior). Set one to `false` and pass a connection string to use
+something you already run: the value lands in a Secrets Manager entry and
+reaches gateway, backend, and the migration task as `DATABASE_URL` /
+`REDIS_URL`, both of which outrank the discrete `DATABASE_*` / `REDIS_*` vars
+in the proxy, so nothing appears in plain text in a task definition.
+
+```hcl
+create_database = false
+database_url    = "postgresql://litellm:...@db.internal:5432/litellm"
+create_redis    = false
+redis_url       = "rediss://:...@cache.internal:6379"
+```
+
+The schema migration still runs on every apply against an existing database;
+only the Aurora-specific IAM-user bootstrap drops out, since those credentials
+are already in the URL.
+
+Leaving the URL empty runs without the component entirely:
+
+- No database: no virtual keys, teams, spend tracking, or UI persistence, and
+  `STORE_MODEL_IN_DB` is not set, so models come from `proxy_config`. Requests
+  authenticate with `LITELLM_MASTER_KEY` only.
+- No Redis: rate limits, budgets, and router cooldowns are per-task rather
+  than cluster-wide, which is only sane at one task per service.
+
 ## Aurora + IAM auth
 
 The cluster runs with `iam_database_authentication_enabled = true`. Enabling
@@ -345,7 +397,7 @@ trial / dev stacks only.
 
 ## Storage and database retention
 
-Three opt-in tripwires guard against accidental data loss on
+Two opt-in tripwires guard against accidental data loss on
 `terraform destroy`:
 
 - **`skip_final_snapshot`** (Aurora; default `false`) — destroying the
@@ -354,6 +406,9 @@ Three opt-in tripwires guard against accidental data loss on
   `/v1/files` content, and the S3 cache backend; default `false`) —
   `terraform destroy` against a non-empty bucket fails.
 
+Neither applies to a database you brought yourself: its lifecycle stays with
+whoever provisioned it, and `terraform destroy` leaves it alone.
+
 Flip either to `true` only for ephemeral / CI stacks where you accept
 losing the contents.
 
@@ -365,7 +420,7 @@ losing the contents.
 | `examples/default/` | Thin root: `aws` provider (with an optional `default_tags` slot for org-wide tags) + a call to the module. The one-command deploy path. |
 | `variables.tf`    | All input variables                                                   |
 | `locals.tf`       | Path-prefix lists for ALB routing (mirror of `helm/.../ingress.yaml`) |
-| `network.tf`      | VPC, subnets, IGW, NAT, route tables, security groups                 |
+| `network.tf`      | VPC, subnets, IGW, NAT, route tables (all optional), security groups   |
 | `secrets.tf`      | Secrets Manager entries + random passwords                            |
 | `rds.tf`          | Aurora Postgres cluster + writer / reader instances                   |
 | `redis.tf`        | ElastiCache Redis                                                     |

terraform/litellm/aws/alb.tf

@@ -3,10 +3,17 @@ resource "aws_lb" "this" {
   load_balancer_type = "application"
   internal           = false
   security_groups    = [aws_security_group.alb.id]
-  subnets            = aws_subnet.public[*].id
+  subnets            = local.public_subnet_ids
 
   idle_timeout = 120
 
+  lifecycle {
+    precondition {
+      condition     = length(local.public_subnet_ids) >= 2
+      error_message = "The ALB needs at least 2 public subnets in different AZs. Set `public_subnet_ids` when using `vpc_id`, or list at least 2 `azs` when the module creates the VPC."
+    }
+  }
+
   tags = local.tags
 }
 
@@ -25,7 +32,7 @@ resource "aws_lb_target_group" "gateway" {
   port        = 4000
   protocol    = "HTTP"
   target_type = "ip"
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   health_check {
     path                = "/health/readiness"
@@ -46,7 +53,7 @@ resource "aws_lb_target_group" "backend" {
   port        = 4001
   protocol    = "HTTP"
   target_type = "ip"
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   health_check {
     path                = "/health/readiness"
@@ -67,7 +74,7 @@ resource "aws_lb_target_group" "ui" {
   port        = 3000
   protocol    = "HTTP"
   target_type = "ip"
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   health_check {
     path                = "/healthz"

terraform/litellm/aws/bootstrap.tf

@@ -1,9 +1,12 @@
 # Auto-runs the two manual steps that used to follow `terraform apply`:
 #
 #   1. Create the IAM-authed Postgres user (litellm_app) — uses the postgres:16
-#      image with the master password from Secrets Manager.
+#      image with the master password from Secrets Manager. Only relevant to
+#      the Aurora cluster this module creates, so it is skipped when
+#      create_database = false.
 #   2. Run prisma migrate deploy — reuses the existing aws_ecs_task_definition
-#      .migrations task def from migrations.tf.
+#      .migrations task def from migrations.tf. Runs against an existing
+#      database too, and only disappears when there is no database at all.
 #
 # Both are invoked via `terraform_data` provisioners. Gateway/backend services
 # in ecs.tf depend on `terraform_data.migration`, so on a fresh apply they
@@ -23,13 +26,14 @@
 # extras — see iam.tf). The DB master password lives in a separate secret used
 # only here, so we grant access in an additive policy.
 resource "aws_iam_policy" "bootstrap_secrets" {
-  name = "${local.name}-bootstrap-secrets-access"
+  count = var.create_database ? 1 : 0
+  name  = "${local.name}-bootstrap-secrets-access"
   policy = jsonencode({
     Version = "2012-10-17"
     Statement = [{
       Effect   = "Allow"
       Action   = ["secretsmanager:GetSecretValue"]
-      Resource = [aws_secretsmanager_secret.db_master_password.arn]
+      Resource = [aws_secretsmanager_secret.db_master_password[0].arn]
     }]
   })
 
@@ -37,12 +41,14 @@ resource "aws_iam_policy" "bootstrap_secrets" {
 }
 
 resource "aws_iam_role_policy_attachment" "task_execution_bootstrap_secrets" {
+  count      = var.create_database ? 1 : 0
   role       = aws_iam_role.task_execution.name
-  policy_arn = aws_iam_policy.bootstrap_secrets.arn
+  policy_arn = aws_iam_policy.bootstrap_secrets[0].arn
 }
 
 # ---------- Bootstrap task def ----------
 resource "aws_cloudwatch_log_group" "bootstrap_db" {
+  count             = var.create_database ? 1 : 0
   name              = "/ecs/${local.name}/bootstrap-db"
   retention_in_days = var.log_retention_days
 
@@ -68,6 +74,7 @@ locals {
 }
 
 resource "aws_ecs_task_definition" "bootstrap_db" {
+  count                    = var.create_database ? 1 : 0
   family                   = "${local.name}-bootstrap-db"
   network_mode             = "awsvpc"
   requires_compatibilities = ["FARGATE"]
@@ -82,15 +89,15 @@ resource "aws_ecs_task_definition" "bootstrap_db" {
     essential = true
 
     environment = [
-      { name = "PGHOST", value = aws_rds_cluster.this.endpoint },
-      { name = "PGPORT", value = tostring(aws_rds_cluster.this.port) },
+      { name = "PGHOST", value = aws_rds_cluster.this[0].endpoint },
+      { name = "PGPORT", value = tostring(aws_rds_cluster.this[0].port) },
       { name = "PGUSER", value = var.db_master_username },
       { name = "PGDATABASE", value = var.db_name },
       { name = "BOOTSTRAP_SQL", value = local.bootstrap_sql },
     ]
     secrets = [
       # `:password::` extracts the password field out of the JSON secret.
-      { name = "PGPASSWORD", valueFrom = "${aws_secretsmanager_secret.db_master_password.arn}:password::" },
+      { name = "PGPASSWORD", valueFrom = "${aws_secretsmanager_secret.db_master_password[0].arn}:password::" },
     ]
 
     entryPoint = ["sh", "-c"]
@@ -99,7 +106,7 @@ resource "aws_ecs_task_definition" "bootstrap_db" {
     logConfiguration = {
       logDriver = "awslogs"
       options = {
-        awslogs-group         = aws_cloudwatch_log_group.bootstrap_db.name
+        awslogs-group         = aws_cloudwatch_log_group.bootstrap_db[0].name
         awslogs-region        = var.region
         awslogs-stream-prefix = "bootstrap"
       }
@@ -111,20 +118,22 @@ resource "aws_ecs_task_definition" "bootstrap_db" {
 
 # ---------- Bootstrap trigger ----------
 resource "terraform_data" "bootstrap_db" {
+  count = var.create_database ? 1 : 0
+
   triggers_replace = {
-    cluster_resource_id = aws_rds_cluster.this.cluster_resource_id
-    task_def_revision   = aws_ecs_task_definition.bootstrap_db.revision
+    cluster_resource_id = aws_rds_cluster.this[0].cluster_resource_id
+    task_def_revision   = aws_ecs_task_definition.bootstrap_db[0].revision
   }
 
   provisioner "local-exec" {
     interpreter = ["bash", "-c"]
     environment = {
       CLUSTER  = aws_ecs_cluster.this.name
-      TASK_DEF = aws_ecs_task_definition.bootstrap_db.arn
-      SUBNETS  = join(",", aws_subnet.private[*].id)
-      SG       = aws_security_group.tasks.id
+      TASK_DEF = aws_ecs_task_definition.bootstrap_db[0].arn
+      SUBNETS  = join(",", local.private_subnet_ids)
+      SG       = join(",", local.task_security_group_ids)
       REGION   = var.region
-      LOG_GRP  = aws_cloudwatch_log_group.bootstrap_db.name
+      LOG_GRP  = aws_cloudwatch_log_group.bootstrap_db[0].name
     }
     command = <<-EOT
       set -euo pipefail
@@ -144,9 +153,13 @@ resource "terraform_data" "bootstrap_db" {
     EOT
   }
 
+  # Same secret-by-ARN gap as the migration below. The margin here is wide,
+  # since the writer instance takes minutes while the version write does not,
+  # but both hang off the cluster in parallel and nothing orders them.
   depends_on = [
     aws_rds_cluster_instance.writer,
     aws_iam_role_policy_attachment.task_execution_bootstrap_secrets,
+    aws_secretsmanager_secret_version.db_master_password,
   ]
 }
 
@@ -154,20 +167,22 @@ resource "terraform_data" "bootstrap_db" {
 # Reuses the task definition from migrations.tf — this resource just invokes
 # it and waits.
 resource "terraform_data" "migration" {
+  count = local.database_enabled ? 1 : 0
+
   triggers_replace = {
-    task_def_revision = aws_ecs_task_definition.migrations.revision
-    bootstrap_id      = terraform_data.bootstrap_db.id
+    task_def_revision = aws_ecs_task_definition.migrations[0].revision
+    bootstrap_id      = join(",", terraform_data.bootstrap_db[*].id)
   }
 
   provisioner "local-exec" {
     interpreter = ["bash", "-c"]
     environment = {
       CLUSTER  = aws_ecs_cluster.this.name
-      TASK_DEF = aws_ecs_task_definition.migrations.arn
-      SUBNETS  = join(",", aws_subnet.private[*].id)
-      SG       = aws_security_group.tasks.id
+      TASK_DEF = aws_ecs_task_definition.migrations[0].arn
+      SUBNETS  = join(",", local.private_subnet_ids)
+      SG       = join(",", local.task_security_group_ids)
       REGION   = var.region
-      LOG_GRP  = aws_cloudwatch_log_group.migrations.name
+      LOG_GRP  = aws_cloudwatch_log_group.migrations[0].name
     }
     command = <<-EOT
       set -euo pipefail
@@ -187,5 +202,14 @@ resource "terraform_data" "migration" {
     EOT
   }
 
-  depends_on = [terraform_data.bootstrap_db]
+  # A container reads a secret by ARN, so Terraform sees no edge from the
+  # ARN to the _version that gives it a value. The managed-Aurora path hides
+  # that: the cluster create takes long enough that the version always lands
+  # first. A bring-your-own database has nothing slow in between, so without
+  # this the run-task below can fire against a valueless secret and fail the
+  # apply with ResourceInitializationError.
+  depends_on = [
+    terraform_data.bootstrap_db,
+    aws_secretsmanager_secret_version.database_url,
+  ]
 }

terraform/litellm/aws/ecs.tf

@@ -31,6 +31,7 @@ resource "aws_cloudwatch_log_group" "ui" {
 }
 
 resource "aws_cloudwatch_log_group" "migrations" {
+  count             = local.database_enabled ? 1 : 0
   name              = "/ecs/${local.name}/migrations"
   retention_in_days = var.log_retention_days
 
@@ -38,11 +39,13 @@ resource "aws_cloudwatch_log_group" "migrations" {
 }
 
 # Shared env block fed to gateway, backend, and the migration task. Mirrors
-# the helm chart's `litellm.serverEnv` helper on the IAM-auth branch:
-# DATABASE_URL is assembled at runtime by
+# the helm chart's `litellm.serverEnv` helper on the IAM-auth branch: for the
+# module-created Aurora, DATABASE_URL is assembled at runtime by
 # litellm/proxy/auth/rds_iam_token.py::init_iam_db_url_from_env from
 # HOST/PORT/USER/NAME plus an IAM-signed token, so no DB password is needed
-# in the task definition.
+# in the task definition. An existing database instead arrives as a
+# DATABASE_URL secret (var.database_url), which run.py and the proxy both
+# take as-is.
 locals {
   # OTel v2 is opt-in and gated on otel_endpoint, matching the GCP stack.
   # When set, LITELLM_OTEL_V2 flips on alongside the OTEL_* block, with
@@ -103,29 +106,50 @@ locals {
     ] : [],
   )
 
-  shared_env = [
+  managed_db_env = var.create_database ? [
     { name = "IAM_TOKEN_DB_AUTH", value = "true" },
-    { name = "DATABASE_HOST", value = aws_rds_cluster.this.endpoint },
-    { name = "DATABASE_PORT", value = tostring(aws_rds_cluster.this.port) },
+    { name = "DATABASE_HOST", value = aws_rds_cluster.this[0].endpoint },
+    { name = "DATABASE_PORT", value = tostring(aws_rds_cluster.this[0].port) },
     { name = "DATABASE_USER", value = var.db_username },
     { name = "DATABASE_NAME", value = var.db_name },
-    { name = "DATABASE_HOST_READ_REPLICA", value = aws_rds_cluster.this.reader_endpoint },
-    { name = "DATABASE_PORT_READ_REPLICA", value = tostring(aws_rds_cluster.this.port) },
-    { name = "REDIS_HOST", value = aws_elasticache_replication_group.this.primary_endpoint_address },
-    { name = "REDIS_PORT", value = tostring(aws_elasticache_replication_group.this.port) },
+    { name = "DATABASE_HOST_READ_REPLICA", value = aws_rds_cluster.this[0].reader_endpoint },
+    { name = "DATABASE_PORT_READ_REPLICA", value = tostring(aws_rds_cluster.this[0].port) },
+  ] : []
+
+  managed_redis_env = var.create_redis ? [
+    { name = "REDIS_HOST", value = aws_elasticache_replication_group.this[0].primary_endpoint_address },
+    { name = "REDIS_PORT", value = tostring(aws_elasticache_replication_group.this[0].port) },
     # transit_encryption_enabled = true on the replication group means the
     # proxy must connect via rediss://. _redis.get_redis_url_from_environment
     # honors REDIS_SSL to flip the scheme.
     { name = "REDIS_SSL", value = "true" },
-    # S3 bucket — referenced from proxy_config via os.environ/S3_BUCKET_NAME
-    # (e.g. cache backend, request log archival, /files passthrough).
-    { name = "S3_BUCKET_NAME", value = aws_s3_bucket.this.bucket },
-    { name = "S3_REGION_NAME", value = var.region },
-    # boto3 inside generate_iam_auth_token reads AWS_REGION_NAME first, then
-    # AWS_REGION. Set both for compatibility.
-    { name = "AWS_REGION", value = var.region },
-    { name = "AWS_REGION_NAME", value = var.region },
-  ]
+  ] : []
+
+  shared_env = concat(
+    local.managed_db_env,
+    local.managed_redis_env,
+    [
+      # S3 bucket — referenced from proxy_config via os.environ/S3_BUCKET_NAME
+      # (e.g. cache backend, request log archival, /files passthrough).
+      { name = "S3_BUCKET_NAME", value = aws_s3_bucket.this.bucket },
+      { name = "S3_REGION_NAME", value = var.region },
+      # boto3 inside generate_iam_auth_token reads AWS_REGION_NAME first, then
+      # AWS_REGION. Set both for compatibility.
+      { name = "AWS_REGION", value = var.region },
+      { name = "AWS_REGION_NAME", value = var.region },
+    ],
+  )
+
+  # DATABASE_URL / REDIS_URL both outrank the discrete host/port vars in the
+  # proxy, so the BYO branch needs nothing removed from shared_env: the
+  # managed_*_env blocks are already empty whenever these are set.
+  byo_database_secrets = local.byo_database ? [
+    { name = "DATABASE_URL", valueFrom = aws_secretsmanager_secret.database_url[0].arn },
+  ] : []
+
+  byo_redis_secrets = local.byo_redis ? [
+    { name = "REDIS_URL", valueFrom = aws_secretsmanager_secret.redis_url[0].arn },
+  ] : []
 
   shared_secrets = concat(
     [
@@ -134,6 +158,8 @@ locals {
     var.litellm_license == "" ? [] : [
       { name = "LITELLM_LICENSE", valueFrom = aws_secretsmanager_secret.license[0].arn },
     ],
+    local.byo_database_secrets,
+    local.byo_redis_secrets,
     local.otel_secrets,
     local.billing_metrics_secrets,
   )
@@ -151,9 +177,11 @@ locals {
     for k, v in var.backend_extra_env : { name = k, value = v }
   ]
 
-  backend_default_env = [
+  # Storing models in the DB needs a DB. Without one the backend reads its
+  # model list from proxy_config only.
+  backend_default_env = local.database_enabled ? [
     { name = "STORE_MODEL_IN_DB", value = "true" },
-  ]
+  ] : []
   gateway_extra_secrets_list = [
     for k, v in var.gateway_extra_secrets : { name = k, valueFrom = v }
   ]
@@ -286,8 +314,8 @@ resource "aws_ecs_service" "gateway" {
   launch_type     = "FARGATE"
 
   network_configuration {
-    subnets          = aws_subnet.private[*].id
-    security_groups  = [aws_security_group.tasks.id]
+    subnets          = local.private_subnet_ids
+    security_groups  = local.task_security_group_ids
     assign_public_ip = false
   }
 
@@ -308,10 +336,20 @@ resource "aws_ecs_service" "gateway" {
 
   # Don't start until the schema migration has run. Otherwise the proxy
   # boots, Prisma fails on the missing tables, and ECS thrashes the task.
+  # The _version entries are listed because a task reads its secrets by ARN,
+  # which gives Terraform no edge to the resource that writes the value; the
+  # migration covers that ordering only while a database exists.
   depends_on = [
     aws_lb_listener.http,
     aws_lb_listener.https,
     terraform_data.migration,
+    aws_secretsmanager_secret_version.master_key,
+    aws_secretsmanager_secret_version.license,
+    aws_secretsmanager_secret_version.database_url,
+    aws_secretsmanager_secret_version.redis_url,
+    aws_secretsmanager_secret_version.billing_metrics_client_cert,
+    aws_secretsmanager_secret_version.billing_metrics_client_key,
+    aws_secretsmanager_secret_version.billing_metrics_ca_cert,
   ]
 
   tags = local.tags
@@ -381,8 +419,8 @@ resource "aws_ecs_service" "backend" {
   launch_type     = "FARGATE"
 
   network_configuration {
-    subnets          = aws_subnet.private[*].id
-    security_groups  = [aws_security_group.tasks.id]
+    subnets          = local.private_subnet_ids
+    security_groups  = local.task_security_group_ids
     assign_public_ip = false
   }
 
@@ -399,10 +437,20 @@ resource "aws_ecs_service" "backend" {
     ignore_changes = [desired_count]
   }
 
+  # Same secret-version ordering as the gateway, plus UI_PASSWORD, which only
+  # the backend consumes.
   depends_on = [
     aws_lb_listener.http,
     aws_lb_listener.https,
     terraform_data.migration,
+    aws_secretsmanager_secret_version.master_key,
+    aws_secretsmanager_secret_version.license,
+    aws_secretsmanager_secret_version.ui_password,
+    aws_secretsmanager_secret_version.database_url,
+    aws_secretsmanager_secret_version.redis_url,
+    aws_secretsmanager_secret_version.billing_metrics_client_cert,
+    aws_secretsmanager_secret_version.billing_metrics_client_key,
+    aws_secretsmanager_secret_version.billing_metrics_ca_cert,
   ]
 
   tags = local.tags
@@ -451,8 +499,8 @@ resource "aws_ecs_service" "ui" {
   launch_type     = "FARGATE"
 
   network_configuration {
-    subnets          = aws_subnet.private[*].id
-    security_groups  = [aws_security_group.tasks.id]
+    subnets          = local.private_subnet_ids
+    security_groups  = local.task_security_group_ids
     assign_public_ip = false
   }
 

terraform/litellm/aws/examples/default/main.tf

@@ -24,6 +24,16 @@ module "litellm" {
   env    = var.env
   azs    = var.azs
 
+  vpc_id                             = var.vpc_id
+  public_subnet_ids                  = var.public_subnet_ids
+  private_subnet_ids                 = var.private_subnet_ids
+  additional_task_security_group_ids = var.additional_task_security_group_ids
+
+  create_database = var.create_database
+  database_url    = var.database_url
+  create_redis    = var.create_redis
+  redis_url       = var.redis_url
+
   litellm_master_key = var.litellm_master_key
   litellm_license    = var.litellm_license
   ui_password        = var.ui_password

terraform/litellm/aws/examples/default/outputs.tf

@@ -13,6 +13,16 @@ output "ecs_cluster" {
   value       = module.litellm.ecs_cluster
 }
 
+output "vpc_id" {
+  description = "VPC the stack runs in, whether module-created or supplied."
+  value       = module.litellm.vpc_id
+}
+
+output "task_security_group_id" {
+  description = "Tasks security group. Allow this inbound on an existing database or Redis."
+  value       = module.litellm.task_security_group_id
+}
+
 output "aurora_writer_endpoint" {
   description = "Aurora writer endpoint."
   value       = module.litellm.aurora_writer_endpoint

terraform/litellm/aws/examples/default/terraform.tfvars.example

@@ -1,5 +1,35 @@
 region = "us-west-2"
-azs    = ["us-west-2a", "us-west-2b"]
+
+# Networking: by default the module creates a VPC, public/private subnets in
+# each AZ listed here, an internet gateway, a NAT gateway, and route tables.
+azs = ["us-west-2a", "us-west-2b"]
+
+# To deploy into networking you already own, drop `azs` and set these
+# instead. Nothing network-related is created then, so the private subnets
+# need their own egress for LLM providers, image pulls, and Secrets Manager.
+# vpc_id             = "vpc-0123456789abcdef0"
+# public_subnet_ids  = ["subnet-aaa", "subnet-bbb"]
+# private_subnet_ids = ["subnet-ccc", "subnet-ddd"]
+#
+# The tasks get their own security group either way. To reach a store that
+# only allows a group you already have, attach it here as well; the
+# `task_security_group_id` output names the module's own group.
+# additional_task_security_group_ids = ["sg-0123456789abcdef0"]
+
+# Data stores: Aurora Postgres and ElastiCache Redis are created by default.
+# Set create_* = false to point at your own, passing a connection string
+# (stored in Secrets Manager, injected as DATABASE_URL / REDIS_URL). Make
+# sure they allow inbound from the stack's tasks security group, which the
+# `task_security_group_id` output names.
+# create_database = false
+# database_url    = "postgresql://litellm:...@db.internal:5432/litellm"
+# create_redis    = false
+# redis_url       = "rediss://:...@cache.internal:6379"
+#
+# Leaving the URL empty runs without that component: no database means no
+# virtual keys, spend tracking, or UI persistence (master-key auth only), and
+# no Redis means rate limits, budgets, and router cooldowns go per-task
+# instead of cluster-wide.
 
