From 676f841534e7c83bcf5d9afb65f5c37bf741af44 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 03:47:38 -0700
Subject: [PATCH 01/79] feat(mistral): add text-to-speech support for
 /v1/audio/speech

---
 .../mistral/audio_speech/transformation.py    | 209 ++++++++++++++++++
 litellm/main.py                               |  28 +++
 ...odel_prices_and_context_window_backup.json |   4 +-
 litellm/router.py                             |   4 +-
 litellm/utils.py                              |   6 +
 model_prices_and_context_window.json          |   4 +-
 ...est_mistral_audio_speech_transformation.py | 198 +++++++++++++++++
 tests/test_litellm/test_cost_calculator.py    |  12 +
 tests/test_litellm/test_main.py               |  28 +++
 tests/test_litellm/test_router.py             |  26 +++
 10 files changed, 513 insertions(+), 6 deletions(-)
 create mode 100644 litellm/llms/mistral/audio_speech/transformation.py
 create mode 100644 tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py

diff --git a/litellm/llms/mistral/audio_speech/transformation.py b/litellm/llms/mistral/audio_speech/transformation.py
new file mode 100644
index 00000000000..6d1a693268a
--- /dev/null
+++ b/litellm/llms/mistral/audio_speech/transformation.py
@@ -0,0 +1,209 @@
+"""
+Support for Mistral Voxtral text-to-speech via ``/v1/audio/speech``.
+
+API reference: https://docs.mistral.ai/api/#tag/audio/operation/audio_speech_v1_audio_speech_post
+"""
+
+import base64
+import json
+from collections.abc import Mapping
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final
+
+import httpx
+
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.base_llm.text_to_speech.transformation import (
+    BaseTextToSpeechConfig,
+    TextToSpeechRequestData,
+)
+from litellm.secret_managers.main import get_secret_str
+
+if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.types.llms.openai import HttpxBinaryResponseContent
+
+
+class MistralTextToSpeechException(BaseLLMException):
+    pass
+
+
+class MistralTextToSpeechConfig(BaseTextToSpeechConfig):
+    TTS_BASE_URL: Final[str] = "https://api.mistral.ai/v1"
+    AUDIO_CONTENT_TYPES: Final[MappingProxyType[str, str]] = MappingProxyType(
+        {
+            "mp3": "audio/mpeg",
+            "wav": "audio/wav",
+            "pcm": "audio/pcm",
+            "flac": "audio/flac",
+            "opus": "audio/ogg",
+        }
+    )
+    DROPPED_RESPONSE_HEADERS: Final[frozenset[str]] = frozenset(
+        {"content-encoding", "transfer-encoding", "content-length", "content-type"}
+    )
+    OPENAI_VOICE_ALIASES: Final[MappingProxyType[str, str]] = MappingProxyType(
+        {
+            "alloy": "en_paul_neutral",
+            "echo": "gb_oliver_neutral",
+            "fable": "en_paul_cheerful",
+            "onyx": "en_paul_confident",
+            "nova": "gb_jane_sarcasm",
+            "shimmer": "gb_jane_sarcasm",
+        }
+    )
+
+    def get_supported_openai_params(self, model: str) -> list:  # mutable-ok: base class contract returns a plain list
+        return ["voice", "response_format"]  # mutable-ok: base class contract returns a plain list
+
+    def _map_openai_voice(self, voice_id: str) -> str:
+        return self.OPENAI_VOICE_ALIASES.get(voice_id.lower(), voice_id)
+
+    def _resolve_voice_id(self, voice: object) -> str | None:
+        if isinstance(voice, str) and voice.strip():
+            return self._map_openai_voice(voice.strip())
+        if isinstance(voice, Mapping):
+            candidates: Final = (voice.get(key) for key in ("voice_id", "id", "name"))
+            resolved: Final = next(
+                (candidate.strip() for candidate in candidates if isinstance(candidate, str) and candidate.strip()),
+                None,
+            )
+            return self._map_openai_voice(resolved) if resolved else None
+        return None
+
+    def map_openai_params(
+        self,
+        model: str,
+        optional_params: Mapping[str, object],
+        voice: object = None,
+        drop_params: bool = False,
+        kwargs: Mapping[str, object] | None = None,