 # Resource naming: every AWS resource the stack creates is named
 # `${tenant}-litellm-${env}` (or that plus a per-resource suffix). E.g.

terraform/litellm/aws/examples/default/variables.tf

@@ -21,8 +21,64 @@ variable "env" {
 }
 
 variable "azs" {
-  description = "Availability zones for subnets. At least 2 (RDS + ALB)."
+  description = "Availability zones for the subnets the module creates. At least 2 (RDS + ALB). Unused when vpc_id is set."
   type        = list(string)
+  default     = []
+}
+
+# Bring-your-own networking. Leave vpc_id empty to have the module create the
+# VPC, subnets, NAT gateway, and route tables.
+variable "vpc_id" {
+  description = "Existing VPC to deploy into. Empty → module creates its own networking."
+  type        = string
+  default     = ""
+}
+
+variable "public_subnet_ids" {
+  description = "Existing public subnets for the ALB (≥ 2 AZs). Required with vpc_id."
+  type        = list(string)
+  default     = []
+}
+
+variable "private_subnet_ids" {
+  description = "Existing private subnets for tasks, Aurora, and Redis. Required with vpc_id."
+  type        = list(string)
+  default     = []
+}
+
+variable "additional_task_security_group_ids" {
+  description = "Extra security groups for the tasks, e.g. one an existing database already allows."
+  type        = list(string)
+  default     = []
+}
+
+# Bring-your-own data stores. create_* false with an empty URL runs without
+# that component: no DB means no key management or spend tracking, no Redis
+# means per-task rate limits instead of cluster-wide.
+variable "create_database" {
+  description = "Create the Aurora Postgres cluster. False → use database_url, or run DB-less."
+  type        = bool
+  default     = true
+}
+
+variable "database_url" {
+  description = "Postgres connection string for an existing database. Read only when create_database = false."
+  type        = string
+  default     = ""
+  sensitive   = true
+}
+
+variable "create_redis" {
+  description = "Create the ElastiCache Redis group. False → use redis_url, or run without Redis."
+  type        = bool
+  default     = true
+}
+
+variable "redis_url" {
+  description = "Connection string for an existing Redis. Read only when create_redis = false."
+  type        = string
+  default     = ""
+  sensitive   = true
 }
 
 # Sensitive — prefer TF_VAR_litellm_master_key / TF_VAR_litellm_license /

terraform/litellm/aws/iam.tf

@@ -56,6 +56,8 @@ data "aws_iam_policy_document" "secrets_access" {
       aws_secretsmanager_secret.billing_metrics_client_cert[*].arn,
       aws_secretsmanager_secret.billing_metrics_client_key[*].arn,
       aws_secretsmanager_secret.billing_metrics_ca_cert[*].arn,
+      aws_secretsmanager_secret.database_url[*].arn,
+      aws_secretsmanager_secret.redis_url[*].arn,
       local.extra_secret_arns,
       var.otel_headers_secret_arn == "" ? [] : [var.otel_headers_secret_arn],
     )
@@ -79,6 +81,9 @@ resource "aws_iam_role_policy_attachment" "task_execution_secrets" {
 # Assumed by the running container. Gets `rds-db:connect` so the proxy can
 # mint IAM-signed Postgres tokens for the app user. Layer additional
 # policies here (e.g. Bedrock invoke, S3 read) when the proxy needs them.
+# IAM auth only applies to the Aurora cluster this module creates: an
+# existing database is reached with the credentials embedded in
+# var.database_url, so the policy is skipped there.
 
 resource "aws_iam_role" "task" {
   name               = "${local.name}-task"
@@ -90,24 +95,28 @@ resource "aws_iam_role" "task" {
 data "aws_caller_identity" "current" {}
 
 data "aws_iam_policy_document" "rds_iam_connect" {
+  count = var.create_database ? 1 : 0
+
   statement {
     actions = ["rds-db:connect"]
     resources = [
-      "arn:aws:rds-db:${var.region}:${data.aws_caller_identity.current.account_id}:dbuser:${aws_rds_cluster.this.cluster_resource_id}/${var.db_username}",
+      "arn:aws:rds-db:${var.region}:${data.aws_caller_identity.current.account_id}:dbuser:${aws_rds_cluster.this[0].cluster_resource_id}/${var.db_username}",
     ]
   }
 }
 
 resource "aws_iam_policy" "rds_iam_connect" {
+  count  = var.create_database ? 1 : 0
   name   = "${local.name}-rds-iam-connect"
-  policy = data.aws_iam_policy_document.rds_iam_connect.json
+  policy = data.aws_iam_policy_document.rds_iam_connect[0].json
 
   tags = local.tags
 }
 
 resource "aws_iam_role_policy_attachment" "task_rds_iam_connect" {
+  count      = var.create_database ? 1 : 0
   role       = aws_iam_role.task.name
-  policy_arn = aws_iam_policy.rds_iam_connect.arn
+  policy_arn = aws_iam_policy.rds_iam_connect[0].arn
 }
 
 # ---------- UI task role ----------

terraform/litellm/aws/locals.tf

@@ -25,6 +25,36 @@ locals {
     var.tags,
   )
 
+  # Networking, database, and cache are each either module-owned or
+  # bring-your-own. Everything downstream reads these locals rather than the
+  # resources, so a resource going to zero instances doesn't ripple.
+  create_vpc         = var.vpc_id == ""
+  vpc_id             = local.create_vpc ? aws_vpc.this[0].id : var.vpc_id
+  public_subnet_ids  = local.create_vpc ? aws_subnet.public[*].id : var.public_subnet_ids
+  private_subnet_ids = local.create_vpc ? aws_subnet.private[*].id : var.private_subnet_ids
+
+  task_security_group_ids = concat([aws_security_group.tasks.id], var.additional_task_security_group_ids)
+
+  # `byo_*` is the existing-store branch, `database_enabled` is either branch.
+  # Neither branch means the component is absent: no DB (no key management,
+  # spend tracking, or UI persistence) or no Redis (per-task rate limits and
+  # cooldowns instead of cluster-wide).
+  # nonsensitive() on the emptiness check only: without it the sensitivity of
+  # the URLs propagates into every value derived from these flags, redacting
+  # unrelated task-definition and output diffs in the plan.
+  byo_database     = !var.create_database && nonsensitive(var.database_url != "")
+  byo_redis        = !var.create_redis && nonsensitive(var.redis_url != "")
+  database_enabled = var.create_database || local.byo_database
+  redis_enabled    = var.create_redis || local.byo_redis
+
+  # Aurora and ElastiCache subnet groups both demand two AZs, so supplied
+  # private subnets have to cover two whenever either store is module-created.
+  managed_stores_need_two_azs = var.create_database || var.create_redis
+
+  # Every uvicorn worker in every gateway task counts its own rate limits when
+  # there is no Redis to share them through, so the ceiling is tasks x workers.
+  max_gateway_processes = (var.gateway_autoscaling_enabled ? var.gateway_max_capacity : var.gateway_desired_count) * var.gateway_num_workers
+
   gateway_path_prefixes = [
     "/v1/chat/*", "/chat/*",
     "/v1/completions*", "/completions*",

terraform/litellm/aws/migrations.tf

@@ -13,6 +13,7 @@
 # every apply (after the IAM-authed user has been created). The
 # `migration_run_command` output is preserved for break-glass manual re-runs.
 resource "aws_ecs_task_definition" "migrations" {
+  count                    = local.database_enabled ? 1 : 0
   family                   = "${local.name}-migrations"
   network_mode             = "awsvpc"
   requires_compatibilities = ["FARGATE"]
@@ -32,11 +33,12 @@ resource "aws_ecs_task_definition" "migrations" {
 
     # No entryPoint/command override — the image's ENTRYPOINT runs run.py.
     environment = local.shared_env
+    secrets     = local.byo_database_secrets
 
     logConfiguration = {
       logDriver = "awslogs"
       options = {
-        awslogs-group         = aws_cloudwatch_log_group.migrations.name
+        awslogs-group         = aws_cloudwatch_log_group.migrations[0].name
         awslogs-region        = var.region
         awslogs-stream-prefix = "migrations"
       }

terraform/litellm/aws/network.tf

@@ -1,24 +1,34 @@
-data "aws_availability_zones" "available" {
-  state = "available"
-}
+# Networking is created only when the caller didn't supply a VPC. With
+# var.vpc_id set, every resource in this file except the security groups has
+# zero instances and the stack consumes the caller's subnets through
+# local.public_subnet_ids / local.private_subnet_ids (see locals.tf).
 
 resource "aws_vpc" "this" {
+  count                = local.create_vpc ? 1 : 0
   cidr_block           = var.vpc_cidr
   enable_dns_hostnames = true
   enable_dns_support   = true
 
+  lifecycle {
+    precondition {
+      condition     = length(var.azs) >= 2
+      error_message = "Provide at least 2 availability zones in `azs`, or set `vpc_id` + `public_subnet_ids` + `private_subnet_ids` to deploy into an existing VPC."
+    }
+  }
+
   tags = merge(local.tags, { Name = local.name })
 }
 
 resource "aws_internet_gateway" "this" {
-  vpc_id = aws_vpc.this.id
+  count  = local.create_vpc ? 1 : 0
+  vpc_id = aws_vpc.this[0].id
   tags   = merge(local.tags, { Name = local.name })
 }
 
 # Public subnets (ALB + NAT). One per AZ.
 resource "aws_subnet" "public" {
-  count                   = length(var.azs)
-  vpc_id                  = aws_vpc.this.id
+  count                   = local.create_vpc ? length(var.azs) : 0
+  vpc_id                  = aws_vpc.this[0].id
   cidr_block              = cidrsubnet(var.vpc_cidr, 8, count.index)
   availability_zone       = var.azs[count.index]
   map_public_ip_on_launch = true
@@ -29,8 +39,8 @@ resource "aws_subnet" "public" {
 # Private subnets (ECS tasks, RDS, ElastiCache). One per AZ, separate from
 # public range.
 resource "aws_subnet" "private" {
-  count             = length(var.azs)
-  vpc_id            = aws_vpc.this.id
+  count             = local.create_vpc ? length(var.azs) : 0
+  vpc_id            = aws_vpc.this[0].id
   cidr_block        = cidrsubnet(var.vpc_cidr, 8, count.index + 10)
   availability_zone = var.azs[count.index]
 
@@ -38,6 +48,7 @@ resource "aws_subnet" "private" {
 }
 
 resource "aws_eip" "nat" {
+  count  = local.create_vpc ? 1 : 0
   domain = "vpc"
   tags   = merge(local.tags, { Name = "${local.name}-nat" })
 
@@ -47,7 +58,8 @@ resource "aws_eip" "nat" {
 # Single NAT gateway in the first public subnet. For HA, replicate per AZ —
 # adds ~$30/mo per gateway, so off by default for a baseline deployment.
 resource "aws_nat_gateway" "this" {
-  allocation_id = aws_eip.nat.id
+  count         = local.create_vpc ? 1 : 0
+  allocation_id = aws_eip.nat[0].id
   subnet_id     = aws_subnet.public[0].id
 
   tags = merge(local.tags, { Name = local.name })
@@ -56,45 +68,53 @@ resource "aws_nat_gateway" "this" {
 }
 
 resource "aws_route_table" "public" {
-  vpc_id = aws_vpc.this.id
+  count  = local.create_vpc ? 1 : 0
+  vpc_id = aws_vpc.this[0].id
 
   route {
     cidr_block = "0.0.0.0/0"
-    gateway_id = aws_internet_gateway.this.id
+    gateway_id = aws_internet_gateway.this[0].id
   }
 
   tags = merge(local.tags, { Name = "${local.name}-public" })
 }
 
 resource "aws_route_table_association" "public" {
-  count          = length(var.azs)
+  count          = local.create_vpc ? length(var.azs) : 0
   subnet_id      = aws_subnet.public[count.index].id
-  route_table_id = aws_route_table.public.id
+  route_table_id = aws_route_table.public[0].id
 }
 
 resource "aws_route_table" "private" {
-  vpc_id = aws_vpc.this.id
+  count  = local.create_vpc ? 1 : 0
+  vpc_id = aws_vpc.this[0].id
 
   route {
     cidr_block     = "0.0.0.0/0"
-    nat_gateway_id = aws_nat_gateway.this.id
+    nat_gateway_id = aws_nat_gateway.this[0].id
   }
 
   tags = merge(local.tags, { Name = "${local.name}-private" })
 }
 
 resource "aws_route_table_association" "private" {
-  count          = length(var.azs)
+  count          = local.create_vpc ? length(var.azs) : 0
   subnet_id      = aws_subnet.private[count.index].id
-  route_table_id = aws_route_table.private.id
+  route_table_id = aws_route_table.private[0].id
 }
 
 # ---------- Security groups ----------
+#
+# Always module-owned, in local.vpc_id, so the stack keeps a least-privilege
+# path between its own components even when it borrows someone else's VPC.
+# Existing databases and caches reached over var.database_url / var.redis_url
+# need to allow inbound from the tasks group (or from a group passed via
+# var.additional_task_security_group_ids).
 
 resource "aws_security_group" "alb" {
   name        = "${local.name}-alb"
   description = "Inbound HTTP/HTTPS to the LiteLLM ALB."
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   ingress {
     description = "HTTP from anywhere"
@@ -126,7 +146,7 @@ resource "aws_security_group" "alb" {
 resource "aws_security_group" "tasks" {
   name        = "${local.name}-tasks"
   description = "ECS tasks (gateway/backend/ui)."
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   ingress {
     description     = "ALB to tasks"
@@ -144,13 +164,23 @@ resource "aws_security_group" "tasks" {
     cidr_blocks = ["0.0.0.0/0"]
   }
 
+  # The tasks group is created in every mode, so this is where the
+  # bring-your-own-VPC inputs get checked.
+  lifecycle {
+    precondition {
+      condition     = local.create_vpc || length(var.private_subnet_ids) >= (local.managed_stores_need_two_azs ? 2 : 1)
+      error_message = "`private_subnet_ids` is required when `vpc_id` is set: the tasks, Aurora, and ElastiCache all live in private subnets. Aurora and ElastiCache subnet groups need subnets in at least 2 AZs, so pass 2 unless both `create_database` and `create_redis` are false."
+    }
+  }
+
   tags = local.tags
 }
 
 resource "aws_security_group" "rds" {
+  count       = var.create_database ? 1 : 0
   name        = "${local.name}-rds"
   description = "RDS Postgres - tasks only."
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   ingress {
     description     = "Postgres from ECS tasks"
@@ -164,9 +194,10 @@ resource "aws_security_group" "rds" {
 }
 
 resource "aws_security_group" "redis" {
+  count       = var.create_redis ? 1 : 0
   name        = "${local.name}-redis"
   description = "ElastiCache Redis - tasks only."
-  vpc_id      = aws_vpc.this.id
+  vpc_id      = local.vpc_id
 
   ingress {
     description     = "Redis from ECS tasks"

terraform/litellm/aws/outputs.tf

@@ -13,19 +13,29 @@ output "ecs_cluster" {
   value       = aws_ecs_cluster.this.name
 }
 
+output "vpc_id" {
+  description = "VPC the stack runs in, whether module-created or passed in via `vpc_id`."
+  value       = local.vpc_id
+}
+
+output "task_security_group_id" {
+  description = "Security group attached to the ECS tasks. Allow inbound from this group on an existing database or Redis reached over `database_url` / `redis_url`."
+  value       = aws_security_group.tasks.id
+}
+
 output "aurora_writer_endpoint" {
-  description = "Aurora writer endpoint (cluster endpoint). Used by gateway/backend as DATABASE_HOST."
-  value       = aws_rds_cluster.this.endpoint
+  description = "Aurora writer endpoint (cluster endpoint). Used by gateway/backend as DATABASE_HOST. Null when `create_database = false`."
+  value       = one(aws_rds_cluster.this[*].endpoint)
 }
 
 output "aurora_reader_endpoint" {
-  description = "Aurora reader endpoint. Used by gateway/backend as DATABASE_HOST_READ_REPLICA."
-  value       = aws_rds_cluster.this.reader_endpoint
+  description = "Aurora reader endpoint. Used by gateway/backend as DATABASE_HOST_READ_REPLICA. Null when `create_database = false`."
+  value       = one(aws_rds_cluster.this[*].reader_endpoint)
 }
 
 output "redis_endpoint" {
-  description = "ElastiCache Redis primary endpoint (TLS, transit_encryption_enabled = true)."
-  value       = "${aws_elasticache_replication_group.this.primary_endpoint_address}:${aws_elasticache_replication_group.this.port}"
+  description = "ElastiCache Redis primary endpoint (TLS, transit_encryption_enabled = true). Null when `create_redis = false`."
+  value       = one([for r in aws_elasticache_replication_group.this : "${r.primary_endpoint_address}:${r.port}"])
 }
 
 output "s3_bucket" {
@@ -39,15 +49,17 @@ output "master_key_secret_arn" {
 }
 
 output "db_master_password_secret_arn" {
-  description = "Secrets Manager ARN holding the Aurora master credentials (bootstrap-only). Used to create the IAM-authed application user."
-  value       = aws_secretsmanager_secret.db_master_password.arn
+  description = "Secrets Manager ARN holding the Aurora master credentials (bootstrap-only). Used to create the IAM-authed application user. Null when `create_database = false`."
+  value       = one(aws_secretsmanager_secret.db_master_password[*].arn)
 }
 
 # Pre-baked SQL to run once as the master user, creating the IAM-authed
 # application user that gateway/backend/migration tasks will authenticate as.
+# Irrelevant to an existing database reached over `database_url`, whose
+# credentials are already in the URL.
 output "db_bootstrap_sql" {
-  description = "Run this once as the master DB user (after the first apply) to create the IAM-authed app user."
-  value       = <<-SQL
+  description = "Run this once as the master DB user (after the first apply) to create the IAM-authed app user. Empty when `create_database = false`."
+  value       = !var.create_database ? "" : <<-SQL
     CREATE USER ${var.db_username};
     GRANT rds_iam TO ${var.db_username};
     GRANT ALL PRIVILEGES ON DATABASE ${var.db_name} TO ${var.db_username};
@@ -60,13 +72,13 @@ output "db_bootstrap_sql" {
 # Pre-baked command for running the one-off migration task. ECS run-task
 # needs the subnet + SG IDs at call time, so we render the full command.
 output "migration_run_command" {
-  description = "Shell command that runs the one-off prisma migration task against Aurora. Run this once, after the bootstrap SQL above, before sending traffic."
-  value = format(
+  description = "Shell command that runs the one-off prisma migration task against the database. Run this once, after the bootstrap SQL above, before sending traffic. Empty when the stack has no database."
+  value = !local.database_enabled ? "" : format(
     "aws ecs run-task --cluster %s --launch-type FARGATE --task-definition %s --network-configuration 'awsvpcConfiguration={subnets=[%s],securityGroups=[%s],assignPublicIp=DISABLED}' --region %s",
     aws_ecs_cluster.this.name,
-    aws_ecs_task_definition.migrations.arn,
-    join(",", aws_subnet.private[*].id),
-    aws_security_group.tasks.id,
+    aws_ecs_task_definition.migrations[0].arn,
+    join(",", local.private_subnet_ids),
+    join(",", local.task_security_group_ids),
     var.region,
   )
 }

terraform/litellm/aws/rds.tf

@@ -1,5 +1,7 @@
 # Aurora Postgres cluster with one writer + one reader instance, IAM
-# database authentication enabled.
+# database authentication enabled. Skipped entirely when
+# create_database = false, in which case the stack either talks to the
+# database named by var.database_url or runs without one.
 #
 # Important: enabling IAM auth on the cluster does not by itself grant any
 # Postgres user the ability to log in with an IAM token. After the first
@@ -17,13 +19,15 @@
 # superusers — keep it for break-glass only.
 
 resource "aws_db_subnet_group" "this" {
+  count      = var.create_database ? 1 : 0
   name       = "${local.name}-db"
-  subnet_ids = aws_subnet.private[*].id
+  subnet_ids = local.private_subnet_ids
 
   tags = local.tags
 }
 
 resource "aws_rds_cluster_parameter_group" "this" {
+  count       = var.create_database ? 1 : 0
   name        = "${local.name}-cluster-pg"
   family      = "aurora-postgresql${split(".", var.db_engine_version)[0]}"
   description = "LiteLLM Aurora Postgres cluster parameters."
@@ -32,16 +36,17 @@ resource "aws_rds_cluster_parameter_group" "this" {
 }
 
 resource "aws_rds_cluster" "this" {
+  count                           = var.create_database ? 1 : 0
   cluster_identifier              = local.name
   engine                          = "aurora-postgresql"
   engine_mode                     = "provisioned"
   engine_version                  = var.db_engine_version
   database_name                   = var.db_name
   master_username                 = var.db_master_username
-  master_password                 = random_password.db_master_password.result
-  db_subnet_group_name            = aws_db_subnet_group.this.name
-  vpc_security_group_ids          = [aws_security_group.rds.id]
-  db_cluster_parameter_group_name = aws_rds_cluster_parameter_group.this.name
+  master_password                 = random_password.db_master_password[0].result
+  db_subnet_group_name            = aws_db_subnet_group.this[0].name
+  vpc_security_group_ids          = [aws_security_group.rds[0].id]
+  db_cluster_parameter_group_name = aws_rds_cluster_parameter_group.this[0].name
 
   iam_database_authentication_enabled = true
   storage_encrypted                   = true
@@ -61,11 +66,12 @@ resource "aws_rds_cluster" "this" {
 }
 
 resource "aws_rds_cluster_instance" "writer" {
+  count              = var.create_database ? 1 : 0
   identifier         = "${local.name}-writer"
-  cluster_identifier = aws_rds_cluster.this.id
+  cluster_identifier = aws_rds_cluster.this[0].id
   instance_class     = var.db_instance_class
-  engine             = aws_rds_cluster.this.engine
-  engine_version     = aws_rds_cluster.this.engine_version
+  engine             = aws_rds_cluster.this[0].engine
+  engine_version     = aws_rds_cluster.this[0].engine_version
 
   publicly_accessible          = false
   performance_insights_enabled = true
@@ -78,11 +84,12 @@ resource "aws_rds_cluster_instance" "writer" {
 }
 
 resource "aws_rds_cluster_instance" "reader" {
+  count              = var.create_database ? 1 : 0
   identifier         = "${local.name}-reader"
-  cluster_identifier = aws_rds_cluster.this.id
+  cluster_identifier = aws_rds_cluster.this[0].id
   instance_class     = var.db_instance_class
-  engine             = aws_rds_cluster.this.engine
-  engine_version     = aws_rds_cluster.this.engine_version
+  engine             = aws_rds_cluster.this[0].engine
+  engine_version     = aws_rds_cluster.this[0].engine_version
 
   publicly_accessible          = false
   performance_insights_enabled = true

terraform/litellm/aws/redis.tf

@@ -1,6 +1,7 @@
 resource "aws_elasticache_subnet_group" "this" {
+  count      = var.create_redis ? 1 : 0
   name       = "${local.name}-redis"
-  subnet_ids = aws_subnet.private[*].id
+  subnet_ids = local.private_subnet_ids
 
   tags = local.tags
 }
@@ -13,6 +14,7 @@ resource "aws_elasticache_subnet_group" "this" {
 # TLS-protected — the proxy connects via the rediss:// scheme thanks to
 # REDIS_SSL=true in the shared task env (see ecs.tf).
 resource "aws_elasticache_replication_group" "this" {
+  count                = var.create_redis ? 1 : 0
   replication_group_id = "${local.name}-redis"
   description          = "LiteLLM ElastiCache Redis"
 