+    ) -> tuple[str | None, dict]:  # mutable-ok: base class contract returns a plain dict
+        response_format: Final = optional_params.get("response_format")
+        ref_audio: Final = kwargs.get("ref_audio") if kwargs else None
+        voice_id_kwarg: Final = kwargs.get("voice_id") if kwargs else None
+        mapped_voice: Final = self._resolve_voice_id(voice) or self._resolve_voice_id(voice_id_kwarg)
+        mapped_params: Final = {  # mutable-ok: base class contract returns a plain dict
+            key: value
+            for key, value in (("response_format", response_format), ("ref_audio", ref_audio))
+            if isinstance(value, str)
+        }
+        return mapped_voice, mapped_params
+
+    def validate_environment(
+        self,
+        headers: Mapping[str, str],
+        model: str,
+        api_key: str | None = None,
+        api_base: str | None = None,
+    ) -> dict:  # mutable-ok: base class contract returns a plain dict
+        resolved_key: Final = api_key or get_secret_str("MISTRAL_API_KEY")
+        if resolved_key is None:
+            raise MistralTextToSpeechException(
+                status_code=401,
+                message="Mistral API key is required. Set MISTRAL_API_KEY or pass api_key.",
+            )
+        return {  # mutable-ok: base class contract returns a plain dict
+            **headers,
+            "Authorization": f"Bearer {resolved_key}",
+            "Content-Type": "application/json",
+        }
+
+    def get_complete_url(
+        self,
+        model: str,
+        api_base: str | None,
+        litellm_params: Mapping[str, object],
+    ) -> str:
+        base_url: Final = api_base or get_secret_str("MISTRAL_API_BASE") or self.TTS_BASE_URL
+        return f"{base_url.rstrip('/')}/audio/speech"
+
+    def transform_text_to_speech_request(
+        self,
+        model: str,
+        input: str,
+        voice: str | None,
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        headers: Mapping[str, str],
+    ) -> TextToSpeechRequestData:
+        response_format: Final = optional_params.get("response_format")
+        ref_audio: Final = optional_params.get("ref_audio")
+        request_data: Final[TextToSpeechRequestData] = {
+            "dict_body": {
+                "model": model,
+                "input": input,
+                **({"voice_id": voice} if voice else {}),
+                **({"response_format": response_format} if isinstance(response_format, str) else {}),
+                **({"ref_audio": ref_audio} if isinstance(ref_audio, str) else {}),
+            },
+            "headers": {"Content-Type": "application/json"},
+        }
+        return request_data
+
+    def _requested_content_type(self, request: httpx.Request) -> str:
+        request_body: Final = json.loads(request.content or b"{}")
+        requested_format: Final = request_body.get("response_format")
+        if not isinstance(requested_format, str):
+            return "audio/mpeg"
+        return self.AUDIO_CONTENT_TYPES.get(requested_format, "audio/mpeg")
+
+    def transform_text_to_speech_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        logging_obj: "LiteLLMLoggingObj",
+    ) -> "HttpxBinaryResponseContent":
+        from litellm.types.llms.openai import HttpxBinaryResponseContent
+
+        try:
+            response_json: Final = raw_response.json()
+        except (json.JSONDecodeError, ValueError):
+            raise MistralTextToSpeechException(
+                status_code=raw_response.status_code,
+                message=f"Non-JSON response from Mistral speech API: {raw_response.text[:500]}",
+                headers=raw_response.headers,
+            )
+        audio_b64: Final = response_json.get("audio_data")
+        if not isinstance(audio_b64, str) or not audio_b64:
+            raise MistralTextToSpeechException(
+                status_code=500,
+                message=f"No audio_data in Mistral speech response. Response keys: {tuple(response_json.keys())}",
+                headers=raw_response.headers,