@@ -23,8 +25,8 @@ resource "aws_elasticache_replication_group" "this" {
   parameter_group_name = "default.redis7"
   port                 = 6379
 
-  subnet_group_name  = aws_elasticache_subnet_group.this.name
-  security_group_ids = [aws_security_group.redis.id]
+  subnet_group_name  = aws_elasticache_subnet_group.this[0].name
+  security_group_ids = [aws_security_group.redis[0].id]
 
   automatic_failover_enabled = var.redis_num_replicas >= 1
   multi_az_enabled           = var.redis_num_replicas >= 1
@@ -35,3 +37,15 @@ resource "aws_elasticache_replication_group" "this" {
 
   tags = local.tags
 }
+
+# Rate limits, budgets, and router cooldowns are shared through Redis. Without
+# it each gateway process counts on its own, so a caller spread across tasks
+# collects the full per-key allowance from every one of them. A `check` rather
+# than a precondition: running without Redis is a legitimate choice when you do
+# not rely on per-key limits, so this warns instead of blocking the plan.
+check "redis_less_rate_limits_are_per_process" {
+  assert {
+    condition     = local.redis_enabled || local.max_gateway_processes <= 1
+    error_message = "No Redis is configured while the gateway can run up to ${local.max_gateway_processes} processes, so per-key RPM/TPM limits, budgets, and cooldowns apply per process and a caller can multiply them across tasks. Set `create_redis = true`, pass `redis_url`, or hold the gateway to one process (`gateway_autoscaling_enabled = false`, `gateway_desired_count = 1`, `gateway_num_workers = 1`)."
+  }
+}

terraform/litellm/aws/secrets.tf

@@ -10,6 +10,7 @@ resource "random_password" "master_key" {
 # user (see rds.tf header). Runtime services authenticate via IAM tokens
 # and never read this secret.
 resource "random_password" "db_master_password" {
+  count       = var.create_database ? 1 : 0
   length      = 32
   special     = false
   min_lower   = 4
@@ -130,6 +131,7 @@ resource "aws_secretsmanager_secret_version" "billing_metrics_ca_cert" {
 }
 
 resource "aws_secretsmanager_secret" "db_master_password" {
+  count                   = var.create_database ? 1 : 0
   name                    = "${local.name}-db-master-password"
   description             = "Aurora master-user password - bootstrap only. Runtime auth is IAM-token."
   recovery_window_in_days = 0
@@ -138,12 +140,50 @@ resource "aws_secretsmanager_secret" "db_master_password" {
 }
 
 resource "aws_secretsmanager_secret_version" "db_master_password" {
-  secret_id = aws_secretsmanager_secret.db_master_password.id
+  count     = var.create_database ? 1 : 0
+  secret_id = aws_secretsmanager_secret.db_master_password[0].id
   secret_string = jsonencode({
     username = var.db_master_username
-    password = random_password.db_master_password.result
-    host     = aws_rds_cluster.this.endpoint
-    port     = aws_rds_cluster.this.port
+    password = random_password.db_master_password[0].result
+    host     = aws_rds_cluster.this[0].endpoint
+    port     = aws_rds_cluster.this[0].port
     dbname   = var.db_name
   })
 }
+
+# Bring-your-own connection strings. Both hold credentials, so they go to
+# Secrets Manager and reach the containers as ECS `secrets` rather than as
+# plain-text env in the task definition.
+resource "aws_secretsmanager_secret" "database_url" {
+  count = local.byo_database ? 1 : 0
+
+  name                    = "${local.name}-database-url"
+  description             = "DATABASE_URL for an existing Postgres, used when create_database = false."
+  recovery_window_in_days = 0
+
+  tags = local.tags
+}
+
+resource "aws_secretsmanager_secret_version" "database_url" {
+  count = local.byo_database ? 1 : 0
+
+  secret_id     = aws_secretsmanager_secret.database_url[0].id
+  secret_string = var.database_url
+}
+
+resource "aws_secretsmanager_secret" "redis_url" {
+  count = local.byo_redis ? 1 : 0
+
+  name                    = "${local.name}-redis-url"
+  description             = "REDIS_URL for an existing Redis, used when create_redis = false."
+  recovery_window_in_days = 0
+
+  tags = local.tags
+}
+
+resource "aws_secretsmanager_secret_version" "redis_url" {
+  count = local.byo_redis ? 1 : 0
+
+  secret_id     = aws_secretsmanager_secret.redis_url[0].id
+  secret_string = var.redis_url
+}

terraform/litellm/aws/tests/byo_infrastructure.tftest.hcl

@@ -0,0 +1,272 @@
+# Plan-only coverage for the four networking/database/cache permutations.
+# `mock_provider` keeps this offline: no AWS credentials, no API calls, no
+# resources. Run from terraform/litellm/aws with `terraform test`.
+
+mock_provider "aws" {
+  # IAM policy documents are validated as JSON by the provider, so the
+  # generated placeholder string has to be replaced with a parsable one.
+  mock_data "aws_iam_policy_document" {
+    defaults = {
+      json = "{\"Version\":\"2012-10-17\",\"Statement\":[]}"
+    }
+  }
+}
+mock_provider "random" {}
+
+variables {
+  region              = "us-east-1"
+  tenant              = "acme"
+  env                 = "test"
+  allow_plaintext_alb = true
+}
+
+run "module_owns_everything_by_default" {
+  command = plan
+
+  variables {
+    azs = ["us-east-1a", "us-east-1b"]
+  }
+
+  assert {
+    condition     = length(aws_vpc.this) == 1 && length(aws_nat_gateway.this) == 1 && length(aws_subnet.private) == 2
+    error_message = "The default path must still create its own VPC, NAT gateway, and one private subnet per AZ."
+  }
+
+  assert {
+    condition     = length(aws_rds_cluster.this) == 1 && length(aws_elasticache_replication_group.this) == 1
+    error_message = "The default path must still create Aurora and ElastiCache."
+  }
+
+  assert {
+    condition     = length(aws_secretsmanager_secret.database_url) == 0 && length(aws_secretsmanager_secret.redis_url) == 0
+    error_message = "Connection-string secrets belong to the bring-your-own path only."
+  }
+
+  assert {
+    condition     = length(local.managed_db_env) == 7 && length(local.managed_redis_env) == 3
+    error_message = "Gateway, backend, and migration tasks must keep the discrete DATABASE_*/REDIS_* env for the module-created stores."
+  }
+
+  assert {
+    condition     = length(terraform_data.bootstrap_db) == 1 && length(aws_ecs_task_definition.migrations) == 1
+    error_message = "The IAM-user bootstrap and the schema migration must both run against the module-created Aurora."
+  }
+}
+
+run "existing_vpc_creates_no_networking" {
+  command = plan
+
+  variables {
+    vpc_id                             = "vpc-00000000000000001"
+    public_subnet_ids                  = ["subnet-pub-a", "subnet-pub-b"]
+    private_subnet_ids                 = ["subnet-priv-a", "subnet-priv-b"]
+    additional_task_security_group_ids = ["sg-caller-owned"]
+  }
+
+  assert {
+    condition = alltrue([
+      length(aws_vpc.this) == 0,
+      length(aws_subnet.public) == 0,
+      length(aws_subnet.private) == 0,
+      length(aws_internet_gateway.this) == 0,
+      length(aws_nat_gateway.this) == 0,
+      length(aws_eip.nat) == 0,
+      length(aws_route_table.public) == 0,
+      length(aws_route_table.private) == 0,
+    ])
+    error_message = "An existing vpc_id must suppress every network resource, including the route tables and NAT gateway."
+  }
+
+  assert {
+    condition     = aws_lb.this.subnets == toset(var.public_subnet_ids)
+    error_message = "The ALB must land in the caller's public subnets."
+  }
+
+  assert {
+    condition = alltrue([
+      aws_db_subnet_group.this[0].subnet_ids == toset(var.private_subnet_ids),
+      aws_elasticache_subnet_group.this[0].subnet_ids == toset(var.private_subnet_ids),
+      aws_ecs_service.gateway.network_configuration[0].subnets == toset(var.private_subnet_ids),
+    ])
+    error_message = "Tasks, Aurora, and ElastiCache must land in the caller's private subnets."
+  }
+
+  assert {
+    condition     = length(local.task_security_group_ids) == 2
+    error_message = "additional_task_security_group_ids must be attached alongside the module's own tasks group."
+  }
+}
+
+run "existing_database_and_redis_replace_the_managed_ones" {
+  command = plan
+
+  variables {
+    azs             = ["us-east-1a", "us-east-1b"]
+    create_database = false
+    database_url    = "postgresql://litellm:pw@db.internal:5432/litellm"
+    create_redis    = false
+    redis_url       = "rediss://:pw@cache.internal:6379"
+  }
+
+  assert {
+    condition = alltrue([
+      length(aws_rds_cluster.this) == 0,
+      length(aws_rds_cluster_instance.writer) == 0,
+      length(aws_db_subnet_group.this) == 0,
+      length(aws_security_group.rds) == 0,
+      length(aws_elasticache_replication_group.this) == 0,
+      length(aws_elasticache_subnet_group.this) == 0,
+      length(aws_security_group.redis) == 0,
+    ])
+    error_message = "Pointing at an existing database and cache must create neither Aurora nor ElastiCache."
+  }
+
+  assert {
+    condition     = length(local.managed_db_env) == 0 && length(local.managed_redis_env) == 0
+    error_message = "The discrete DATABASE_*/REDIS_* env vars must be dropped so DATABASE_URL/REDIS_URL are the only connection targets."
+  }
+
+  assert {
+    condition = alltrue([
+      length([for s in local.shared_secrets : s if s.name == "DATABASE_URL"]) == 1,
+      length([for s in local.shared_secrets : s if s.name == "REDIS_URL"]) == 1,
+    ])
+    error_message = "Both connection strings must reach the containers as Secrets Manager references, not plain-text env."
+  }
+
+  assert {
+    condition     = length(terraform_data.bootstrap_db) == 0 && length(aws_ecs_task_definition.migrations) == 1
+    error_message = "An existing database still needs the schema migration, but not the Aurora IAM-user bootstrap."
+  }
+
+  assert {
+    condition     = length([for e in local.backend_default_env : e if e.name == "STORE_MODEL_IN_DB"]) == 1
+    error_message = "STORE_MODEL_IN_DB must stay set when a database is reachable."
+  }
+}
+
+run "vpc_without_subnets_fails_at_plan" {
+  command = plan
+
+  variables {
+    vpc_id = "vpc-00000000000000001"
+  }
+
+  expect_failures = [
+    aws_lb.this,
+    aws_security_group.tasks,
+  ]
+}
+
+run "neither_vpc_nor_azs_fails_at_plan" {
+  command = plan
+
+  expect_failures = [
+    aws_vpc.this,
+  ]
+}
+
+# Aurora and ElastiCache subnet groups need two AZs, so one private subnet is
+# only enough when neither store is module-created.
+run "one_private_subnet_fails_while_a_managed_store_needs_two_azs" {
+  command = plan
+
+  variables {
+    vpc_id             = "vpc-00000000000000001"
+    public_subnet_ids  = ["subnet-pub-a", "subnet-pub-b"]
+    private_subnet_ids = ["subnet-priv-a"]
+  }
+
+  expect_failures = [
+    aws_security_group.tasks,
+  ]
+}
+
+run "one_private_subnet_is_enough_without_managed_stores" {
+  command = plan
+
+  variables {
+    vpc_id             = "vpc-00000000000000001"
+    public_subnet_ids  = ["subnet-pub-a", "subnet-pub-b"]
+    private_subnet_ids = ["subnet-priv-a"]
+    create_database    = false
+    create_redis       = false
+    # Single process, so the Redis-less rate-limit check stays quiet and this
+    # run is only exercising the subnet rule.
+    gateway_autoscaling_enabled = false
+    gateway_desired_count       = 1
+    gateway_num_workers         = 1
+  }
+
+  assert {
+    condition     = length(aws_security_group.tasks.vpc_id) > 0
+    error_message = "With no module-created database or cache, a single private subnet must plan cleanly."
+  }
+}
+
+# The default sizing is 10 tasks under autoscaling, so a Redis-less stack must
+# warn that per-key limits are counted per process.
+run "redis_less_multi_process_gateway_is_flagged" {
+  command = plan
+
+  variables {
+    azs          = ["us-east-1a", "us-east-1b"]
+    create_redis = false
+  }
+
+  expect_failures = [
+    check.redis_less_rate_limits_are_per_process,
+  ]
+}
+
+run "redis_less_single_process_gateway_is_not_flagged" {
+  command = plan
+
+  variables {
+    azs                         = ["us-east-1a", "us-east-1b"]
+    create_redis                = false
+    gateway_autoscaling_enabled = false
+    gateway_desired_count       = 1
+    gateway_num_workers         = 1
+  }
+
+  assert {
+    condition     = local.max_gateway_processes == 1
+    error_message = "One task with one worker is a single process, which is the supported way to run without Redis."
+  }
+}
+
+run "no_database_and_no_redis_drops_the_schema_migration" {
+  command = plan
+
+  variables {
+    azs             = ["us-east-1a", "us-east-1b"]
+    create_database = false
+    create_redis    = false
+    # Single process, so the Redis-less rate-limit check stays quiet here; it
+    # has its own run above.
+    gateway_autoscaling_enabled = false
+    gateway_desired_count       = 1
+    gateway_num_workers         = 1
+  }
+
+  assert {
+    condition = alltrue([
+      length(aws_ecs_task_definition.migrations) == 0,
+      length(terraform_data.migration) == 0,
+      length(aws_iam_policy.rds_iam_connect) == 0,
+      length(aws_secretsmanager_secret.db_master_password) == 0,
+    ])
+    error_message = "With no database at all there is nothing to migrate, bootstrap, or grant rds-db:connect on."
+  }
+
+  assert {
+    condition     = length(local.backend_default_env) == 0
+    error_message = "STORE_MODEL_IN_DB must not be set without a database to store models in."
+  }
+
+  assert {
+    condition     = length(local.shared_env) == 4
+    error_message = "The shared env must narrow to the S3 bucket and region pair when both data stores are gone."
+  }
+}

terraform/litellm/aws/variables.tf

@@ -74,20 +74,63 @@ variable "ui_password" {
 }
 
 # ---------- Networking ----------
+#
+# Two modes:
+#
+#   1. Module-owned (default, `vpc_id = ""`): the stack creates a VPC, public
+#      and private subnets per AZ, an internet gateway, a NAT gateway, and
+#      the route tables wiring them together. `vpc_cidr` + `azs` drive it.
+#   2. Bring-your-own (`vpc_id` set): the stack creates no networking and
+#      places the ALB in `public_subnet_ids` and every task, plus the Aurora
+#      and ElastiCache subnet groups, in `private_subnet_ids`. `vpc_cidr` and
+#      `azs` are then unused.
+
+variable "vpc_id" {
+  description = <<-EOT
+    Existing VPC to deploy into. Leave empty ("") to have the module create
+    its own VPC, subnets, NAT gateway, and route tables. When set,
+    `public_subnet_ids` and `private_subnet_ids` are required and no
+    networking is created: the private subnets must already have egress
+    (NAT gateway or equivalent) so tasks can reach LLM providers, ECR/GHCR,
+    and Secrets Manager.
+  EOT
+  type        = string
+  default     = ""
+}
+
+variable "public_subnet_ids" {
+  description = "Existing public subnets for the ALB, in at least 2 AZs. Required when `vpc_id` is set, ignored otherwise."
+  type        = list(string)
+  default     = []
+}
+
+variable "private_subnet_ids" {
+  description = "Existing private subnets for the ECS tasks, Aurora, and ElastiCache. Required when `vpc_id` is set, ignored otherwise."
+  type        = list(string)
+  default     = []
+}
+
+variable "additional_task_security_group_ids" {
+  description = <<-EOT
+    Extra security groups to attach to the ECS tasks, on top of the one the
+    module creates. Useful with `vpc_id`: attach a group your existing
+    database or cache already allows inbound from, instead of editing their
+    ingress rules.
+  EOT
+  type        = list(string)
+  default     = []
+}
 
 variable "vpc_cidr" {
-  description = "CIDR block for the VPC."
+  description = "CIDR block for the VPC the module creates. Unused when `vpc_id` is set."
   type        = string
   default     = "10.40.0.0/16"
 }
 
 variable "azs" {
-  description = "Availability zones to spread subnets across. At least 2 required for RDS and ALB."
+  description = "Availability zones to spread the module-created subnets across. At least 2 required for Aurora and the ALB. Unused when `vpc_id` is set."
   type        = list(string)
-  validation {
-    condition     = length(var.azs) >= 2
-    error_message = "Provide at least 2 availability zones."
-  }
+  default     = []
 }
 
 # ---------- Component images ----------
@@ -279,6 +322,34 @@ variable "ui_cpu_target" {
 
 # ---------- RDS ----------
 
+variable "create_database" {
+  description = <<-EOT
+    Create the Aurora Postgres cluster (default). Set false to skip it and
+    either point the stack at an existing database via `database_url`, or
+    run without a database at all when `database_url` is also empty. The
+    DB-less mode drops key management, spend tracking, and the admin UI's
+    persistence: the proxy then serves traffic authenticated by
+    LITELLM_MASTER_KEY only.
+  EOT
+  type        = bool
+  default     = true
+}
+
+variable "database_url" {
+  description = <<-EOT
+    Postgres connection string for an existing database, e.g.
+    `postgresql://user:pass@host:5432/litellm`. Only read when
+    `create_database = false`. Stored in a
+    `<tenant>-litellm-<env>-database-url` Secrets Manager entry and injected
+    into gateway, backend, and the migration task as DATABASE_URL, so the
+    value never lands in a task definition. The schema migration still runs
+    against it on every apply.
+  EOT
+  type        = string
+  default     = ""
+  sensitive   = true
+}
+
 variable "db_instance_class" {
   description = "Aurora instance class for both writer and reader."
   type        = string
@@ -311,6 +382,31 @@ variable "db_username" {
 
 # ---------- Redis ----------
 
+variable "create_redis" {
+  description = <<-EOT
+    Create the ElastiCache Redis replication group (default). Set false to
+    skip it and either point the stack at an existing cache via `redis_url`,
+    or run with no Redis at all when `redis_url` is also empty. Without
+    Redis the proxy loses cross-task state: rate limits, budgets, and the
+    router's cooldowns become per-task instead of cluster-wide.
+  EOT
+  type        = bool
+  default     = true
+}
+
+variable "redis_url" {
+  description = <<-EOT
+    Connection string for an existing Redis, e.g.
+    `rediss://:password@host:6379`. Only read when `create_redis = false`.
+    Stored in a `<tenant>-litellm-<env>-redis-url` Secrets Manager entry and
+    injected as REDIS_URL, which takes precedence over REDIS_HOST/REDIS_PORT
+    in the proxy.
+  EOT
+  type        = string
+  default     = ""
+  sensitive   = true
+}
+
 variable "redis_node_type" {
   description = "ElastiCache node type."
   type        = string

tests/audio_tests/test_whisper.py

@@ -160,25 +160,6 @@ async def test_whisper_log_pre_call():
         mock_log_pre_call.assert_called_once()
 
 
-@pytest.mark.asyncio
-async def test_whisper_log_pre_call():
-    from litellm.litellm_core_utils.litellm_logging import Logging
-    from datetime import datetime
-    from unittest.mock import patch, MagicMock
-    from litellm.integrations.custom_logger import CustomLogger
-
-    custom_logger = CustomLogger()
-
-    litellm.callbacks = [custom_logger]
-
-    with patch.object(custom_logger, "log_pre_api_call") as mock_log_pre_call:
-        await litellm.atranscription(
-            model="whisper-1",
-            file=_audio_file(),
-        )
-        mock_log_pre_call.assert_called_once()
-
-
 @pytest.mark.asyncio
 async def test_gpt_4o_transcribe():
     from litellm.litellm_core_utils.litellm_logging import Logging

tests/batches_tests/test_hosted_vllm_batches_and_files.py

@@ -1,105 +0,0 @@
-"""
-Unit Tests for hosted_vllm Batches and Files API
-
-Tests the integration of hosted_vllm provider with LiteLLM's batch and file operations.
-Tests against a real OpenAI-compatible endpoint.
-"""
-
-import json
-import os
-import sys
-import time
-import uuid
-
-import httpx
-import pytest
-from dotenv import load_dotenv
-
-load_dotenv()
-sys.path.insert(0, os.path.abspath("../.."))
-
-import litellm
-
-
-SERVER_URL = "https://exampleopenaiendpoint-production-0ee2.up.railway.app/v1"
-
-
-@pytest.mark.asyncio()
-@pytest.mark.skip(reason="Local only test")
-async def test_hosted_vllm_full_workflow():
-    """
-    Test the complete workflow: create file -> create batch -> retrieve batch -> retrieve file.
-    Tests against real OpenAI-compatible endpoint.
-    """
-    litellm._turn_on_debug()
-    file_name = "openai_batch_completions.jsonl"
-    _current_dir = os.path.dirname(os.path.abspath(__file__))
-    file_path = os.path.join(_current_dir, file_name)
-
-    # Step 1: Create file
-    print("\n=== Step 1: Creating file ===")
-    file_obj = await litellm.acreate_file(
-        file=open(file_path, "rb"),
-        purpose="batch",
-        custom_llm_provider="hosted_vllm",
-        api_base=SERVER_URL,
-        api_key="test-api-key",
-    )
-
-    print(f"✓ Created file: {file_obj.id}")
-    assert file_obj.id is not None
-    assert file_obj.object == "file"
-    assert file_obj.purpose == "batch"
-
-    # Step 2: Create batch
-    print("\n=== Step 2: Creating batch ===")
-    batch_obj = await litellm.acreate_batch(
-        completion_window="24h",
-        endpoint="/v1/chat/completions",
-        input_file_id=file_obj.id,
-        custom_llm_provider="hosted_vllm",
-        metadata={"test": "hosted_vllm_integration"},
-        api_base=SERVER_URL,
-        api_key="test-api-key",
-    )
-
-    print(f"✓ Created batch: {batch_obj.id}")
-    print(f"  Status: {batch_obj.status}")
-    print(f"  Input file: {batch_obj.input_file_id}")
-    assert batch_obj.id is not None
-    assert batch_obj.object == "batch"
-    assert batch_obj.input_file_id == file_obj.id
-    assert batch_obj.endpoint == "/v1/chat/completions"
-
-    # Step 3: Retrieve batch
-    print("\n=== Step 3: Retrieving batch ===")
-    retrieved_batch = await litellm.aretrieve_batch(
-        batch_id=batch_obj.id,
-        custom_llm_provider="hosted_vllm",
-        api_base=SERVER_URL,
-        api_key="test-api-key",
-    )
-
-    print(f"✓ Retrieved batch: {retrieved_batch.id}")
-    print(f"  Status: {retrieved_batch.status}")
-    print(f"  Output file: {retrieved_batch.output_file_id}")
-    assert retrieved_batch.id == batch_obj.id
-    assert retrieved_batch.object == "batch"
-    assert retrieved_batch.input_file_id == file_obj.id
-
-    # Step 4: Retrieve file (verify file still accessible)
-    print("\n=== Step 4: Retrieving original file ===")
-    retrieved_file = await litellm.afile_retrieve(
-        file_id=file_obj.id,
-        custom_llm_provider="hosted_vllm",
-        api_base=SERVER_URL,
-        api_key="test-api-key",
-    )
-
-    print(f"✓ Retrieved file: {retrieved_file.id}")
-    print(f"  Filename: {retrieved_file.filename}")
-    print(f"  Bytes: {retrieved_file.bytes}")
-    assert retrieved_file.id == file_obj.id
-    assert retrieved_file.object == "file"
-
-    print("\n✅ Full workflow test completed successfully!")

tests/e2e/claude_code/_builder_unit_tests/test_matrix_builder.py

@@ -0,0 +1,148 @@
+"""Unit tests for `find_regressions`, the green→red detector that gates
+auto-merge on the daily compat-matrix docs PR (see `cron_vm/`).
+
+Markerless harness tests: they exercise publisher plumbing, not a product
+feature, so they run without a proxy and carry no `e2e` marker.
+"""
+
+from __future__ import annotations
+
+from typing import Mapping, Union
+
+from claude_code.matrix_builder import find_regressions
+
+_CellSpec = Union[str, Mapping[str, str]]
+
+
+def _matrix(
+    cells: Mapping[tuple[str, str], _CellSpec],
+    *,
+    names: Mapping[str, str] | None = None,
+) -> dict[str, object]:
+    """Build a minimal matrix dict from a {(feature_id, provider): status}
+    or {(feature_id, provider): cell_dict} mapping."""
+    names = names or {}
+    features: dict[str, dict[str, dict[str, str]]] = {}
+    for (feature_id, provider), value in cells.items():
+        cell = {"status": value} if isinstance(value, str) else dict(value)
+        features.setdefault(feature_id, {})[provider] = cell
+    return {
+        "features": [
+            {
+                "id": feature_id,
+                "name": names.get(feature_id, feature_id.upper()),
+                "providers": providers,
+            }
+            for feature_id, providers in features.items()
+        ]
+    }
+
+
+def test_find_regressions_flags_pass_to_fail() -> None:
+    old = _matrix({("vision", "anthropic"): "pass"})
+    new = _matrix(
+        {("vision", "anthropic"): {"status": "fail", "error": "credit balance too low"}}
+    )
+    regressions = find_regressions(old, new)
+    assert len(regressions) == 1
+    r = regressions[0]
+    assert r["feature_id"] == "vision"
+    assert r["provider"] == "anthropic"
+    assert r["old_status"] == "pass"
+    assert r["new_status"] == "fail"
+    assert r["error"] == "credit balance too low"
+
+
+def test_find_regressions_ignores_red_to_red() -> None:
+    """An already-failing cell that stays failing is NOT a regression — a
+    provider that's independently broken (e.g. out of credits) must not
+    block the daily auto-merge forever."""
+    old = _matrix({("vision", "anthropic"): "fail"})
+    new = _matrix({("vision", "anthropic"): "fail"})
+    assert find_regressions(old, new) == []
+
+
+def test_find_regressions_ignores_improvements_and_steady_green() -> None:
+    old = _matrix(
+        {
+            ("vision", "anthropic"): "fail",  # red -> green
+            ("tool_use", "azure"): "pass",  # green -> green
+        }
+    )
+    new = _matrix(
+        {
+            ("vision", "anthropic"): "pass",
+            ("tool_use", "azure"): "pass",
+        }
+    )
+    assert find_regressions(old, new) == []
+
+
+def test_find_regressions_ignores_green_to_grey() -> None:
+    """green→not_tested / green→not_applicable are degradations but not
+    *red* regressions; we deliberately don't block on them."""
+    old = _matrix(
+        {
+            ("vision", "azure"): "pass",
+            ("tool_use", "azure"): "pass",
+        }
+    )
+    new = _matrix(
+        {
+            ("vision", "azure"): "not_tested",
+            ("tool_use", "azure"): {"status": "not_applicable", "reason": "skip"},
+        }
+    )
+    assert find_regressions(old, new) == []
+
+
+def test_find_regressions_ignores_new_cells_without_baseline() -> None:
+    """A cell only present in the new matrix (new feature/provider) has no
+    baseline, so a fail there can't be a regression."""
+    old = _matrix({("vision", "anthropic"): "pass"})
+    new = _matrix(
+        {
+            ("vision", "anthropic"): "pass",
+            ("brand_new_feature", "anthropic"): "fail",
+        }
+    )
+    assert find_regressions(old, new) == []
+
+
+def test_find_regressions_matches_by_id_not_name() -> None:
+    """Renaming a feature's display name must not hide a regression: cells
+    are matched on the stable id."""
+    old = _matrix({("thinking", "anthropic"): "pass"}, names={"thinking": "Old Name"})
+    new = _matrix(
+        {("thinking", "anthropic"): "fail"}, names={"thinking": "Totally New Name"}
+    )
+    regressions = find_regressions(old, new)
+    assert len(regressions) == 1
+    assert regressions[0]["feature_id"] == "thinking"
+    assert regressions[0]["feature_name"] == "Totally New Name"
+
+
+def test_find_regressions_reports_multiple_sorted() -> None:
+    old = _matrix(
+        {
+            ("vision", "anthropic"): "pass",
+            ("tool_use", "anthropic"): "pass",
+            ("vision", "azure"): "pass",
+        }
+    )
+    new = _matrix(
+        {
+            ("vision", "anthropic"): "fail",
+            ("tool_use", "anthropic"): "fail",
+            ("vision", "azure"): "pass",  # stays green
+        }
+    )
+    regressions = find_regressions(old, new)
+    keys = [(r["feature_id"], r["provider"]) for r in regressions]
+    assert keys == [("tool_use", "anthropic"), ("vision", "anthropic")]
+
+
+def test_find_regressions_empty_old_matrix_is_safe() -> None:
+    """No baseline at all (first publish) yields no regressions."""
+    new = _matrix({("vision", "anthropic"): "fail"})
+    assert find_regressions({}, new) == []

tests/e2e/claude_code/cron_vm/README.md

@@ -0,0 +1,195 @@
+# Cron VM setup for the Claude Code compatibility-matrix populator
+
+The populator runs daily on a dedicated GCP VM
+(`litellm-compatibility-matrix-populator`) rather than as a GitHub
+Action. Trade-offs:
+
+- ✅ Real VM means we can `gh auth login` against an account that's
+  already a collaborator on `BerriAI/litellm-docs`, instead of
+  provisioning a GitHub App with `pull-requests: write`.
+- ✅ Persistent state (a single `~/litellm-cron-worktree/` and its `.venv`)
+  is reused across runs, so each daily run does a fast `git checkout` +
+  incremental `uv sync` rather than a fresh clone + cold sync.
+- ✅ No Docker dependency — the proxy runs directly via `uv run litellm`.
+- ⚠️ The VM has to actually be on. systemd's `Persistent=true` recovers
+  from short outages, but a multi-day outage means the matrix goes
+  stale until the VM is back.
+- ⚠️ Provider credentials live on the VM filesystem
+  (`/etc/litellm-compat-matrix.env`) instead of GitHub secrets. Treat
+  the VM as an environment with comparable blast radius to a CI runner.
+
+This directory used to live at `tests/claude_code/cron_vm/` (paired with
+the standalone `tests/claude_code/` suite); it now runs the maintained
+`tests/e2e/claude_code/` suite instead. The pytest env interface changed
+accordingly: the runner exports `LITELLM_PROXY_URL` / `LITELLM_MASTER_KEY`
+(previously `LITELLM_PROXY_BASE_URL` / `LITELLM_PROXY_API_KEY`), the azure
+column reads `AZURE_AI_API_KEY` / `AZURE_AI_API_BASE` (previously
+`AZURE_FOUNDRY_*`), and the GPT columns need `OPENAI_API_KEY` and
+`AZURE_API_BASE` / `AZURE_API_KEY` — see `litellm-compat-matrix.env.example`.
+
+## Layout
+
+| File | Purpose |
+| --- | --- |
+| `run_daily.sh` | The actual cron job. Resolves versions, updates the worktree, boots the proxy, runs pytest, builds the JSON, opens (or updates) a docs PR, sweeps stale compat-matrix PRs. |
+| `build_matrix.py` | Tiny Python CLI that wraps `claude_code.matrix_builder.build_from_paths`. Exists only because the bash script needs *some* way to render the per-cell aggregation, and the builder is already Python. |
+| `check_regressions.py` | Tiny Python CLI that wraps `claude_code.matrix_builder.find_regressions`. Diffs the freshly built matrix against the currently-published one and exits `3` if any cell flipped green→red, which gates auto-merge. |
+| `litellm-compat-matrix.service` | systemd oneshot that invokes `run_daily.sh`. |
+| `litellm-compat-matrix.timer` | `OnCalendar=*-*-* 06:00:00 UTC`, `Persistent=true`. |
+| `litellm-compat-matrix.env.example` | Template for `/etc/litellm-compat-matrix.env`. |
+
+## What `run_daily.sh` does
+
+1. **Resolves the latest LiteLLM final release tag** (newest bare
+   `vX.Y.Z`, skipping `-rc.N`/`-dev.N` pre-releases) by paging the
+   GitHub Releases API (`curl | jq`).
+2. **Reads the local Claude Code CLI version** via `claude --version`.
+   The cron does not auto-upgrade the CLI — operators do that
+   out-of-band by running `npm install -g @anthropic-ai/claude-code@latest`.
+3. **Updates the persistent worktree** at `~/litellm-cron-worktree/`:
+   `git fetch --tags --force`, `git reset --hard`,
+   `git clean -fdx -e .venv -e .uv-bin`, `git checkout --force <tag>`.
+   The `.venv` is preserved across runs so `uv sync --frozen` is
+   incremental. Then **shims the test suite**: `tests/e2e/` in the
+   worktree is rebuilt from the dev checkout — the `claude_code/` suite
+   plus the five shared transport helpers it imports (`proxy_client.py`,
+   `e2e_http.py`, `models.py`, `e2e_config.py`, `transport.py`) — so the
+   cron always runs *today's* tests against the latest stable proxy. The
+   tag's own `tests/e2e/` tree (including the EKS-harness `conftest.py`,
+   whose imports the stable venv doesn't install) is deliberately not
+   used.
+4. **Boots the proxy** as a `setsid` background process on port `4100`
+   (so it can't collide with a developer's `:4000`), then polls
+   `/health/liveliness` until it's up.
+5. **Runs pytest** on `tests/e2e/claude_code/` with `LITELLM_PROXY_URL`
+   pointed at the proxy and `COMPAT_RESULTS_PATH` set so the conftest
+   hook writes the per-test results artifact. Test failures become
+   `fail` cells in the JSON, not script errors.
+6. **Builds `compatibility-matrix.json`** by handing the artifact +
+   manifest to `build_matrix.py`.
+7. **Opens or updates a docs PR**: `gh repo clone` of `litellm-docs`
+   into a tempdir, deterministic head branch
+   (`compat-matrix/<litellm-version>-<claude-code-version>-<UTC-date>`),
+   `--force` push **directly to `BerriAI/litellm-docs`** (the
+   `mateo-berri` token has write access, so this is a same-repo branch,
+   not a fork), `gh pr create`. A re-run on the same day fast-forwards
+   the existing branch and `gh pr create` no-ops ("a pull request for
+   branch ... already exists" is treated as success). These PRs are no
+   longer gated on a second human review.
+8. **Gates auto-merge on a regression check**: before enabling
+   auto-merge, `check_regressions.py` diffs the new matrix against the
+   one currently on `main`. Auto-merge (`gh pr merge --auto --squash`)
+   is only enabled when **no cell flipped green→red** — i.e. every
+   transition is red→green, green→green, or red→red. A pre-existing red
+   cell (e.g. a provider that's out of API credits) is `red→red` and
+   does **not** block; only a `pass`→`fail` flip does. When a regression
+   is detected the PR is still opened/updated (with a warning banner
+   naming the offending cells) but auto-merge is left **off** — and any
+   auto-merge a prior same-day run enabled is explicitly disabled — so a
+   human reviews before it lands on the public table. The check fails
+   *closed*: if it errors, auto-merge is withheld.
+9. **Sweeps stale compat-matrix PRs**: once today's PR exists, every
+   other open `compat-matrix/*` PR on the docs repo is closed (and its
+   bot-owned branch deleted), so at most one compat-matrix PR is ever
+   open — the newest.
+
+## One-time VM setup
+
+Run as `mateo` on the cron VM:
+
+```bash
+# 1. Toolchain
+sudo apt-get update
+sudo apt-get install -y git nodejs npm jq curl
+curl -LsSf https://astral.sh/uv/install.sh | sh
+sudo apt-get install -y gh   # or follow https://cli.github.com/
+
+# 2. Claude Code CLI (the cron does NOT auto-upgrade this; rerun this
+#    line out-of-band when you want a fresh CLI to be tested)
+sudo npm install -g @anthropic-ai/claude-code@latest
+
+# 3. Litellm checkout. Used by systemd's WorkingDirectory and as the
+#    source of the .service / .timer files. The cron itself runs out
+#    of the separate worktree at ~/litellm-cron-worktree/.
+mkdir -p ~/litellm
+git clone https://github.com/BerriAI/litellm.git ~/litellm/litellm
+git -C ~/litellm/litellm checkout litellm_internal_staging
+
+# 4. gh auth — must be a collaborator on BerriAI/litellm-docs.
+gh auth login   # follow prompts; pick HTTPS + token paste flow
+
+# 5. Provider credentials + the publish token.
+sudo cp ~/litellm/litellm/tests/e2e/claude_code/cron_vm/litellm-compat-matrix.env.example \
+        /etc/litellm-compat-matrix.env
+sudoedit /etc/litellm-compat-matrix.env   # fill in real values
+sudo chmod 0600 /etc/litellm-compat-matrix.env
+# The mateo-berri PAT lives in its own file, mapped into the service via
+# systemd LoadCredential so it stays out of the test processes' env
+# (see the env.example comment for why).
+sudo install -m 0600 /dev/null /etc/litellm-compat-matrix-github-token
+sudoedit /etc/litellm-compat-matrix-github-token   # single line: the PAT
+
+# 6. systemd units.
+sudo cp ~/litellm/litellm/tests/e2e/claude_code/cron_vm/litellm-compat-matrix.service /etc/systemd/system/
+sudo cp ~/litellm/litellm/tests/e2e/claude_code/cron_vm/litellm-compat-matrix.timer   /etc/systemd/system/
+sudo systemctl daemon-reload
+sudo systemctl enable --now litellm-compat-matrix.timer
+```
+
+## Operating it
+
+```bash
+# When does it run next?
+systemctl list-timers litellm-compat-matrix.timer
+
+# Trigger a real run right now (PRs to litellm-docs).
+sudo systemctl start litellm-compat-matrix.service
+
+# Trigger a run that does NOT open a PR (good for first-time validation).
+SKIP_PUBLISH=1 ~/litellm/litellm/tests/e2e/claude_code/cron_vm/run_daily.sh
+
+# Narrow to one cell while debugging.
+SKIP_PUBLISH=1 PYTEST_K='basic_messaging_non_streaming and anthropic' \
+  ~/litellm/litellm/tests/e2e/claude_code/cron_vm/run_daily.sh
+
+# Watch the most recent run.
+journalctl -u litellm-compat-matrix.service -f
+
+# Read older runs.
+journalctl -u litellm-compat-matrix.service --since '2 days ago'
+
+# Disable until further notice (e.g. while debugging).
+sudo systemctl disable --now litellm-compat-matrix.timer
+```
+
+## Gotchas
+
+- **The venv is pinned to Python 3.12 (`CRON_PYTHON_VERSION`).** The
+  e2e suite uses PEP 695 `type` aliases, which the VM's system Python
+  (3.11) can't parse; `run_daily.sh` has uv fetch a managed CPython
+  into `~/litellm-cron-worktree/.uv-python/` and syncs the venv against
+  it. The first run after a version bump is a cold venv rebuild.
+- **The proxy port is `4100`, not `4000`.** This is so a developer SSH'd
+  into the same VM with their own `:4000` proxy doesn't collide with a
+  cron run. Override with `PROXY_PORT=...` in `/etc/litellm-compat-matrix.env`
+  if you need to.
+- **`uv sync --frozen` requires the resolved tag to be tagged on
+  GitHub.** If the latest stable release was made but not pushed as a
+  git tag, the `git checkout` step fails. Push the tag, then rerun.
+- **Publish-token rotation is your problem.** The cron does not
+  refresh the token; if `mateo-berri`'s PAT in
+  `/etc/litellm-compat-matrix-github-token` expires, the run fails at
+  the `git push`/`gh pr create` step with a 401 ("Bad credentials" /
+  "Authentication failed"). Mint a fresh PAT and update that file.
+  The token needs write access to `BerriAI/litellm-docs` (classic
+  `repo` scope, or fine-grained Contents:RW + Pull requests:RW). It is
+  delivered via systemd `LoadCredential`, not the env file, so pytest,
+  the proxy, and the claude CLI never inherit it; manual runs export
+  `GITHUB_TOKEN` instead.
+- **First run after upgrading the Claude Code CLI is the riskiest one.**
+  If the new CLI changes its wire format the matrix run can produce
+  systematic failures. Always run with `SKIP_PUBLISH=1` after a CLI
+  upgrade before letting the next scheduled fire happen.
+- **Disk:** the worktree's `.venv` is ~1.3 GB and the `.git` directory
+  is ~1 GB. Plan for at least 5 GB free on the VM, otherwise
+  `uv sync` will fail mid-run and leave you with a half-installed venv.

tests/e2e/claude_code/cron_vm/build_matrix.py

@@ -0,0 +1,52 @@
+"""Tiny CLI wrapper around `claude_code.matrix_builder.build_from_paths`.
+
+Exists only so `run_daily.sh` can hand the version metadata + paths into
+the matrix builder without re-implementing it in bash. All real logic
+lives in `matrix_builder.py`.
+
+The suite imports its own modules with `tests/e2e/` on sys.path (that is
+how pytest resolves them: `tests/e2e/` has no `__init__.py`, while
+`claude_code/` does), so this script bootstraps the same root — two
+levels up from this file — before importing.
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+
+from claude_code.matrix_builder import (
+    build_from_paths,
+)  # noqa: E402  # needs the sys.path bootstrap above
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--manifest", type=Path, required=True)
+    parser.add_argument("--results", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--litellm-version", required=True)
+    parser.add_argument("--claude-code-version", required=True)
+    args = parser.parse_args()
+
+    generated_at = datetime.datetime.now(datetime.timezone.utc).strftime(
+        "%Y-%m-%dT%H:%M:%SZ"
+    )
+    build_from_paths(
+        manifest_path=args.manifest,
+        results_path=args.results,
+        litellm_version=args.litellm_version,
+        claude_code_version=args.claude_code_version,
+        generated_at=generated_at,
+        output_path=args.output,
+    )
+    print(f"wrote {args.output}")  # noqa: T201  # CLI output read by run_daily.sh
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/e2e/claude_code/cron_vm/check_regressions.py

@@ -0,0 +1,80 @@
+"""CLI: detect green→red regressions between the published matrix and a
+freshly built one, so `run_daily.sh` can decide whether to enable
+auto-merge on the daily docs PR.
+
+All real logic lives in `claude_code.matrix_builder.find_regressions`;
+this file only does the I/O and maps the result onto an exit code the
+bash caller can branch on.
+
+Exit codes (the bash gate depends on these exact values):
+
+  0  no green→red regressions          -> safe to auto-merge
+  3  one or more green→red regressions  -> do NOT auto-merge (human review)
+  2  argparse/usage error (argparse default)
+
+The `--old` file is allowed to be missing: on the first-ever publish there
+is no baseline to regress against, so we exit 0.
+
+Imports resolve with `tests/e2e/` on sys.path, mirroring build_matrix.py.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+
+from claude_code.matrix_builder import (
+    find_regressions,
+)  # noqa: E402  # needs the sys.path bootstrap above
+
+REGRESSION_EXIT = 3
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--old",
+        type=Path,
+        required=True,
+        help="currently published matrix JSON (may be absent on first publish)",
+    )
+    parser.add_argument(
+        "--new",
+        type=Path,
+        required=True,
+        help="freshly built matrix JSON",
+    )
+    args = parser.parse_args()
+
+    if not args.old.exists():
+        print(  # noqa: T201  # CLI output read by run_daily.sh
+            "no published matrix to compare against "
+            "(first publish); treating as no regressions"
+        )
+        return 0
+
+    old_matrix = json.loads(args.old.read_text())
+    new_matrix = json.loads(args.new.read_text())
+
+    regressions = find_regressions(old_matrix, new_matrix)
+    if not regressions:
+        print("no green->red regressions detected")  # noqa: T201  # CLI output
+        return 0
+
+    print(  # noqa: T201  # CLI output read by run_daily.sh
+        f"detected {len(regressions)} green->red regression(s):"
+    )
+    for r in regressions:
+        line = f"  - {r['feature_name']} [{r['provider']}]: pass -> fail"
+        if r["error"]:
+            line += f" ({r['error'][:160]})"
+        print(line)  # noqa: T201  # CLI output read by run_daily.sh
+    return REGRESSION_EXIT
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/e2e/claude_code/cron_vm/litellm-compat-matrix.env.example

@@ -0,0 +1,68 @@
+# Environment file consumed by `litellm-compat-matrix.service`.
+#
+# Install at `/etc/litellm-compat-matrix.env` and chmod 0600.
+# `EnvironmentFile=-` in the unit means the service is allowed to start
+# even if this file is missing, but the populator will fail at the
+# first provider request without these credentials.
+
+# Anthropic
+ANTHROPIC_API_KEY=
+
+# Bedrock (invoke + converse columns; also bedrock_mantle when enabled).
+# Use Anthropic's Bedrock API-key passthrough (long-lived bearer token).
+# No AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY required for the matrix --
+# both the LiteLLM invoke and converse routes pick up
+# AWS_BEARER_TOKEN_BEDROCK when present.
+AWS_BEARER_TOKEN_BEDROCK=
+AWS_REGION_NAME=us-east-1
+
+# Vertex AI (vertex_ai + vertex_ai_gpt columns).
+# On the GCP VM, the default service-account ADC from the metadata server
+# is used -- no JSON key file is needed. If you ever need to run outside
+# GCP, also export GOOGLE_APPLICATION_CREDENTIALS=/path/to/sa.json.
+VERTEXAI_PROJECT=
+VERTEXAI_LOCATION=global
+
+# Azure AI Foundry (azure column — Claude models on Foundry)
+AZURE_AI_API_KEY=
+AZURE_AI_API_BASE=
+
+# OpenAI (openai GPT column)
+OPENAI_API_KEY=
+
+# Azure OpenAI (azure_openai GPT column)
+AZURE_API_BASE=
+AZURE_API_KEY=
+
+# The publish PAT (mateo-berri, write access on BerriAI/litellm-docs)
+# deliberately does NOT live in this file. Everything here lands in the
+# process environment of pytest, the proxy, and the model-driven claude
+# CLI, where any same-UID reader can lift it from /proc/<pid>/environ.
+# Instead, install the token at /etc/litellm-compat-matrix-github-token
+# (chmod 0600, single line); the service maps it in via systemd
+# LoadCredential and run_daily.sh keeps it out of every child process
+# env. Used to (a) resolve the latest stable release, (b) push the
+# daily compat-matrix branch directly to BerriAI/litellm-docs, (c) open
+# the same-repo PR, and (d) enable squash auto-merge on it. Scopes:
+# classic `repo` + `workflow`, or fine-grained on BerriAI/litellm-docs
+# with Contents:RW + Pull requests:RW + Workflows:RW.
+# Manual runs export GITHUB_TOKEN instead, or skip publishing entirely
+# with SKIP_PUBLISH=1 (only writes the matrix JSON locally).
+
+# Optional: the bedrock_mantle column is opt-in because the AWS account
+# needs the Mantle (OpenAI-on-Bedrock) models enabled. Without this the
+# mantle cells are skipped and recorded as not_tested rather than fail.
+# COMPAT_MANTLE_CELLS=1
+
+# Optional: the openai column is likewise opt-in; its cells hit CLI
+# timeouts under the concurrent stage suite, but the serial cron can
+# usually run them. Skipped cells are recorded as not_tested.
+# COMPAT_OPENAI_GPT_CELLS=1
+
+# Optional overrides; defaults are sensible for the cron VM.
+# PROXY_PORT=4100
+# LITELLM_WORKTREE=/home/mateo/litellm-cron-worktree
+# DOCS_REPO=BerriAI/litellm-docs
+# DOCS_BRANCH=main
+# DOCS_TARGET_PATH=src/data/compatibility-matrix.json
+# AUTO_MERGE_METHOD=squash

tests/e2e/claude_code/cron_vm/litellm-compat-matrix.service

@@ -0,0 +1,113 @@
+# systemd service for the Claude Code compatibility-matrix populator.
+#
+# Triggered by `litellm-compat-matrix.timer`; not started directly. The
+# unit is a `Type=oneshot` so the timer's `OnCalendar=` semantics
+# describe "run once per day" cleanly — there's no long-lived daemon to
+# supervise; each invocation runs the populator end-to-end and exits.
+#
+# Install
+# -------
+#
+#   sudo cp tests/e2e/claude_code/cron_vm/litellm-compat-matrix.service /etc/systemd/system/
+#   sudo cp tests/e2e/claude_code/cron_vm/litellm-compat-matrix.timer   /etc/systemd/system/
+#   sudo systemctl daemon-reload
+#   sudo systemctl enable --now litellm-compat-matrix.timer
+#
+# Paths are hard-coded to /home/mateo rather than using systemd's %h
+# specifier. Why: in *system* units (this one), %h is expanded at
+# parse time against the *manager's* home -- which is /root for PID 1
+# -- and *not* against the User= directive. That mismatch makes
+# ReadWritePaths point at /root/.cache (which doesn't exist), causing
+# the namespace setup to fail with status=226/NAMESPACE before the
+# script ever runs. The runtime user (`User=mateo`) must:
+#
+#   * have a checkout of `BerriAI/litellm` at `~/litellm/litellm` so the
+#     publisher module is importable;
+#   * have a uv venv at `~/litellm/litellm/.venv` (created by
+#     `uv sync --frozen` inside that checkout once);
+#   * have `gh` already authenticated against an account with
+#     `pull-requests: write` on `BerriAI/litellm-docs`;
+#   * have provider credentials exported in `/etc/litellm-compat-matrix.env`
+#     (see `litellm-compat-matrix.env.example` in this directory);
+#   * have the mateo-berri publish PAT at
+#     `/etc/litellm-compat-matrix-github-token` (chmod 0600, single
+#     line), delivered via `LoadCredential=` below.
+
+[Unit]
+Description=Claude Code compatibility-matrix populator (oneshot)
+Documentation=file:///home/mateo/litellm/litellm/tests/e2e/claude_code/cron_vm/README.md
+Wants=network-online.target
+After=network-online.target
+
+[Service]
+Type=oneshot
+User=mateo
+Group=mateo
+
+# Provider credentials + any gh/PROXY_PORT overrides live here. Format
+# is the standard `KEY=value` one line per env var.
+EnvironmentFile=-/etc/litellm-compat-matrix.env
+
+# The mateo-berri publish PAT is mapped in via the credential store, NOT
+# the EnvironmentFile, so it never lands in the process environment that
+# pytest, the proxy, and the model-driven claude CLI inherit (any
+# same-UID process can read /proc/<pid>/environ). run_daily.sh reads
+# ${CREDENTIALS_DIRECTORY}/github-token and hands it to gh per call.
+# Unlike EnvironmentFile= above, this is deliberately NOT optional: a
+# missing token file fails the unit at start instead of 30 minutes in.
+LoadCredential=github-token:/etc/litellm-compat-matrix-github-token
+
+# systemd starts with a minimal PATH (~/usr/local/bin:/usr/bin:/bin).
+# `uv` and `claude` are installed under the runtime user's `~/.local/bin`
+# so we have to prepend it explicitly; otherwise run_daily.sh fails at
+# the up-front command-presence check.
+Environment=PATH=/home/mateo/.local/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+
+# `HOME` is auto-set to /home/mateo when User=mateo is honored, but be
+# explicit so anything that reads $HOME (e.g. uv's cache lookup, the
+# claude CLI's per-session dir) sees the right value even if a future
+# refactor flips DynamicUser= or PrivateUsers= on.
+Environment=HOME=/home/mateo
+
+WorkingDirectory=/home/mateo/litellm/litellm
+
+ExecStart=/home/mateo/litellm/litellm/tests/e2e/claude_code/cron_vm/run_daily.sh
+
+# 90 minutes is generous: cold runs do `git clone` + `uv sync` of a new
+# tag's lockfile, which can take a couple of minutes on a 2-vCPU VM,
+# plus the full feature x provider grid of pytest cells hitting several
+# cloud providers.
+TimeoutStartSec=90min
+
+# A failed run shouldn't restart automatically — the next timer fire is
+# the right retry. Reruns of the same day's matrix are idempotent.
+Restart=no
+
+# Security hardening: the populator only reads the litellm checkout and
+# the env-file; everything else it writes lives in either the worktree
+# (managed) or `/tmp` (cleaned up by tempfile).
+#
+# ReadWritePaths whitelist:
+#   * litellm-cron-worktree    - the long-lived stable-tag checkout +
+#                                its `.venv` (`uv sync` rewrites every
+#                                run) + `.uv-bin` (pinned `uv` binary
+#                                cache).
+#   * .cache                   - uv's wheel cache (~/.cache/uv) so we
+#                                don't redownload pinned deps each run.
+#   * .claude                  - `claude` CLI's per-session state under
+#                                `~/.claude/projects/<sha>/`; created
+#                                on every `claude --print` invocation.
+#   * .config/gh               - `gh` CLI host config; technically not
+#                                needed when we pass GH_TOKEN inline,
+#                                but cheap to whitelist and prevents
+#                                future regressions if a code path
+#                                ever falls back to the host config.
+#   * /tmp                     - mktemp -d workdir + proxy logs.
+NoNewPrivileges=true
+ProtectSystem=strict
+ProtectHome=read-only
+ReadWritePaths=/home/mateo/litellm-cron-worktree /home/mateo/.cache /home/mateo/.claude /home/mateo/.config/gh /tmp
+PrivateTmp=true
+
+[Install]
+WantedBy=multi-user.target

tests/e2e/claude_code/cron_vm/litellm-compat-matrix.timer

@@ -0,0 +1,25 @@
+# Daily timer for the compatibility-matrix populator.
+#
+# 06:00 UTC matches the original GitHub Actions cron schedule; chosen so
+# operators in US/EU timezones see fresh PRs at the start of their work
+# day.
+#
+# `Persistent=true` causes a missed run (VM was off / suspended) to
+# fire the next time the timer is started, which is the property we
+# want for a once-a-day job: the matrix should refresh as soon as the
+# VM is reachable again, not wait another 24h.
+#
+# `RandomizedDelaySec=10min` smears load if multiple matrix-style
+# pipelines are ever colocated on the same VM in the future.
+
+[Unit]
+Description=Run the Claude Code compatibility-matrix populator daily
+
+[Timer]
+OnCalendar=*-*-* 06:00:00 UTC
+Persistent=true
+RandomizedDelaySec=10min
+Unit=litellm-compat-matrix.service
+
+[Install]
+WantedBy=timers.target

tests/e2e/claude_code/cron_vm/run_daily.sh

@@ -0,0 +1,672 @@
+#!/usr/bin/env bash
+# Daily Claude Code compatibility-matrix populator.
+#
+# Runs from the GCP VM `litellm-compatibility-matrix-populator` via the
+# systemd timer in this directory. The flow is:
+#
+#   1. Resolve the latest LiteLLM final release tag from the GitHub
+#      Releases API.
+#   2. Update a long-lived worktree at $WORKTREE to that tag and `uv sync` it.
+#   3. Boot the proxy as a background subprocess on $PROXY_PORT (default
+#      4100; a separate port from the human-tended :4000 proxy).
+#   4. Run `pytest tests/e2e/claude_code/` against the proxy. Test
+#      failures become `fail` cells in the JSON, not script errors.
+#   5. Hand the per-test results artifact + manifest to a small Python
+#      CLI (`build_matrix.py`) that wraps the existing
+#      `matrix_builder.build_from_paths` to produce the published
+#      compatibility-matrix.json.
+#   6. `gh repo clone` litellm-docs, write the JSON to a deterministic
+#      branch (`compat-matrix/<litellm>-<claude>-<UTC-date>`), commit,
+#      push the branch straight to BerriAI/litellm-docs (mateo-berri has
+#      write access), `gh pr create`, then — *only if no cell regressed
+#      green→red versus the currently-published matrix* — enable squash
+#      auto-merge so the PR merges itself once required checks pass. A
+#      green→red regression leaves auto-merge off for human review; an
+#      already-red cell (red→red) does not block.
+#   7. Sweep stale compat-matrix PRs: once today's PR exists, close any
+#      other open `compat-matrix/*` PR (and delete its bot-owned branch)
+#      so at most ONE compat-matrix PR is ever open — the newest. A
+#      gate-withheld PR that nobody triages is superseded by the next
+#      day's run rather than accumulating in the queue.
+#
+# Same-day reruns land on the same branch so they update the existing PR
+# rather than spawning a new one. If the JSON is byte-identical to the
+# docs branch, we skip the push entirely.
+#
+# Required commands on $PATH: git, uv, gh, jq, curl, claude, npm.
+# Required state: a litellm checkout at $LITELLM_REPO (this file lives in
+# it), $WORKTREE is created on first run, gh is already authenticated.
+#
+# Override any default by setting the matching env var; see the systemd
+# unit for the production wiring.
+
+set -Eeuo pipefail
+
+LITELLM_REPO="${LITELLM_REPO:-${HOME}/litellm/litellm}"
+WORKTREE="${LITELLM_WORKTREE:-${HOME}/litellm-cron-worktree}"
+PROXY_PORT="${PROXY_PORT:-4100}"
+PROXY_API_KEY="${PROXY_API_KEY:-sk-cron-matrix}"
+DOCS_REPO="${DOCS_REPO:-BerriAI/litellm-docs}"
+DOCS_BRANCH="${DOCS_BRANCH:-main}"
+DOCS_TARGET_PATH="${DOCS_TARGET_PATH:-src/data/compatibility-matrix.json}"
+SKIP_PUBLISH="${SKIP_PUBLISH:-0}"
+PYTEST_K="${PYTEST_K:-}"
+# The e2e suite uses PEP 695 `type` aliases, so the venv needs Python
+# >= 3.12 (also what repo CI runs) even when the VM's system python is
+# older. uv fetches a managed CPython of this version on first use --
+# checksum-verified against the manifest baked into the pinned uv
+# binary -- and installs it under ${WORKTREE}/.uv-python (see
+# UV_PYTHON_INSTALL_DIR below) so it lives inside the one tree the
+# systemd sandbox lets us write to.
+CRON_PYTHON_VERSION="${CRON_PYTHON_VERSION:-3.12}"
+# Merge method for auto-merge. BerriAI/litellm-docs only allows squash
+# merges (merge-commit and rebase are disabled at the repo level), so
+# `squash` is the only valid value here unless that changes upstream.
+AUTO_MERGE_METHOD="${AUTO_MERGE_METHOD:-squash}"
+
+POPULATOR_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+WORKDIR="$(mktemp -d -t litellm-compat-matrix.XXXXXX)"
+PROXY_PID_FILE="${WORKDIR}/proxy.pid"
+
+# Cleanup is intentionally aggressive: it can run on normal exit, on a
+# signal received by the script, or after a partial failure where the
+# proxy is up but ${PROXY_PID_FILE} is stale. We try four things in
+# order and stop as soon as the proxy port is free:
+#
+#   1. SIGTERM the pid recorded in proxy.pid.
+#   2. SIGKILL anything from `pgrep -f "litellm.*--port ${PROXY_PORT}"`
+#      that survived. This catches the common case where the recorded
+#      pid was the sh wrapper, not the long-lived python child.
+#   3. ss -K on the port (kernel kills sockets but not processes;
+#      mostly useful for catching lingering CLOSE_WAITs).
+#   4. wipe ${WORKDIR}.
+cleanup() {
+  local rc=$?
+  set +e
+  local proxy_pid
+  if [[ -f "${PROXY_PID_FILE}" ]]; then
+    proxy_pid="$(cat "${PROXY_PID_FILE}")"
+    if [[ -n "${proxy_pid}" ]]; then
+      kill -TERM "-${proxy_pid}" 2>/dev/null || kill -TERM "${proxy_pid}" 2>/dev/null || true
+      for _ in 1 2 3 4 5; do
+        kill -0 "${proxy_pid}" 2>/dev/null || break
+        sleep 1
+      done
+    fi
+  fi
+  # Belt-and-braces: any python or uv talking to ${PROXY_PORT} that
+  # survived the SIGTERM gets SIGKILL'd by name.
+  pgrep -f "litellm.*--port[ =]?${PROXY_PORT}([^0-9]|$)" 2>/dev/null \
+    | xargs -r kill -KILL 2>/dev/null || true
+  pgrep -f "${WORKTREE}/.uv-bin/uv.*run litellm" 2>/dev/null \
+    | xargs -r kill -KILL 2>/dev/null || true
+  rm -rf "${WORKDIR}"
+  exit "${rc}"
+}
+trap cleanup EXIT INT TERM
+
+log() { printf '==> %s\n' "$*" >&2; }
+die() { printf 'ERROR: %s\n' "$*" >&2; exit 1; }
+
+for cmd in git uv gh jq curl claude; do
+  command -v "${cmd}" >/dev/null 2>&1 || die "missing required command: ${cmd}"
+done
+
+# Publishing pushes the branch straight to BerriAI/litellm-docs and opens
+# the PR as mateo-berri, who has write access on the docs repo. Under
+# systemd the PAT arrives as a file via LoadCredential=, NOT via the
+# EnvironmentFile: several suite cells let the model-driven claude CLI
+# read arbitrary files as this user, and /proc/<pid>/environ of the
+# script, pytest, and the proxy would hand an env-borne token to any
+# same-UID reader. Kept as an unexported shell variable and passed per
+# invocation (GH_TOKEN=... / curl header / push URL), it never enters a
+# child's environment. Manual runs may export GITHUB_TOKEN instead.
+# Require it up front -- failing 30 minutes into a run is a waste of CI
+# quota.
+if [[ -z "${GITHUB_TOKEN:-}" && -n "${CREDENTIALS_DIRECTORY:-}" && -f "${CREDENTIALS_DIRECTORY}/github-token" ]]; then
+  GITHUB_TOKEN="$(<"${CREDENTIALS_DIRECTORY}/github-token")"
+  log "publish token source: systemd credential store"
+elif [[ -n "${GITHUB_TOKEN:-}" ]]; then
+  log "publish token source: process environment"
+fi
+if [[ "${SKIP_PUBLISH}" != "1" ]]; then
+  [[ -n "${GITHUB_TOKEN:-}" ]] \
+    || die "publish token required: /etc/litellm-compat-matrix-github-token via LoadCredential under systemd, or an exported GITHUB_TOKEN for manual runs (or set SKIP_PUBLISH=1)"
+fi
+
+# ---------------------------------------------------------------------------
+# 1. Resolve versions
+# ---------------------------------------------------------------------------
+
+# Newest PEP 440 *final* release on BerriAI/litellm. LiteLLM moved off
+# the legacy `vX.Y.Z-stable` tag convention to PEP 440: a final/stable
+# release is now a bare `vX.Y.Z` tag, while pre-releases carry a
+# `-rc.N` / `-dev.N` segment (and the old `…-stable` / `…-stable.patch.N`
+# tags are legacy and frozen at v1.83.x). We therefore select the newest
+# tag with no pre-release segment -- matching `^v[0-9]+\.[0-9]+\.[0-9]+$`
+# -- and skip drafts. The numeric version_key sort handles 1.10 > 1.9.
+#
+# Paginate through the releases endpoint instead of grabbing only page 1
+# (default page_size=30). LiteLLM ships multiple pre-releases per day, so
+# it's common to need to walk past 30+ entries before hitting the most
+# recent final release. We cap at 5 pages (500 releases) which is
+# conservatively beyond the worst observed gap.
+GH_AUTH_HEADER=()
+if [[ -n "${GITHUB_TOKEN:-}" ]]; then
+  GH_AUTH_HEADER=(-H "Authorization: Bearer ${GITHUB_TOKEN}")
+fi
+RELEASES_JSON="${WORKDIR}/releases.json"
+echo "[]" >"${RELEASES_JSON}"
+for page in 1 2 3 4 5; do
+  PAGE_JSON="${WORKDIR}/releases.page${page}.json"
+  curl -fsS \
+    -H 'Accept: application/vnd.github+json' \
+    -H 'User-Agent: litellm-compat-matrix' \
+    "${GH_AUTH_HEADER[@]}" \
+    "https://api.github.com/repos/BerriAI/litellm/releases?per_page=100&page=${page}" \
+    >"${PAGE_JSON}"
+  jq -s '.[0] + .[1]' "${RELEASES_JSON}" "${PAGE_JSON}" >"${RELEASES_JSON}.merged"
+  mv "${RELEASES_JSON}.merged" "${RELEASES_JSON}"
+  # Stop early once we've seen at least one final release tag — no point
+  # paging further for a daily script that only needs the newest.
+  if jq -e '[.[] | select((.draft // false) == false) | .tag_name // "" | select(test("^v[0-9]+\\.[0-9]+\\.[0-9]+$"))] | length > 0' "${PAGE_JSON}" >/dev/null; then
+    break
+  fi
+  # No more pages? GitHub returns an empty array past the last page.
+  if [[ "$(jq 'length' "${PAGE_JSON}")" == "0" ]]; then
+    break
+  fi
+done
+LITELLM_VERSION="$(
+  jq -r '
+    [ .[]
+      | select((.draft // false) == false)
+      | .tag_name // empty
+      | select(test("^v[0-9]+\\.[0-9]+\\.[0-9]+$"))
+    ]
+    | sort_by(
+        capture("^v(?<a>[0-9]+)\\.(?<b>[0-9]+)\\.(?<c>[0-9]+)$")
+        | [(.a|tonumber), (.b|tonumber), (.c|tonumber)]
+      )
+    | last // empty
+  ' "${RELEASES_JSON}"
+)"
+[[ -n "${LITELLM_VERSION}" ]] || die "could not resolve latest PEP 440 final release (vX.Y.Z) in 5 pages of releases"
+log "resolved litellm: ${LITELLM_VERSION}"
+
+CLAUDE_CODE_VERSION="$(claude --version 2>/dev/null | awk '{print $1}')"
+[[ -n "${CLAUDE_CODE_VERSION}" ]] || die "could not read 'claude --version'"
+log "local claude code: ${CLAUDE_CODE_VERSION}"
+
+# ---------------------------------------------------------------------------
+# 2. Update the worktree to that tag
+# ---------------------------------------------------------------------------
+
+if [[ ! -d "${WORKTREE}/.git" ]]; then
+  log "first run: cloning litellm into ${WORKTREE}"
+  mkdir -p "$(dirname "${WORKTREE}")"
+  git clone https://github.com/BerriAI/litellm.git "${WORKTREE}"
+fi
+
+log "updating worktree to ${LITELLM_VERSION}"
+git -C "${WORKTREE}" fetch --tags --force
+git -C "${WORKTREE}" reset --hard
+# Keep the venv, the .uv-bin cache, and the .uv-python managed
+# interpreter around — uv sync will reconcile the venv on every run,
+# and we don't want to re-download the pinned uv binary or the managed
+# CPython each time. Drop everything else (including any prior
+# tests/e2e/ shim) so each run starts clean before the shim below
+# rewrites it from the dev checkout.
+git -C "${WORKTREE}" clean -fdx -e .venv -e .uv-bin -e .uv-python
+git -C "${WORKTREE}" checkout --force "${LITELLM_VERSION}"
+
+# Always rebuild tests/e2e/ in the worktree from the dev checkout,
+# regardless of what the resolved ${LITELLM_VERSION} tag ships. Two
+# reasons:
+#
+#   * The matrix populator's job is to exercise *today's* tests against
+#     the latest stable proxy. The dev checkout carries the most recent
+#     test fixes that haven't yet rolled into a stable release, and we
+#     want every cron run to pick those up the moment they land on
+#     ${LITELLM_REPO}, not whenever the next stable release happens.
+#   * The tag's own tests/e2e/ ships the full EKS e2e harness, whose
+#     top-level conftest.py imports modules (e2e_db, lifecycle,
+#     otel_client, ...) that the stable venv does not install. Copying
+#     the whole tree would make pytest collection blow up on those
+#     imports.
+#
+# So the shim is a fresh `rm -rf` of tests/e2e/ followed by copying ONLY
+# the claude_code suite plus the shared transport helpers it imports.
+# pytest puts tests/e2e/ itself on sys.path (it has no __init__.py, while
+# claude_code/ does), which is what resolves both the `claude_code.*`
+# and the bare `proxy_client` / `e2e_http` imports inside the suite.
+E2E_HELPER_FILES=(proxy_client.py e2e_http.py models.py e2e_config.py transport.py)
+if [[ ! -d "${LITELLM_REPO}/tests/e2e/claude_code" ]]; then
+  die "no shim source at ${LITELLM_REPO}/tests/e2e/claude_code"
+fi
+for helper in "${E2E_HELPER_FILES[@]}"; do
+  [[ -f "${LITELLM_REPO}/tests/e2e/${helper}" ]] \
+    || die "missing shim helper: ${LITELLM_REPO}/tests/e2e/${helper}"
+done
+log "shimming tests/e2e/claude_code/ + helpers from ${LITELLM_REPO} (always-overwrite)"
+rm -rf "${WORKTREE}/tests/e2e"
+mkdir -p "${WORKTREE}/tests/e2e"
+cp -r "${LITELLM_REPO}/tests/e2e/claude_code" "${WORKTREE}/tests/e2e/"
+for helper in "${E2E_HELPER_FILES[@]}"; do
+  cp "${LITELLM_REPO}/tests/e2e/${helper}" "${WORKTREE}/tests/e2e/"
+done
+
+# litellm pins an exact uv version in pyproject.toml's [tool.uv]
+# `required-version` field, so a system uv that's newer or older
+# refuses to sync. We pin our own local copy at the version the
+# checked-out tag asks for, cached under .uv-bin/ inside the worktree
+# so subsequent runs skip the download.
+PINNED_UV_VERSION="$(
+  awk -F'"' '
+    /^required-version[[:space:]]*=/ {
+      # Field 2 is the value between the quotes, e.g. ">=0.10.9" or
+      # "0.10.9". Strip any leading specifier prefix so we end up with
+      # the bare version string, which is what /releases/download/<v>/
+      # expects.
+      v = $2
+      sub(/^[[:space:]=<>!~]+/, "", v)
+      if (v != "") { print v; exit }
+    }
+  ' "${WORKTREE}/pyproject.toml"
+)"
+if [[ -z "${PINNED_UV_VERSION}" ]]; then
+  log "no uv version pin in pyproject.toml; using system uv"
+  WORKTREE_UV="$(command -v uv)"
+else
+  WORKTREE_UV="${WORKTREE}/.uv-bin/uv-${PINNED_UV_VERSION}"
+  if [[ ! -x "${WORKTREE_UV}" ]]; then
+    log "downloading uv ${PINNED_UV_VERSION} for the worktree"
+    mkdir -p "${WORKTREE}/.uv-bin"
+    UV_TARBALL_NAME="uv-x86_64-unknown-linux-gnu.tar.gz"
+    UV_DOWNLOAD_URL="https://github.com/astral-sh/uv/releases/download/${PINNED_UV_VERSION}/${UV_TARBALL_NAME}"
+    UV_TMPDIR="$(mktemp -d -t uv-download.XXXXXX)"
+    # Download the tarball and Astral's official .sha256 sidecar to disk
+    # and verify the digest before extracting/executing anything. This
+    # closes the supply-chain trust gap of piping a remote binary
+    # straight into `tar -xzO ... > file ; chmod +x` (see CLAUDE.md
+    # "CI Supply-Chain Safety").
+    curl -fsSL --output "${UV_TMPDIR}/${UV_TARBALL_NAME}" "${UV_DOWNLOAD_URL}"
+    curl -fsSL --output "${UV_TMPDIR}/${UV_TARBALL_NAME}.sha256" "${UV_DOWNLOAD_URL}.sha256"
+    (cd "${UV_TMPDIR}" && sha256sum -c "${UV_TARBALL_NAME}.sha256") \
+      || { rm -rf "${UV_TMPDIR}"; die "uv ${PINNED_UV_VERSION} sha256 mismatch — refusing to install"; }
+    tar -xzf "${UV_TMPDIR}/${UV_TARBALL_NAME}" -C "${UV_TMPDIR}" "uv-x86_64-unknown-linux-gnu/uv"
+    mv "${UV_TMPDIR}/uv-x86_64-unknown-linux-gnu/uv" "${WORKTREE_UV}.tmp"
+    chmod +x "${WORKTREE_UV}.tmp"
+    mv "${WORKTREE_UV}.tmp" "${WORKTREE_UV}"
+    rm -rf "${UV_TMPDIR}"
+  fi
+fi
+# `--extra proxy` pulls fastapi/uvicorn/etc. so `uv run litellm` can
+# actually serve. `--group proxy-dev` brings in pytest and the rest of
+# what tests/e2e/claude_code/ needs. `--python` pins the venv to
+# ${CRON_PYTHON_VERSION}; the first run after a version bump recreates
+# the venv from scratch (a one-time cold sync).
+export UV_PYTHON_INSTALL_DIR="${WORKTREE}/.uv-python"
+log "uv sync --frozen --group proxy-dev --extra proxy --python ${CRON_PYTHON_VERSION} (uv ${PINNED_UV_VERSION:-system})"
+(cd "${WORKTREE}" && "${WORKTREE_UV}" sync --frozen --group proxy-dev --extra proxy --python "${CRON_PYTHON_VERSION}")
+
+PROXY_CONFIG="${WORKTREE}/tests/e2e/claude_code/test_config.yaml"
+[[ -f "${PROXY_CONFIG}" ]] || die "proxy config not found at ${PROXY_CONFIG} (shim incomplete?)"
+
+# ---------------------------------------------------------------------------
+# 3. Boot the proxy
+# ---------------------------------------------------------------------------
+
+log "starting proxy on 127.0.0.1:${PROXY_PORT}"
+# Bind the proxy to loopback only. The populator proxy is talked to
+# exclusively by the pytest run on the same host (the health check and
+# the test env set `LITELLM_PROXY_URL=http://127.0.0.1:...`),
+# so there's no reason to expose it on the VM's external interfaces.
+# Without `--host`, `litellm` defaults to 0.0.0.0, which combined with
+# the predictable default `LITELLM_MASTER_KEY=sk-cron-matrix` would
+# allow anything that can reach :${PROXY_PORT} on the VM to authenticate
+# and burn upstream provider credentials.
+#
+# `setsid` puts the proxy in its own session+pgroup so cleanup() can
+# SIGTERM the whole tree by passing the pgid as a negative pid. We
+# write that pid to a file so cleanup() doesn't need to remember a
+# variable that might be stale by the time the trap fires.
+setsid env LITELLM_MASTER_KEY="${PROXY_API_KEY}" bash -c '
+  echo "$$" > "$0"
+  cd "$1"
+  exec "$2" run litellm --config "$3" --host 127.0.0.1 --port "$4"
+' "${PROXY_PID_FILE}" "${WORKTREE}" "${WORKTREE_UV}" "${PROXY_CONFIG}" "${PROXY_PORT}" \
+  >"${WORKDIR}/proxy.log" 2>&1 &
+disown
+
+HEALTH_URL="http://127.0.0.1:${PROXY_PORT}/health/liveliness"
+for _ in $(seq 1 45); do
+  if curl -fsS "${HEALTH_URL}" >/dev/null 2>&1; then
+    break
+  fi
+  sleep 2
+done
+curl -fsS "${HEALTH_URL}" >/dev/null \
+  || { tail -50 "${WORKDIR}/proxy.log" >&2; die "proxy did not become healthy"; }
+
+# ---------------------------------------------------------------------------
+# 4. Run pytest
+# ---------------------------------------------------------------------------
+
+RESULTS_JSON="${WORKDIR}/compat-results.json"
+# The `_*_unit_tests` ignore is defensive: those harness-only trees are
+# markerless (they run without a proxy) and don't feed matrix cells, so
+# the cron skips them if/when they land in the suite.
+PYTEST_ARGS=(
+  tests/e2e/claude_code/
+  "--ignore-glob=*_unit_tests*"
+)
+if [[ -n "${PYTEST_K}" ]]; then
+  log "PYTEST_K set; narrowing to: ${PYTEST_K}"
+  PYTEST_ARGS+=(-k "${PYTEST_K}")
+fi
+
+log "running pytest"
+set +e
+(
+  cd "${WORKTREE}" \
+    && LITELLM_PROXY_URL="http://127.0.0.1:${PROXY_PORT}" \
+       LITELLM_MASTER_KEY="${PROXY_API_KEY}" \
+       COMPAT_RESULTS_PATH="${RESULTS_JSON}" \
+       "${WORKTREE_UV}" run pytest "${PYTEST_ARGS[@]}"
+)
+PYTEST_EXIT=$?
+set -e
+log "pytest exit code: ${PYTEST_EXIT} (failures become 'fail' cells, not script errors)"
+# 0=green, 1=test failures (fail cells); >=2 = interrupted/internal/usage/no
+# tests, i.e. a partial run whose missing cells would publish as not_tested.
+[[ ${PYTEST_EXIT} -le 1 ]] \
+  || die "pytest exited abnormally (${PYTEST_EXIT}); refusing to publish a partial matrix"
+[[ -f "${RESULTS_JSON}" ]] || die "pytest did not produce ${RESULTS_JSON}"
+
+# ---------------------------------------------------------------------------
+# 5. Build the matrix JSON
+# ---------------------------------------------------------------------------
+
+MATRIX_JSON="${WORKDIR}/compatibility-matrix.json"
+log "building ${MATRIX_JSON}"
+(
+  cd "${WORKTREE}" \
+    && "${WORKTREE_UV}" run python "${POPULATOR_DIR}/build_matrix.py" \
+       --manifest "${WORKTREE}/tests/e2e/claude_code/manifest.yaml" \
+       --results "${RESULTS_JSON}" \
+       --output "${MATRIX_JSON}" \
+       --litellm-version "${LITELLM_VERSION}" \
+       --claude-code-version "${CLAUDE_CODE_VERSION}"
+)
+
+# ---------------------------------------------------------------------------
+# 6. Open a docs-repo PR
+# ---------------------------------------------------------------------------
+
+if [[ "${SKIP_PUBLISH}" == "1" ]]; then
+  cp "${MATRIX_JSON}" "${LITELLM_REPO}/compatibility-matrix.json"
+  log "SKIP_PUBLISH=1; matrix written to ${LITELLM_REPO}/compatibility-matrix.json"
+  exit 0
+fi
+
+DATE_UTC="$(date -u +%Y-%m-%d)"
+BRANCH_NAME="compat-matrix/${LITELLM_VERSION}-${CLAUDE_CODE_VERSION}-${DATE_UTC}"
+DOCS_CLONE="${WORKDIR}/litellm-docs"
+
+log "cloning ${DOCS_REPO}@${DOCS_BRANCH}"
+gh repo clone "${DOCS_REPO}" "${DOCS_CLONE}" -- --depth 1 --branch "${DOCS_BRANCH}"
+
+cd "${DOCS_CLONE}"
+git config user.email "litellm-bot@berri.ai"
+git config user.name "litellm-compat-matrix-bot"
+git checkout -b "${BRANCH_NAME}"
+
+# Snapshot the currently-published matrix *before* we overwrite it, so the
+# auto-merge gate below can diff old→new cell statuses. On the first-ever
+# publish the file won't exist yet; we leave ${PUBLISHED_MATRIX} pointing
+# at a path that doesn't exist and let check_regressions.py treat that as
+# "no baseline → no regressions".
+PUBLISHED_MATRIX="${WORKDIR}/published-matrix.json"
+if [[ -f "${DOCS_TARGET_PATH}" ]]; then
+  cp "${DOCS_TARGET_PATH}" "${PUBLISHED_MATRIX}"
+fi
+
+mkdir -p "$(dirname "${DOCS_TARGET_PATH}")"
+cp "${MATRIX_JSON}" "${DOCS_TARGET_PATH}"
+git add "${DOCS_TARGET_PATH}"
+
+if git diff --cached --quiet; then
+  log "matrix JSON unchanged from ${DOCS_BRANCH}; skipping PR"
+  exit 0
+fi
+
+# --- Auto-merge regression gate --------------------------------------------
+# Only auto-merge when the new matrix is improvement-or-equal: every cell
+# transition is red→green, green→green, or red→red. If any cell flips
+# green→red (a `pass` that became `fail`), we still open/refresh the PR but
+# leave auto-merge OFF so a human reviews the regression before it lands on
+# the public docs table. A pre-existing red cell (e.g. Anthropic out of API
+# credits) is red→red and does NOT block, so the daily PR keeps flowing.
+log "checking for green->red regressions vs the published matrix"
+set +e
+REGRESSION_REPORT="$(
+  cd "${WORKTREE}" \
+    && "${WORKTREE_UV}" run python "${POPULATOR_DIR}/check_regressions.py" \
+       --old "${PUBLISHED_MATRIX}" \
+       --new "${MATRIX_JSON}"
+)"
+REGRESSION_EXIT=$?
+set -e
+printf '%s\n' "${REGRESSION_REPORT}" | sed 's/^/  /' >&2
+# Exit 0 = clean. Exit 3 = green→red regression(s) found. Any other code
+# means the checker itself errored; fail *closed* (withhold auto-merge) so a
+# bug in the gate can never silently auto-merge a regression.
+if [[ ${REGRESSION_EXIT} -eq 0 ]]; then
+  ALLOW_AUTOMERGE=1
+elif [[ ${REGRESSION_EXIT} -eq 3 ]]; then
+  ALLOW_AUTOMERGE=0
+  log "WARN: green->red regression(s) detected; auto-merge will be left OFF for review"
+else
+  ALLOW_AUTOMERGE=0
+  log "WARN: regression check errored (exit ${REGRESSION_EXIT}); withholding auto-merge to be safe"
+fi
+
+GENERATED_AT="$(jq -r '.generated_at' "${MATRIX_JSON}")"
+COMMIT_MSG="$(cat <<EOF
+Update Claude Code compatibility matrix
+
+litellm_version: ${LITELLM_VERSION}
+claude_code_version: ${CLAUDE_CODE_VERSION}
+generated_at: ${GENERATED_AT}
+EOF
+)"
+git commit -m "${COMMIT_MSG}"
+
+# Push the branch straight to BerriAI/litellm-docs. mateo-berri has write
+# access on the docs repo, so there's no fork hop: the PR is a same-repo
+# branch PR. The temp remote carries the token in its URL, so we add it,
+# push, then immediately remove it so the token never lingers in
+# ${DOCS_CLONE}/.git/config. (${DOCS_CLONE} is also rm -rf'd by the
+# cleanup trap on exit.)
+#
+# Plain --force (not --force-with-lease) is acceptable here: the
+# compat-matrix/* branch is bot-owned, only this script ever writes to
+# it, and runs are serialized by the systemd timer. --force-with-lease
+# would require a fetch to populate the remote-tracking ref before each
+# push and adds no safety in this single-writer setup.
+PUBLISH_PUSH_URL="https://x-access-token:${GITHUB_TOKEN}@github.com/${DOCS_REPO}.git"
+git remote remove publish 2>/dev/null || true
+git remote add publish "${PUBLISH_PUSH_URL}"
+git push --force --set-upstream publish "${BRANCH_NAME}"
+git remote remove publish
+unset PUBLISH_PUSH_URL
+
+# Per-feature status table for the PR body. Reviewers triage from this.
+PR_FEATURE_TABLE="$(jq -r '
+  .features[] as $f
+  | "- **\($f.name)**: " +
+    ([ .providers[] as $p
+       | "\($p)=\($f.providers[$p].status // "not_tested")"
+     ] | join(", "))
+' "${MATRIX_JSON}")"
+
+# When the gate withheld auto-merge, call it out at the top of the PR body
+# (with the offending cells) so a reviewer knows this PR needs a human and
+# why. On the clean path this section is empty. Note `$(...)` strips the
+# trailing newline, so the body below puts explicit blank lines *around*
+# the placeholder rather than relying on the heredoc's own spacing.
+if [[ "${ALLOW_AUTOMERGE}" != "1" ]]; then
+  PR_REGRESSION_SECTION="$(cat <<EOF
+> [!WARNING]
+> **Auto-merge disabled:** one or more cells regressed green→red versus the
+> currently-published matrix. Review the diff before merging.
+
+\`\`\`
+${REGRESSION_REPORT}
+\`\`\`
+EOF
+)"
+else
+  PR_REGRESSION_SECTION=""
+fi
+
+PR_TITLE="chore(compat-matrix): refresh for ${LITELLM_VERSION} + claude-code ${CLAUDE_CODE_VERSION}"
+PR_BODY="$(cat <<EOF
+Automated daily refresh of the Claude Code compatibility matrix.
+
+${PR_REGRESSION_SECTION}
+
+| Field | Value |
+| --- | --- |
+| litellm_version | \`${LITELLM_VERSION}\` |
+| claude_code_version | \`${CLAUDE_CODE_VERSION}\` |
+| generated_at | \`${GENERATED_AT}\` |
+
+## Per-feature results
+
+${PR_FEATURE_TABLE}
+
+---
+
+Generated by \`tests/e2e/claude_code/cron_vm/run_daily.sh\`. Close without merging if the diff looks wrong; the next cron run will reopen with fresh results.
+EOF
+)"
+
+log "opening PR from ${BRANCH_NAME} -> ${DOCS_REPO}:${DOCS_BRANCH} (as mateo-berri)"
+# GH_TOKEN is mateo-berri's write-scoped token, the same identity used
+# for release-listing above. The branch lives on ${DOCS_REPO} itself, so
+# --head is a bare branch name (a same-repo PR), not `OWNER:BRANCH`.
+set +e
+PR_OUT="$(
+  GH_TOKEN="${GITHUB_TOKEN}" gh pr create \
+    --repo "${DOCS_REPO}" \
+    --base "${DOCS_BRANCH}" \
+    --head "${BRANCH_NAME}" \
+    --title "${PR_TITLE}" \
+    --body "${PR_BODY}" 2>&1
+)"
+PR_EXIT=$?
+set -e
+echo "${PR_OUT}"
+
+if [[ ${PR_EXIT} -ne 0 ]]; then
+  if grep -q "a pull request for branch.*already exists" <<<"${PR_OUT}"; then
+    log "PR already exists for ${BRANCH_NAME}; updated branch in place"
+  else
+    die "gh pr create failed (exit ${PR_EXIT})"
+  fi
+fi
+
+# Enable auto-merge so the PR merges itself once the docs repo's required
+# checks pass -- we no longer gate these bot PRs on a second human
+# approval. mateo-berri authors and merges them directly. The repo only
+# permits squash merges and has auto-merge enabled at the repo level
+# (${AUTO_MERGE_METHOD} defaults to squash accordingly).
+#
+# This only fires when the regression gate above is satisfied
+# (${ALLOW_AUTOMERGE}==1): a green→red regression — or a gate error —
+# leaves auto-merge OFF so a human triages the PR.
+#
+# `gh pr merge --auto` is idempotent: re-enabling auto-merge on a PR that
+# already has it set is a no-op, so same-day reruns stay clean. It's
+# non-fatal: if auto-merge can't be enabled (e.g. the PR is already in a
+# clean/mergeable state with nothing left to wait on, or branch
+# protection isn't configured), the matrix JSON has still landed on the
+# PR and the worst case is a manual merge click.
+if [[ "${ALLOW_AUTOMERGE}" == "1" ]]; then
+  log "enabling ${AUTO_MERGE_METHOD} auto-merge on ${BRANCH_NAME}"
+  set +e
+  GH_TOKEN="${GITHUB_TOKEN}" gh pr merge \
+    "${BRANCH_NAME}" \
+    --repo "${DOCS_REPO}" \
+    --auto \
+    "--${AUTO_MERGE_METHOD}" 2>&1 | sed 's/^/  /'
+  AUTOMERGE_EXIT=${PIPESTATUS[0]}
+  set -e
+  if [[ ${AUTOMERGE_EXIT} -ne 0 ]]; then
+    log "WARN: gh pr merge --auto exited ${AUTOMERGE_EXIT} (non-fatal)"
+  fi
+else
+  # Regression (or gate error): make sure auto-merge is OFF. A same-day
+  # rerun may have enabled it on an earlier, clean pass, so explicitly
+  # disable rather than just skipping. The disable call itself is allowed
+  # to error (`--disable-auto` fails harmlessly when auto-merge was never
+  # enabled), but the read-back below is authoritative: a regressed matrix
+  # must never be left armed to merge, so a still-armed PR is fatal.
+  log "leaving ${BRANCH_NAME} for manual review; disabling any prior auto-merge"
+  set +e
+  GH_TOKEN="${GITHUB_TOKEN}" gh pr merge \
+    "${BRANCH_NAME}" \
+    --repo "${DOCS_REPO}" \
+    --disable-auto 2>&1 | sed 's/^/  /'
+  set -e
+  AUTOMERGE_ARMED="$(
+    GH_TOKEN="${GITHUB_TOKEN}" gh pr view \
+      "${BRANCH_NAME}" \
+      --repo "${DOCS_REPO}" \
+      --json autoMergeRequest \
+      --jq '.autoMergeRequest.enabledAt // empty'
+  )" || die "could not read back the auto-merge state on ${BRANCH_NAME}"
+  [[ -z "${AUTOMERGE_ARMED}" ]] \
+    || die "auto-merge still armed on ${BRANCH_NAME} (enabled ${AUTOMERGE_ARMED}) after --disable-auto"
+fi
+
+# --- Stale-PR sweep ----------------------------------------------------------
+# Keep at most ONE compat-matrix PR open: today's. Any other open
+# `compat-matrix/*` PR is a leftover from a day whose regression gate
+# withheld auto-merge and nobody triaged it; the PR we just opened or
+# refreshed above carries strictly fresher results, so the old one is
+# pure queue noise. Closing is non-destructive — the PR record and its
+# regression report stay browsable; only the bot-owned branch is
+# deleted. This runs only after today's PR exists (a `die` above skips
+# it), so a failed publish can never close the queue down to zero.
+#
+# Non-fatal: a sweep failure (rate limit, transient API error) leaves
+# stale PRs for the next run to retry; it must not fail the pipeline.
+log "sweeping stale compat-matrix PRs (keeping ${BRANCH_NAME})"
+set +e
+STALE_PRS="$(
+  GH_TOKEN="${GITHUB_TOKEN}" gh pr list \
+    --repo "${DOCS_REPO}" \
+    --state open \
+    --limit 100 \
+    --json number,headRefName \
+    --jq '.[] | select(.headRefName | startswith("compat-matrix/")) | "\(.number)\t\(.headRefName)"'
+)"
+while IFS=$'\t' read -r stale_pr stale_head; do
+  [[ -z "${stale_pr}" ]] && continue
+  [[ "${stale_head}" == "${BRANCH_NAME}" ]] && continue
+  GH_TOKEN="${GITHUB_TOKEN}" gh pr close "${stale_pr}" \
+    --repo "${DOCS_REPO}" \
+    --delete-branch \
+    --comment "Superseded by the newer daily compat-matrix PR from \`${BRANCH_NAME}\`; the populator keeps only the most recent compat-matrix PR open." 2>&1 | sed 's/^/  /'
+  if [[ ${PIPESTATUS[0]} -eq 0 ]]; then
+    log "closed stale compat-matrix PR #${stale_pr} (${stale_head})"
+  else
+    log "WARN: could not close stale compat-matrix PR #${stale_pr} (non-fatal)"
+  fi
+done <<<"${STALE_PRS}"
+set -e
+
+log "done"

tests/e2e/claude_code/matrix_builder.py

@@ -174,6 +174,86 @@ def _aggregate_cell(results: Sequence[Mapping[str, Any]]) -> Dict[str, Any]:
     return {"status": "not_tested"}
 
 
+def _index_cells(matrix: Mapping[str, Any]) -> dict[tuple[str, str], dict[str, Any]]:
+    """Map ``(feature_id, provider) -> cell dict`` for a built matrix.
+
+    Cells are keyed by the *stable* feature ``id`` (not the display
+    ``name``, which can be reworded without changing the underlying row)
+    and the provider key, so two matrices built at different times line up
+    even if feature names drift.
+    """
+    out: dict[tuple[str, str], dict[str, Any]] = {}
+    for feature in matrix.get("features", []) or []:
+        if not isinstance(feature, Mapping):
+            continue
+        feature_id = feature.get("id")
+        if not feature_id:
+            continue
+        providers = feature.get("providers", {}) or {}
+        if not isinstance(providers, Mapping):
+            continue
+        for provider, cell in providers.items():
+            if isinstance(cell, Mapping):
+                out[(feature_id, provider)] = dict(cell)
+    return out
+
+
+def find_regressions(
+    old_matrix: Mapping[str, Any],
+    new_matrix: Mapping[str, Any],
+) -> list[dict[str, str]]:
+    """Return the cells that flipped green→red (``pass`` → ``fail``).
+
+    A *regression* is defined strictly: a cell that was ``pass`` in
+    ``old_matrix`` and is ``fail`` in ``new_matrix``. Every other
+    transition is intentionally *not* a regression:
+
+      * ``red → green`` / ``green → green`` — the happy path.
+      * ``red → red`` — a cell that is *already* failing for an unrelated
+        reason (e.g. Anthropic out of API credits) must not block
+        publishing, otherwise the daily PR would never auto-merge until
+        that independent issue is fixed.
+      * ``green → not_tested`` / ``green → not_applicable`` — a cell going
+        grey is a degradation but not a *red* regression; treating a
+        skipped/flaky run as a hard block would create false positives.
+
+    Cells present only in ``new_matrix`` (a newly added feature or
+    provider) have no baseline and therefore cannot be regressions.
+
+    Each returned item is a flat str→str mapping so callers (the cron's
+    ``check_regressions.py``) can render it without further lookups:
+    ``feature_id``, ``feature_name``, ``provider``, ``old_status``,
+    ``new_status``, ``error``.
+    """
+    old_cells = _index_cells(old_matrix)
+    feature_names = {
+        f.get("id"): str(f.get("name", f.get("id")))
+        for f in new_matrix.get("features", []) or []
+        if isinstance(f, Mapping) and f.get("id")
+    }
+
+    regressions: list[dict[str, str]] = []
+    for (feature_id, provider), new_cell in sorted(
+        _index_cells(new_matrix).items(), key=lambda kv: (kv[0][0], kv[0][1])
+    ):
+        if new_cell.get("status") != "fail":
+            continue
+        old_cell = old_cells.get((feature_id, provider))
+        if old_cell is None or old_cell.get("status") != "pass":
+            continue
+        regressions.append(
+            {
+                "feature_id": str(feature_id),
+                "feature_name": feature_names.get(feature_id, str(feature_id)),
+                "provider": str(provider),
+                "old_status": "pass",
+                "new_status": "fail",
+                "error": str(new_cell.get("error", "")),
+            }
+        )
+    return regressions
+
+
 def build_from_paths(
     *,
     manifest_path: Path,

tests/e2e/ui/fixtures/mock_llm_server/server.py

@@ -3,6 +3,7 @@ Mock LLM server for UI e2e tests.
 Responds to OpenAI-format endpoints with canned responses.
 """
 
+import os
 import time
 import json
 import uuid
@@ -117,4 +118,12 @@ async def embeddings(request: Request):
 
 
 if __name__ == "__main__":
-    uvicorn.run(app, host="127.0.0.1", port=8090)
+    # The port is overridable so two checkouts can run the harness at the same
+    # time; the default keeps every existing caller (run_e2e.sh, the CircleCI
+    # job, the e2e chart's sidecar) working untouched.
+    #
+    # The HOST is deliberately NOT configurable. Binding loopback is what makes
+    # this reachable at 127.0.0.1:8090 from inside the proxy's own pod, which is
+    # the contract the deployed config.yml and the e2e values file are written
+    # against.
+    uvicorn.run(app, host="127.0.0.1", port=int(os.environ.get("MOCK_LLM_PORT", "8090")))

tests/e2e/ui/helpers/mcp.ts

@@ -0,0 +1,65 @@
+import { expect, Page as PwPage } from "@playwright/test";
+import { navigateToPage } from "./navigation";
+import { Page } from "../fixtures/pages";
+import { masterKey } from "./traffic";
+
+/** Creates an MCP server through the UI's discovery to custom-form flow and returns its name. */
+export async function createMcpServer(page: PwPage, url: string): Promise<string> {
+  await navigateToPage(page, Page.McpServers);
+
+  await page.getByRole("button", { name: /Add New MCP Server/i }).click();
+  const discovery = page.getByRole("dialog").filter({ hasText: "Add MCP Server" });
+  await expect(discovery).toBeVisible({ timeout: 5_000 });
+  await discovery.getByRole("button", { name: /Custom Server/i }).click();
+
+  const formModal = page.locator(".ant-modal:visible").filter({ hasText: "MCP Server Name" });
+  await expect(formModal).toBeVisible({ timeout: 5_000 });
+
+  // validateMCPServerName rejects spaces and hyphens; the worker index avoids a same-millisecond collision.
+  const name = `e2e_mcp_${process.env.TEST_WORKER_INDEX ?? "0"}_${Date.now()}`;
+  await formModal.locator('input[id="server_name"]').fill(name);
+
+  const transportField = formModal.locator(".ant-form-item", { hasText: "Transport Type" });
+  await transportField.locator(".ant-select").click();
+  await page.locator(".ant-select-dropdown:visible").getByText("Streamable HTTP").click();
+
+  await formModal.locator('input[id="url"]').fill(url);
+
+  // The auth_type Form.Item has no label prop, so anchor on the enclosing Collapse panel.
+  const authSection = formModal.locator(".ant-collapse-item", { hasText: /^Authentication/ });
+  await authSection.locator(".ant-form-item").first().locator(".ant-select").click();
+  await page.locator(".ant-select-dropdown:visible").getByText("None", { exact: true }).click();
+
+  await formModal.getByRole("button", { name: /^Add MCP Server$/ }).click();
+  await expect(page.getByText("MCP Server created successfully").first()).toBeVisible({ timeout: 15_000 });
+
+  const card = page.getByTestId("mcp-servers-grid").getByText(name).first();
+  await expect(card).toBeVisible({ timeout: 10_000 });
+  return name;
+}
+
+/**
+ * Deletes every server carrying `serverName`. Leaked servers break unrelated MCP specs: the page
+ * reaches out to each one it lists, so unreachable leftovers stall networkidle until it times out.
+ * Errors are swallowed because this runs from afterEach.
+ */
+export async function deleteMcpServerByName(page: PwPage, serverName: string): Promise<void> {
+  const headers = { Authorization: `Bearer ${masterKey()}` };
+  try {
+    const res = await page.request.get("/v1/mcp/server", { headers });
+    if (!res.ok()) return;
+    const servers = (await res.json()) as { server_id: string; server_name?: string }[];
+    for (const server of servers.filter((candidate) => candidate.server_name === serverName)) {
+      await page.request.delete(`/v1/mcp/server/${server.server_id}`, { headers });
+    }
+  } catch {
+    // best effort, see above
+  }
+}
+
+/** Opens a server from the grid and switches to its MCP Tools tab. */
+export async function openMcpToolsTab(page: PwPage, serverName: string): Promise<void> {
+  await page.getByTestId("mcp-servers-grid").getByText(serverName).first().click();
+  await expect(page.getByRole("button", { name: /Back to All Servers/i })).toBeVisible({ timeout: 10_000 });
+  await page.getByRole("tab", { name: "MCP Tools" }).click();
+}

tests/e2e/ui/helpers/playground.ts

@@ -0,0 +1,46 @@
+import { expect, type Locator, type Page as PlaywrightPage } from "@playwright/test";
+import { navigateToPage, dismissFeedbackPopup } from "./navigation";
+import { Page } from "../fixtures/pages";
+
+/** Controls for the Test Key / Playground page, shared with the router-fallback specs. */
+
+/**
+ * The configuration panel is rendered twice, docked and overlay, with one visible at a time.
+ * Every control is narrowed to the visible copy or it trips strict mode against its hidden twin.
+ */
+export const onlyVisible = (locator: Locator): Locator => locator.filter({ visible: true }).first();
+
+/** The model dropdown, addressed by the placeholder it shows before selection. */
+export const modelSelect = (page: PlaywrightPage): Locator =>
+  onlyVisible(page.locator('.ant-select:has(.ant-select-selection-placeholder:text-is("Select a Model"))'));
+
+/** Send button is icon-only (an up-arrow), so there is no accessible name. */
+export const sendButton = (page: PlaywrightPage): Locator => onlyVisible(page.locator("button:has(.anticon-arrow-up)"));
+
+/** The Virtual Key Source dropdown, addressed by its currently selected label. */
+export const keySourceSelect = (page: PlaywrightPage, current: string): Locator =>
+  onlyVisible(page.locator(`.ant-select:has(.ant-select-selection-item[title="${current}"])`));
+
+export async function openPlayground(page: PlaywrightPage): Promise<void> {
+  await navigateToPage(page, Page.LlmPlayground);
+  await dismissFeedbackPopup(page);
+  await expect(onlyVisible(page.getByText("Virtual Key Source"))).toBeVisible({
+    timeout: 20_000,
+  });
+}
+
+export async function selectModel(page: PlaywrightPage, model: string): Promise<void> {
+  const select = modelSelect(page);
+  await select.click();
+  // Virtualized: options outside the rendered window are absent from the DOM, so search first.
+  await select.locator("input.ant-select-selection-search-input").fill(model);
+  // antd portals its dropdown to the body; options carry the value as `title`.
+  await onlyVisible(page.locator(`.ant-select-item-option[title="${model}"]`)).click({ timeout: 15_000 });
+}
+
+export async function sendMessage(page: PlaywrightPage, message: string): Promise<void> {
+  const input = onlyVisible(page.getByPlaceholder("Type your message", { exact: false }));
+  await expect(input).toBeVisible({ timeout: 15_000 });
+  await input.fill(message);
+  await sendButton(page).click();
+}

tests/e2e/ui/helpers/roundTrip.ts

@@ -0,0 +1,28 @@
+import { expect, Page } from "@playwright/test";
+import { masterKey } from "./traffic";
+
+/**
+ * Runs `action` and returns the parsed body of the first matching request.
+ *
+ * `action` is a callback so the listener is armed before the click; awaiting the
+ * click first lets the request go by, and the test then hangs until timeout.
+ */
+export async function captureRequestBody(
+  page: Page,
+  match: { method: string; urlIncludes: string },
+  action: () => Promise<void>,
+): Promise<Record<string, any>> {
+  const pending = page.waitForRequest((req) => req.method() === match.method && req.url().includes(match.urlIncludes));
+  await action();
+  const request = await pending;
+  return JSON.parse(request.postData() ?? "{}") as Record<string, any>;
+}
+
+/** Reads an endpoint as the master key, so a failure is bad data and not an expired UI token. */
+export async function readBack<T = any>(page: Page, endpoint: string): Promise<T> {
+  const res = await page.request.get(endpoint, {
+    headers: { Authorization: `Bearer ${masterKey()}` },
+  });
+  expect(res.ok(), `GET ${endpoint}`).toBe(true);
+  return (await res.json()) as T;
+}

tests/e2e/ui/helpers/traffic.ts

@@ -0,0 +1,125 @@
+import { APIRequestContext, expect } from "@playwright/test";
+
+/** Model names served by fixtures/config.yml, both backed by the mock LLM server. */
+export const CHAT_MODEL_A = "fake-openai-gpt-4";
+export const CHAT_MODEL_B = "fake-anthropic-claude";
+
+/** The only completion text fixtures/mock_llm_server/server.py ever returns. */
+export const MOCK_RESPONSE_TEXT = "This is a mock response.";
+
+export const masterKey = (): string => process.env.LITELLM_MASTER_KEY || "sk-1234";
+
+const rootPath = (): string => process.env.SERVER_ROOT_PATH ?? "";
+
+interface ChatOptions {
+  model: string;
+  prompt: string;
+  apiKey?: string;
+  /** Sent as `user`, which lands in the spend log's end_user column. */
+  endUser?: string;
+}
+
+/** POST /v1/chat/completions and return the completion id (the Logs Request ID). */
+export async function sendChatCompletion(request: APIRequestContext, opts: ChatOptions): Promise<string> {
+  const res = await request.post(`${rootPath()}/v1/chat/completions`, {
+    headers: {
+      Authorization: `Bearer ${opts.apiKey ?? masterKey()}`,
+      "Content-Type": "application/json",
+    },
+    data: {
+      model: opts.model,
+      messages: [{ role: "user", content: opts.prompt }],
+      ...(opts.endUser ? { user: opts.endUser } : {}),
+    },
+  });
+  expect(res.ok(), `chat completion for ${opts.model} failed (${res.status()}): ${await res.text()}`).toBe(true);
+  const body = await res.json();
+  expect(body.choices?.[0]?.message?.content).toContain(MOCK_RESPONSE_TEXT);
+  return body.id as string;
+}
+
+/** `key` is the sk- value to authenticate with; `token` is its hash, which spend aggregates are keyed by. */
+export async function createVirtualKey(
+  request: APIRequestContext,
+  data: Record<string, unknown> = {},
+): Promise<{ key: string; token: string; alias?: string }> {
+  const res = await request.post(`${rootPath()}/key/generate`, {
+    headers: {
+      Authorization: `Bearer ${masterKey()}`,
+      "Content-Type": "application/json",
+    },
+    data,
+  });
+  expect(res.ok(), `key generate failed (${res.status()}): ${await res.text()}`).toBe(true);
+  const body = await res.json();
+  return {
+    key: body.key as string,
+    token: (body.token ?? body.token_id) as string,
+    alias: body.key_alias as string | undefined,
+  };
+}
+
+/** Spend logs are flushed on a timer, so an assertion straight after a completion races the writer. */
+export async function waitForSpendLog(
+  request: APIRequestContext,
+  requestId: string,
+  timeoutMs = 60_000,
+): Promise<void> {
+  const deadline = Date.now() + timeoutMs;
+  let lastStatus = 0;
+  while (Date.now() < deadline) {
+    const res = await request.get(`${rootPath()}/spend/logs?request_id=${encodeURIComponent(requestId)}`, {
+      headers: { Authorization: `Bearer ${masterKey()}` },
+    });
+    lastStatus = res.status();
+    if (res.ok()) {
+      const body = await res.json();
+      const rows = Array.isArray(body) ? body : (body?.data ?? []);
+      if (rows.length > 0) {
+        return;
+      }
+    }
+    await new Promise((r) => setTimeout(r, 2_000));
+  }
+  throw new Error(`spend log for request ${requestId} never appeared (last /spend/logs status ${lastStatus})`);
+}
+
+const isoDay = (d: Date): string => d.toISOString().slice(0, 10);
+
+/**
+ * The Usage page reads /user/daily/activity, a rollup written by a background job, and fetches it once
+ * on mount. Navigating before the rollup lands leaves a stale render that never refreshes.
+ */
+export async function waitForKeyInDailyActivity(
+  request: APIRequestContext,
+  keyToken: string,
+  timeoutMs = 120_000,
+): Promise<void> {
+  const now = new Date();
+  const start = new Date(now);
+  start.setDate(start.getDate() - 7);
+  const query = `start_date=${isoDay(start)}&end_date=${isoDay(now)}`;
+
+  const deadline = Date.now() + timeoutMs;
+  let lastStatus = 0;
+  while (Date.now() < deadline) {
+    const res = await request.get(`${rootPath()}/user/daily/activity?${query}`, {
+      headers: { Authorization: `Bearer ${masterKey()}` },
+    });
+    lastStatus = res.status();
+    if (res.ok()) {
+      const body = await res.json();
+      const seen = (body?.results ?? []).some(
+        (day: { breakdown?: { api_keys?: Record<string, unknown> } }) => keyToken in (day.breakdown?.api_keys ?? {}),
+      );
+      if (seen) {
+        return;
+      }
+    }
+    await new Promise((r) => setTimeout(r, 3_000));
+  }
+  throw new Error(
+    `key ${keyToken} never appeared in /user/daily/activity (last status ${lastStatus}); ` +
+      "the daily spend rollup may not be running",
+  );
+}

tests/e2e/ui/run_e2e.sh

@@ -12,6 +12,10 @@ set -euo pipefail
 #   ./run_e2e.sh --repeat-each=5    # Run each test 5 times
 #   ./run_e2e.sh --headed           # Run with browser visible
 #
+# Ports default to 4000 / 5432 / 8090 and can be moved when another checkout
+# already holds them:
+#   PROXY_PORT=4100 POSTGRES_PORT=5532 MOCK_LLM_PORT=8190 ./run_e2e.sh
+#
 # In CI (CI=true), expects:
 #   - PostgreSQL already running on 127.0.0.1:5432
 #   - DATABASE_URL already set
@@ -28,12 +32,50 @@ MOCK_PID=""
 PROXY_PID=""
 PROXY_LOG=""
 
+# Ports, overridable so two checkouts can run this harness at the same time --
+# otherwise a second run aborts on "port 4000 is in use" and the only way out is
+# to stop someone else's stack. Defaults are the historical values, so an unset
+# environment behaves exactly as before (CI, the CircleCI job and the docs all
+# assume 4000/5432/8090).
+PROXY_PORT="${PROXY_PORT:-4000}"
+POSTGRES_PORT="${POSTGRES_PORT:-5432}"
+MOCK_LLM_PORT="${MOCK_LLM_PORT:-8090}"
+export MOCK_LLM_PORT
+
 # --- Ensure common tool paths are available (local dev only) ---
 if [ "$IS_CI" = "false" ]; then
   for p in /usr/local/bin /opt/homebrew/bin "$HOME/.local/bin" /opt/homebrew/opt/postgresql@14/bin /opt/homebrew/opt/libpq/bin; do
     [ -d "$p" ] && export PATH="$p:$PATH"
   done
-  [ -s "$HOME/.nvm/nvm.sh" ] && source "$HOME/.nvm/nvm.sh"
+  # Sourcing nvm only makes `nvm` available -- it leaves you on whatever the
+  # default alias points at, which is frequently an older Node than the
+  # dashboard's engines allow. `npm install` then fails EBADENGINE, npm exits
+  # non-zero, and because the install below is `--silent ... || true` the error
+  # is swallowed and the run dies later with the far less obvious
+  # "sh: next: command not found".
+  #
+  # So select a Node that satisfies ui/litellm-dashboard's engines.node, and if
+  # none is available say so here rather than 200 lines downstream.
+  if [ -s "$HOME/.nvm/nvm.sh" ]; then
+    # shellcheck disable=SC1091
+    source "$HOME/.nvm/nvm.sh"
+    required_major="$(sed -nE 's/.*"node"[[:space:]]*:[[:space:]]*">=?([0-9]+).*/\1/p' \
+      "$DASHBOARD_DIR/package.json" 2>/dev/null | head -1)"
+    if [ -n "$required_major" ]; then
+      current_major="$(node --version 2>/dev/null | sed -E 's/^v([0-9]+).*/\1/')"
+      if [ -z "$current_major" ] || [ "$current_major" -lt "$required_major" ]; then
+        echo "Node $(node --version 2>/dev/null || echo 'not found') is below the dashboard's required v${required_major}; selecting a newer one via nvm"
+        nvm use "$required_major" >/dev/null 2>&1 || nvm use --lts >/dev/null 2>&1 || true
+        current_major="$(node --version 2>/dev/null | sed -E 's/^v([0-9]+).*/\1/')"
+        if [ -z "$current_major" ] || [ "$current_major" -lt "$required_major" ]; then
+          echo "Error: ui/litellm-dashboard requires Node >= v${required_major}, and no such version is installed."
+          echo "       Install one with:  nvm install ${required_major}"
+          exit 1
+        fi
+      fi
+      echo "Using Node $(node --version) / npm $(npm --version)"
+    fi
+  fi
 fi
 
 # --- Cleanup on exit ---
@@ -47,7 +89,11 @@ cleanup() {
   fi
   echo "Done."
 }
-trap cleanup EXIT INT TERM
+on_signal() {
+  exit 130
+}
+trap cleanup EXIT
+trap on_signal INT TERM
 
 # --- Pre-flight checks ---
 for cmd in python3 npx uv; do
@@ -59,9 +105,14 @@ if [ "$IS_CI" = "false" ]; then
   for cmd in docker psql; do
     command -v "$cmd" >/dev/null 2>&1 || { echo "Error: $cmd not found."; exit 1; }
   done
-  for port in 4000 5432 8090; do
-    if lsof -ti ":$port" >/dev/null 2>&1; then
-      echo "Error: port $port is in use"
+  # Only a LISTENER conflicts with us. Without -sTCP:LISTEN this also matches
+  # ESTABLISHED sockets, so an unrelated *outbound* connection from this machine
+  # to someone else's :5432 (a psql session, a running app, a Prisma engine
+  # talking to a remote database) aborts the run with "port 5432 is in use"
+  # while nothing is actually bound locally.
+  for port in "$PROXY_PORT" "$POSTGRES_PORT" "$MOCK_LLM_PORT"; do
+    if lsof -nP -iTCP:"$port" -sTCP:LISTEN >/dev/null 2>&1; then
+      echo "Error: port $port is in use (override with PROXY_PORT / POSTGRES_PORT / MOCK_LLM_PORT)"
       exit 1
     fi
   done
@@ -69,12 +120,12 @@ if [ "$IS_CI" = "false" ]; then
   export POSTGRES_USER="e2euser"
   export POSTGRES_PASSWORD="$(openssl rand -hex 32)"
   export POSTGRES_DB="litellm_e2e"
-  export DATABASE_URL="postgresql://${POSTGRES_USER}:${POSTGRES_PASSWORD}@127.0.0.1:5432/${POSTGRES_DB}"
+  export DATABASE_URL="postgresql://${POSTGRES_USER}:${POSTGRES_PASSWORD}@127.0.0.1:${POSTGRES_PORT}/${POSTGRES_DB}"
 
   echo "=== Starting PostgreSQL ==="
   docker run -d --rm --name "$CONTAINER_NAME" \
     -e POSTGRES_USER -e POSTGRES_PASSWORD -e POSTGRES_DB \
-    -p 127.0.0.1:5432:5432 \
+    -p "127.0.0.1:${POSTGRES_PORT}:5432" \
     postgres:16
 
   echo "Waiting for PostgreSQL..."
@@ -91,8 +142,13 @@ fi
 
 # --- Credentials ---
 export LITELLM_MASTER_KEY="sk-1234"
-export MOCK_LLM_URL="http://127.0.0.1:8090/v1"
+export MOCK_LLM_URL="http://127.0.0.1:${MOCK_LLM_PORT}/v1"
 export DISABLE_SCHEMA_UPDATE="true"
+# The suite resolves its target from E2E_UI_BASE_URL (constants.ts), which
+# otherwise defaults to :4000 -- so without this a relocated stack would be
+# built and booted correctly and then tested against whatever happens to be
+# listening on the default port.
+export E2E_UI_BASE_URL="${E2E_UI_BASE_URL:-http://127.0.0.1:${PROXY_PORT}}"
 # Ensure the proxy serves UI at /ui (not behind a subpath)
 export SERVER_ROOT_PATH=""
 # Boot with an external logout URL so proxyLogoutUrl.spec.ts can assert the
@@ -108,7 +164,11 @@ export LITELLM_LICENSE="${LITELLM_LICENSE:-}"
 # --- Rebuild UI from source ---
 echo "=== Building UI from source ==="
 cd "$DASHBOARD_DIR"
-npm install --silent 2>/dev/null || true
+# NOT silenced, and NOT `|| true`. Swallowing this is what turns a one-line
+# EBADENGINE ("dashboard requires node >=24, you have v20") into the
+# considerably less helpful "sh: next: command not found" from the build below,
+# because the deps that provide `next` were never installed.
+npm install
 npm run build
 # Copy the fresh build to the proxy's static UI directory
 cp -r "$DASHBOARD_DIR/out/" "$REPO_ROOT/litellm/proxy/_experimental/out/"
@@ -139,7 +199,7 @@ uv run --no-sync python "$SCRIPT_DIR/fixtures/mock_llm_server/server.py" &
 MOCK_PID=$!
 
 for i in $(seq 1 15); do
-  if curl -sf http://127.0.0.1:8090/health >/dev/null 2>&1; then break; fi
+  if curl -sf http://127.0.0.1:${MOCK_LLM_PORT}/health >/dev/null 2>&1; then break; fi
   sleep 1
 done
 
@@ -149,7 +209,7 @@ cd "$REPO_ROOT"
 PROXY_LOG="${TMPDIR:-/tmp}/litellm-e2e-proxy-$$.log"
 uv run --no-sync python -m litellm.proxy.proxy_cli \
   --config "$SCRIPT_DIR/fixtures/config.yml" \
-  --port 4000 >"$PROXY_LOG" 2>&1 &
+  --port "$PROXY_PORT" >"$PROXY_LOG" 2>&1 &
 PROXY_PID=$!
 
 echo "Waiting for proxy (logs: $PROXY_LOG)..."
@@ -160,7 +220,7 @@ for i in $(seq 1 180); do
     tail -n 100 "$PROXY_LOG"
     exit 1
   fi
-  HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:4000/health -H "Authorization: Bearer $LITELLM_MASTER_KEY" 2>/dev/null || true)
+  HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" http://127.0.0.1:${PROXY_PORT}/health -H "Authorization: Bearer $LITELLM_MASTER_KEY" 2>/dev/null || true)
   if [ "$HTTP_CODE" = "200" ]; then
     PROXY_READY=1
     break
@@ -188,9 +248,38 @@ PGPASSWORD="$DB_PASS" psql -h "$DB_HOST" -p "$DB_PORT" -U "$DB_USER" -d "$DB_NAM
 # --- Playwright ---
 echo "=== Installing Playwright dependencies ==="
 cd "$SCRIPT_DIR"
-npm install --silent 2>/dev/null || true
+# Same reasoning as the dashboard install above: a failure here means the suite
+# has no @playwright/test, and the run should say that rather than fail later.
+npm install
 npx playwright install chromium --with-deps 2>/dev/null || npx playwright install chromium
 
+# Authoring a new spec means running it over and over against a stack that is
+# already up -- rebuilding the UI and re-seeding for every iteration costs
+# minutes each time. E2E_KEEP_ALIVE brings the stack up, then blocks, so you can
+# run `npx playwright test <spec>` yourself from another shell against it.
+# Ctrl-C here tears everything down through the usual trap.
+if [ "${E2E_KEEP_ALIVE:-0}" = "1" ]; then
+  cat <<EOF
+
+=== Stack is up (E2E_KEEP_ALIVE=1); not running tests ===
+  UI / API : http://127.0.0.1:${PROXY_PORT}
+  Mock LLM : http://127.0.0.1:${MOCK_LLM_PORT}/v1
+  Database : $DATABASE_URL
+  Proxy log: $PROXY_LOG
+
+Run specs against it from $SCRIPT_DIR:
+  npx playwright test --config playwright.config.ts <spec>
+
+Press Ctrl-C to tear the stack down.
+EOF
+  while kill -0 "$PROXY_PID" 2>/dev/null; do
+    sleep 5
+  done
+  echo "Error: proxy process exited unexpectedly. Proxy output:"
+  tail -n 100 "$PROXY_LOG"
+  exit 1
+fi
+
 echo "=== Running Playwright tests ==="
 npx playwright test --config playwright.config.ts "$@"
 EXIT_CODE=$?

tests/e2e/ui/tests/logs/logs.spec.ts

@@ -0,0 +1,221 @@
+import { test, expect, type Locator, type Page as PlaywrightPage } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { navigateToPage, dismissFeedbackPopup } from "../../helpers/navigation";
+import { Page } from "../../fixtures/pages";
+import { CHAT_MODEL_A, MOCK_RESPONSE_TEXT, sendChatCompletion, waitForSpendLog } from "../../helpers/traffic";
+
+/**
+ * Anchored to traffic this spec generates itself, with a unique prompt and end user per run, so it
+ * neither depends on seeded spend rows nor collides with other specs under parallelism.
+ */
+
+const uniqueSuffix = (): string => `${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+
+/**
+ * Walking up from the label is the only stable handle: the header carries no role, test id or class,
+ * and its copy button is icon-only with a hover-only tooltip.
+ */
+const sectionHeader = (drawer: Locator, label: "Input" | "Output"): Locator =>
+  drawer.getByText(label, { exact: true }).locator("xpath=../../..");
+
+/** Every tab stays mounted, so the DOM holds four tables at once; scope to the visible one. */
+const requestLogsRows = (page: PlaywrightPage): Locator =>
+  page.locator("table").filter({ visible: true }).first().locator("tbody tr");
+
+const visibleTestId = (page: PlaywrightPage, id: string): Locator => page.getByTestId(id).filter({ visible: true });
+
+/** Open the Logs page and filter the table down to a single request id. */
+async function openLogsForRequest(page: PlaywrightPage, requestId: string): Promise<Locator> {
+  await navigateToPage(page, Page.Logs);
+  await dismissFeedbackPopup(page);
+
+  const search = visibleTestId(page, "datatable-search");
+  await expect(search).toBeVisible({ timeout: 20_000 });
+  await search.fill(requestId);
+
+  const row = requestLogsRows(page).filter({ hasText: requestId });
+  await expect(row, `no logs row for request ${requestId}`).toHaveCount(1, {
+    timeout: 30_000,
+  });
+  return row;
+}
+
+test.describe("Logs page", () => {
+  test.use({
+    storageState: ADMIN_STORAGE_PATH,
+    // The copy buttons go through navigator.clipboard, which rejects without these.
+    permissions: ["clipboard-read", "clipboard-write"],
+  });
+
+  test("a served request expands to its request and response", async ({ page, request }) => {
+    const prompt = `logs-detail-prompt-${uniqueSuffix()}`;
+    const requestId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt,
+    });
+    await waitForSpendLog(request, requestId);
+
+    const row = await openLogsForRequest(page, requestId);
+
+    // Expand: clicking the row opens the detail drawer for that request.
+    await row.click();
+    const drawer = page.locator(".ant-drawer-content").first();
+    await expect(drawer).toBeVisible({ timeout: 20_000 });
+    await expect(drawer.getByText("Request & Response")).toBeVisible({
+      timeout: 20_000,
+    });
+
+    // The prompt we sent and the mock server's reply are both rendered.
+    await expect(drawer.getByText(prompt, { exact: false })).toBeVisible({
+      timeout: 20_000,
+    });
+    await expect(drawer.getByText(MOCK_RESPONSE_TEXT, { exact: false }).first()).toBeVisible({ timeout: 20_000 });
+  });
+
+  // Split out because only the copy path needs a secure context; folding it in would
+  // take the drawer-rendering coverage down with it.
+  test("the drawer copies the request and the response to the clipboard", async ({ page, request }) => {
+    // `navigator.clipboard` is undefined outside a secure context, and handleCopy calls
+    // writeText unguarded, so on plain HTTP served from a hostname the click throws and no
+    // toast renders. Skipped rather than weakened so the product gap stays visible.
+    await page.goto("/ui");
+    const isSecure = await page.evaluate(() => window.isSecureContext);
+    test.skip(!isSecure, "origin is not a secure context, so navigator.clipboard is unavailable");
+
+    const prompt = `logs-copy-prompt-${uniqueSuffix()}`;
+    const requestId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt,
+    });
+    await waitForSpendLog(request, requestId);
+
+    const row = await openLogsForRequest(page, requestId);
+    await row.click();
+    const drawer = page.locator(".ant-drawer-content").first();
+    await expect(drawer).toBeVisible({ timeout: 20_000 });
+
+    // Copy request: the Input card's copy button puts the prompt on the clipboard.
+    await sectionHeader(drawer, "Input").getByRole("button").click();
+    await expect(page.getByText("Input copied")).toBeVisible({
+      timeout: 10_000,
+    });
+    expect(await page.evaluate(() => navigator.clipboard.readText())).toContain(prompt);
+
+    // Copy response: the Output card's copy button puts the completion on it.
+    await sectionHeader(drawer, "Output").getByRole("button").click();
+    await expect(page.getByText("Output copied")).toBeVisible({
+      timeout: 10_000,
+    });
+    expect(await page.evaluate(() => navigator.clipboard.readText())).toContain(MOCK_RESPONSE_TEXT);
+  });
+
+  test("the Input card collapses and expands", async ({ page, request }) => {
+    const prompt = `logs-collapse-prompt-${uniqueSuffix()}`;
+    const requestId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt,
+    });
+    await waitForSpendLog(request, requestId);
+
+    const row = await openLogsForRequest(page, requestId);
+    await row.click();
+
+    const drawer = page.locator(".ant-drawer-content").first();
+    await expect(drawer.getByText("Request & Response")).toBeVisible({
+      timeout: 20_000,
+    });
+
+    // The body collapses via `max-height: 0; overflow: hidden`, which zeroes its own bounding
+    // box, so the wrapper reads as hidden while the clipped text node inside it does not.
+    const header = sectionHeader(drawer, "Input");
+    const body = header.locator("xpath=following-sibling::div[1]");
+    await expect(header.locator(".anticon-up")).toBeVisible();
+    await expect(body).toBeVisible();
+
+    await header.click();
+    await expect(header.locator(".anticon-down")).toBeVisible({
+      timeout: 10_000,
+    });
+    await expect(body).toBeHidden({ timeout: 10_000 });
+
+    await header.click();
+    await expect(header.locator(".anticon-up")).toBeVisible({
+      timeout: 10_000,
+    });
+    await expect(body).toBeVisible({ timeout: 10_000 });
+    await expect(drawer.getByText(prompt, { exact: false })).toBeVisible({
+      timeout: 10_000,
+    });
+  });
+
+  test("the JSON view exposes Request and Response tabs", async ({ page, request }) => {
+    const prompt = `logs-json-prompt-${uniqueSuffix()}`;
+    const requestId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt,
+    });
+    await waitForSpendLog(request, requestId);
+
+    const row = await openLogsForRequest(page, requestId);
+    await row.click();
+
+    const drawer = page.locator(".ant-drawer-content").first();
+    await expect(drawer.getByText("Request & Response")).toBeVisible({
+      timeout: 20_000,
+    });
+
+    // antd Radio.Button hides the <input> under its <label>, which intercepts
+    // the pointer event — click the label, not the radio.
+    await drawer.locator("label.ant-radio-button-wrapper").filter({ hasText: "JSON" }).click();
+
+    const requestTab = drawer.getByRole("tab", { name: "Request" });
+    await expect(requestTab).toBeVisible({ timeout: 10_000 });
+    await requestTab.click();
+    await expect(drawer.getByText(prompt, { exact: false }).first()).toBeVisible({ timeout: 10_000 });
+
+    await drawer.getByRole("tab", { name: "Response" }).click();
+    await expect(drawer.getByText(MOCK_RESPONSE_TEXT, { exact: false }).first()).toBeVisible({ timeout: 10_000 });
+  });
+
+  test("the End User filter narrows the table to that customer", async ({ page, request }) => {
+    const endUser = `logs-end-user-${uniqueSuffix()}`;
+    const minePrompt = `logs-filter-mine-${uniqueSuffix()}`;
+    const otherPrompt = `logs-filter-other-${uniqueSuffix()}`;
+
+    const mineId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt: minePrompt,
+      endUser,
+    });
+    const otherId = await sendChatCompletion(request, {
+      model: CHAT_MODEL_A,
+      prompt: otherPrompt,
+    });
+    await waitForSpendLog(request, mineId);
+    await waitForSpendLog(request, otherId);
+
+    await navigateToPage(page, Page.Logs);
+    await dismissFeedbackPopup(page);
+
+    // Both requests are in the unfiltered table.
+    await expect(requestLogsRows(page).filter({ hasText: mineId })).toHaveCount(1, { timeout: 30_000 });
+    await expect(requestLogsRows(page).filter({ hasText: otherId })).toHaveCount(1, { timeout: 30_000 });
+
+    await visibleTestId(page, "datatable-filters-trigger").click();
+    const filters = page.getByRole("dialog").filter({ hasText: "Narrow down request logs" });
+    await expect(filters).toBeVisible({ timeout: 10_000 });
+
+    const endUserInput = filters.getByPlaceholder("Search an end user");
+    await endUserInput.click();
+    await endUserInput.fill(endUser);
+    // The combobox popup is portaled to the body, so it is outside the filter
+    // dialog's subtree — scope the option lookup to the page, not the dialog.
+    await page.getByRole("option", { name: endUser, exact: true }).click({ timeout: 30_000 });
+    await filters.getByRole("button", { name: "Apply Filters" }).click();
+
+    // Only the request tagged with this end user survives the filter.
+    await expect(requestLogsRows(page).filter({ hasText: otherId })).toHaveCount(0, { timeout: 30_000 });
+    await expect(requestLogsRows(page).filter({ hasText: mineId })).toHaveCount(1);
+    await expect(requestLogsRows(page)).toHaveCount(1);
+  });
+});

tests/e2e/ui/tests/mcp/mcpServerEdit.spec.ts

@@ -0,0 +1,92 @@
+import { test, expect, type Page as PlaywrightPage } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { createMcpServer, deleteMcpServerByName } from "../../helpers/mcp";
+import { captureRequestBody, readBack } from "../../helpers/roundTrip";
+
+/**
+ * Editing and deleting an MCP server, verified against the API. The reported failures are all on
+ * this side: renames that need repeating, deletes that need two attempts, each toasting success on
+ * the failing attempt. The URL is unreachable on purpose; only persistence is under test here.
+ */
+const UNREACHABLE_URL = "https://e2e-fake-mcp.test.local/mcp";
+
+/** GET /v1/mcp/server returns a bare array of servers (useMCPServers types it MCPServer[]). */
+async function findServerByName(page: PlaywrightPage, serverName: string): Promise<Record<string, any> | undefined> {
+  const servers = await readBack<Record<string, any>[]>(page, "/v1/mcp/server");
+  return servers.find((server) => server.server_name === serverName);
+}
+
+test.describe("MCP Servers - edit and delete", () => {
+  test.use({ storageState: ADMIN_STORAGE_PATH });
+
+  let serverName: string;
+
+  test.beforeEach(async ({ page }) => {
+    serverName = await createMcpServer(page, UNREACHABLE_URL);
+  });
+
+  // The rename test leaves an unreachable server behind, which slows the MCP page for later tests.
+  test.afterEach(async ({ page }) => {
+    await deleteMcpServerByName(page, serverName);
+  });
+
+  test("Renaming a server's alias persists", async ({ page }) => {
+    const before = await findServerByName(page, serverName);
+    expect(before, `created server ${serverName} readable from /v1/mcp/server`).toBeTruthy();
+
+    await page.getByTestId("mcp-servers-grid").getByText(serverName).first().click();
+    await expect(page.getByRole("button", { name: /Back to All Servers/i })).toBeVisible({ timeout: 10_000 });
+
+    // exact: the server view also renders a "Network Settings" tab.
+    await page.getByRole("tab", { name: "Settings", exact: true }).click();
+
+    // A card click may land straight in edit mode, so only click the button when it rendered.
+    const editSettings = page.getByRole("button", { name: "Edit Settings" });
+    if (await editSettings.isVisible().catch(() => false)) {
+      await editSettings.click();
+    }
+
+    // The create modal stays mounted behind the view with its own #alias and Save.
+    const settingsPanel = page.getByRole("tabpanel", { name: "Settings" });
+
+    const newAlias = `${serverName}_renamed`;
+    const aliasInput = settingsPanel.locator('input[id="alias"]');
+    await expect(aliasInput).toBeVisible({ timeout: 10_000 });
+    await aliasInput.fill(newAlias);
+
+    const update = await captureRequestBody(page, { method: "PUT", urlIncludes: "/v1/mcp/server" }, async () => {
+      await settingsPanel.getByRole("button", { name: "Save Changes" }).click();
+    });
+    expect(update.alias, "new alias on the wire").toBe(newAlias);
+    // An unidentified target is one way a save succeeds and changes nothing.
+    expect(update.server_id, "update targets the server being edited").toBe(before?.server_id);
+
+    // The reported symptom is a first save that returns success and does not stick.
+    await expect
+      .poll(async () => (await findServerByName(page, serverName))?.alias, {
+        message: `alias for ${serverName} did not persist after one save`,
+        timeout: 15_000,
+      })
+      .toBe(newAlias);
+  });
+
+  test("Deleting a server removes it", async ({ page }) => {
+    expect(await findServerByName(page, serverName), `created server ${serverName} exists`).toBeTruthy();
+
+    const card = page.getByTestId("mcp-servers-grid").locator("div").filter({ hasText: serverName }).first();
+    await card.getByRole("button", { name: "Server actions" }).click();
+    await page.getByRole("menuitem", { name: "Delete" }).click();
+
+    const dialog = page.getByRole("alertdialog");
+    await expect(dialog.getByText("Delete MCP Server?")).toBeVisible({ timeout: 5_000 });
+    await dialog.getByRole("button", { name: "Delete", exact: true }).click();
+
+    // One attempt has to be enough; the report is a delete that needs two.
+    await expect
+      .poll(async () => await findServerByName(page, serverName), {
+        message: `server ${serverName} still present after one delete`,
+        timeout: 15_000,
+      })
+      .toBeUndefined();
+  });
+});

tests/e2e/ui/tests/mcp/mcpServers.spec.ts

@@ -2,6 +2,7 @@ import { test, expect } from "@playwright/test";
 import { ADMIN_STORAGE_PATH } from "../../constants";
 import { navigateToPage } from "../../helpers/navigation";
 import { Page } from "../../fixtures/pages";
+import { deleteMcpServerByName } from "../../helpers/mcp";
 
 // Coverage scope: only the happy-path Streamable HTTP + None auth create flow.
 // See E2E_COVERAGE.md (#29 row) for the full list of uncovered MCP surfaces
@@ -11,6 +12,15 @@ import { Page } from "../../fixtures/pages";
 test.describe("MCP Servers", () => {
   test.use({ storageState: ADMIN_STORAGE_PATH });
 
+  let createdServerName = "";
+
+  // The server this test creates is unreachable, and the MCP page contacts
+  // every server it lists, so leaving it behind slows down every later MCP
+  // test. See deleteMcpServerByName for what that actually cost.
+  test.afterEach(async ({ page }) => {
+    if (createdServerName) await deleteMcpServerByName(page, createdServerName);
+  });
+
   test("Add a custom MCP server via the discovery → custom form", async ({ page }) => {
     await navigateToPage(page, Page.McpServers);
 
@@ -25,6 +35,7 @@ test.describe("MCP Servers", () => {
 
     // Name — no spaces or hyphens per validateMCPServerName
     const uniqueName = `e2e_mcp_${Date.now()}`;
+    createdServerName = uniqueName;
     await formModal.locator('input[id="server_name"]').fill(uniqueName);
 
     // Transport: Streamable HTTP — the only value the proxy actually accepts is "http"
@@ -48,8 +59,6 @@ test.describe("MCP Servers", () => {
     // Submit
     await formModal.getByRole("button", { name: /^Add MCP Server$/ }).click();
 
-    // No teardown needed — the e2e runner spins up a fresh DB per invocation.
-
     // Success toast and the new card in the server grid. Scope the lookup to
     // the MCP servers grid so the form modal's `server_name` input — which
     // still holds the timestamped value during its close animation — can't

tests/e2e/ui/tests/mcp/mcpTools.spec.ts

@@ -0,0 +1,80 @@
+import { test, expect, Locator } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { createMcpServer, deleteMcpServerByName, openMcpToolsTab } from "../../helpers/mcp";
+
+// Listing and calling MCP tools, which needs a server that really answers; the create-only spec
+// points at an unreachable URL on purpose.
+//
+// This spec makes a read-only network call to DeepWiki's public MCP server, from the proxy rather
+// than the browser. It needs no credentials, so there is no secret to leak from a public repo.
+//
+// A DeepWiki outage turns this red for something that is not a litellm regression. That is left
+// visible rather than auto-skipped: skipping on connection trouble also skips when the proxy's own
+// MCP client breaks, which is the regression this exists to catch. E2E_SKIP_EXTERNAL_MCP=1 opts out.
+const MCP_SERVER_URL = "https://mcp.deepwiki.com/mcp";
+const TOOL_NAME = "read_wiki_structure";
+const TOOL_ARG_REPO = "BerriAI/litellm";
+
+// Match the h4 heading, not page text: a tool whose description names another tool trips strict mode.
+const toolCard = (list: Locator, name: string): Locator =>
+  list.locator("h4.font-mono").filter({ hasText: new RegExp(`^${name}$`) });
+
+test.describe("MCP Tools", () => {
+  test.use({ storageState: ADMIN_STORAGE_PATH });
+  test.skip(!!process.env.E2E_SKIP_EXTERNAL_MCP, "E2E_SKIP_EXTERNAL_MCP is set");
+
+  let serverName: string;
+
+  test.beforeEach(async ({ page }) => {
+    serverName = await createMcpServer(page, MCP_SERVER_URL);
+    await openMcpToolsTab(page, serverName);
+  });
+
+  // The MCP page contacts every server it lists, so leaks slow later tests run by run.
+  test.afterEach(async ({ page }) => {
+    await deleteMcpServerByName(page, serverName);
+  });
+
+  test("MCP Tools tab lists the tools the upstream server advertises", async ({ page }) => {
+    // Fetched through the proxy on mount, so allow for a cold upstream connection.
+    const toolList = page.locator(".mcp-tools-scrollable");
+    await expect(toolList).toBeVisible({ timeout: 30_000 });
+
+    // Non-empty would still pass if the proxy returned some other server's tools.
+    await expect(toolCard(toolList, TOOL_NAME)).toBeVisible();
+    await expect(toolCard(toolList, "ask_question")).toBeVisible();
+    await expect(toolCard(toolList, "read_wiki_contents")).toBeVisible();
+
+    // No other tool's name or description contains this string, so exactly one card survives.
+    await page.getByPlaceholder("Search tools...").fill(TOOL_NAME);
+    await expect(toolList.locator("h4.font-mono")).toHaveCount(1);
+    await expect(toolCard(toolList, TOOL_NAME)).toBeVisible();
+  });
+
+  test("Calling a tool from the Test Tool panel returns the upstream result", async ({ page }) => {
+    const toolList = page.locator(".mcp-tools-scrollable");
+    await expect(toolList).toBeVisible({ timeout: 30_000 });
+
+    await toolCard(toolList, TOOL_NAME).click();
+
+    // Selecting a tool swaps the right-hand pane in for the empty state.
+    await expect(page.getByText("Test Tool:", { exact: true })).toBeVisible({ timeout: 10_000 });
+    await expect(page.getByText("Ready to Call Tool")).toBeVisible();
+
+    // The form is generated from the tool's inputSchema, so `repoName` proves the schema
+    // round-tripped through the proxy instead of the panel falling back to a generic field.
+    const repoInput = page.locator('input[id="repoName"]');
+    await expect(repoInput).toBeVisible();
+    await repoInput.fill(TOOL_ARG_REPO);
+
+    await page.getByRole("button", { name: "Call Tool", exact: true }).click();
+
+    await expect(page.getByText("Tool executed successfully")).toBeVisible({ timeout: 60_000 });
+    // read_wiki_structure answers with the repo's outline, so the pane must name the repo.
+    await expect(page.getByText(TOOL_ARG_REPO).first()).toBeVisible();
+
+    // A second call is offered rather than the button resetting to its
+    // first-run label.
+    await expect(page.getByRole("button", { name: "Call Again", exact: true })).toBeVisible();
+  });
+});