+            )
+        try:
+            audio_bytes: Final = base64.b64decode(audio_b64)
+        except ValueError:
+            raise MistralTextToSpeechException(
+                status_code=500,
+                message="Invalid base64 audio_data in Mistral speech response.",
+                headers=raw_response.headers,
+            )
+        retained_headers: Final = tuple(
+            (key, value)
+            for key, value in raw_response.headers.items()
+            if key.lower() not in self.DROPPED_RESPONSE_HEADERS
+        )
+        response_headers: Final = retained_headers + (
+            ("content-length", str(len(audio_bytes))),
+            ("content-type", self._requested_content_type(raw_response.request)),
+        )
+        binary_response: Final = httpx.Response(
+            status_code=200,
+            headers=response_headers,
+            content=audio_bytes,
+            request=raw_response.request,
+        )
+        return HttpxBinaryResponseContent(binary_response)
+
+    def get_error_class(
+        self,
+        error_message: str,
+        status_code: int,
+        headers: dict | httpx.Headers,  # mutable-ok: BaseLLMException takes a plain dict or httpx.Headers
+    ) -> BaseLLMException:
+        return MistralTextToSpeechException(
+            message=error_message,
+            status_code=status_code,
+            headers=headers,
+        )
diff --git a/litellm/main.py b/litellm/main.py
index cafa1e4718f..692df23b3f9 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -8367,6 +8367,34 @@ def speech(
             client=client,
             _is_async=aspeech or False,
         )
+    elif custom_llm_provider == "mistral":
+        from litellm.llms.mistral.audio_speech.transformation import (
+            MistralTextToSpeechConfig,
+        )
+
+        mistral_tts_config: Final = text_to_speech_provider_config or MistralTextToSpeechConfig()
+
+        if api_base is not None:
+            litellm_params_dict["api_base"] = api_base
+        if api_key is not None:
+            litellm_params_dict["api_key"] = api_key
+
+        mistral_voice: Final[str | None] = voice if isinstance(voice, str) else None
+
+        response = base_llm_http_handler.text_to_speech_handler(
+            model=model,
+            input=input,
+            voice=mistral_voice,
+            text_to_speech_provider_config=mistral_tts_config,
+            text_to_speech_optional_params=optional_params,
+            custom_llm_provider=custom_llm_provider,
+            litellm_params=litellm_params_dict,
+            logging_obj=logging_obj,
+            timeout=timeout,
+            extra_headers=extra_headers,
+            client=client,
+            _is_async=aspeech or False,
+        )
     elif custom_llm_provider == "aws_polly":
         from litellm.llms.aws_polly.text_to_speech.transformation import (
             AWSPollyTextToSpeechConfig,
diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json
index bebbcc32181..620fe2f8030 100644
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@@ -31126,9 +31126,9 @@
         "supports_audio_input": true
     },
     "mistral/voxtral-mini-tts-latest": {
+        "input_cost_per_character": 1.6e-05,
         "litellm_provider": "mistral",
         "mode": "audio_speech",
-        "output_cost_per_character": 1.6e-05,
         "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03",
         "supported_endpoints": [
             "/v1/audio/speech"
@@ -51677,9 +51677,9 @@
         "supports_audio_input": true
     },
     "mistral/voxtral-mini-tts-2603": {
+        "input_cost_per_character": 1.6e-05,
         "litellm_provider": "mistral",
         "mode": "audio_speech",
-        "output_cost_per_character": 1.6e-05,
         "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03",
         "supported_endpoints": [
             "/v1/audio/speech"
diff --git a/litellm/router.py b/litellm/router.py
index c93c1753f0e..d6ec5e57467 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -4270,7 +4270,7 @@ class Router:
                 self.fail_calls[model_name] += 1
             raise e
 
-    async def aspeech(self, model: str, input: str, voice: str, **kwargs):
+    async def aspeech(self, model: str, input: str, voice: str | None = None, **kwargs):
         """
         Example Usage:
 
@@ -4322,7 +4322,7 @@ class Router:
             )
             raise e
 
-    async def _aspeech(self, model: str, input: str, voice: str, **kwargs):
+    async def _aspeech(self, model: str, input: str, voice: str | None = None, **kwargs):
         model_name: Final = model
         try:
             verbose_router_logger.debug("Inside _aspeech()- model: %s; kwargs: %s", model, kwargs)
diff --git a/litellm/utils.py b/litellm/utils.py
index 5cd9bfc5f32..74a9b4ce935 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -9408,6 +9408,12 @@ class ProviderConfigManager:
             )
 
             return MinimaxTextToSpeechConfig()
+        elif litellm.LlmProviders.MISTRAL == provider:
+            from litellm.llms.mistral.audio_speech.transformation import (
+                MistralTextToSpeechConfig,
+            )
+
+            return MistralTextToSpeechConfig()
         elif litellm.LlmProviders.AWS_POLLY == provider:
             from litellm.llms.aws_polly.text_to_speech.transformation import (
                 AWSPollyTextToSpeechConfig,
diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json
index bebbcc32181..620fe2f8030 100644
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@@ -31126,9 +31126,9 @@
         "supports_audio_input": true
     },
     "mistral/voxtral-mini-tts-latest": {
+        "input_cost_per_character": 1.6e-05,
         "litellm_provider": "mistral",
         "mode": "audio_speech",
-        "output_cost_per_character": 1.6e-05,
         "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03",
         "supported_endpoints": [
             "/v1/audio/speech"
@@ -51677,9 +51677,9 @@
         "supports_audio_input": true
     },
     "mistral/voxtral-mini-tts-2603": {
+        "input_cost_per_character": 1.6e-05,
         "litellm_provider": "mistral",
         "mode": "audio_speech",
-        "output_cost_per_character": 1.6e-05,
         "source": "https://docs.mistral.ai/models/model-cards/voxtral-tts-26-03",
         "supported_endpoints": [
             "/v1/audio/speech"
diff --git a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
new file mode 100644
index 00000000000..20d07699cb8
--- /dev/null
+++ b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
@@ -0,0 +1,198 @@
+import base64
+from typing import Final
+from unittest.mock import MagicMock
+
+import httpx
+import pytest
+
+import litellm
+from litellm.llms.base_llm.text_to_speech.transformation import BaseTextToSpeechConfig
+from litellm.llms.mistral.audio_speech.transformation import (
+    MistralTextToSpeechConfig,
+    MistralTextToSpeechException,
+)
+from litellm.utils import ProviderConfigManager
+
+SPEECH_URL: Final = "https://api.mistral.ai/v1/audio/speech"
+
+
+def test_mistral_text_to_speech_config_installed():
+    config: Final = ProviderConfigManager.get_provider_text_to_speech_config(
+        model="voxtral-mini-tts-2603",
+        provider=litellm.LlmProviders.MISTRAL,
+    )
+    assert isinstance(config, BaseTextToSpeechConfig)
+    assert isinstance(config, MistralTextToSpeechConfig)
+
+
+def test_map_openai_params_drops_speed_and_instructions():
+    config: Final = MistralTextToSpeechConfig()
+    voice, params = config.map_openai_params(
+        model="voxtral-mini-tts-2603",
+        optional_params={"response_format": "wav", "speed": 1.5, "instructions": "sound cheerful"},
+        voice="en_paul_neutral",
+    )
+    assert voice == "en_paul_neutral"
+    assert params == {"response_format": "wav"}
+
+
+def test_map_openai_params_accepts_voice_dict_and_ref_audio():
+    config: Final = MistralTextToSpeechConfig()
+    voice, params = config.map_openai_params(
+        model="voxtral-mini-tts-2603",
+        optional_params={},
+        voice={"voice_id": "1f3a8b0c-voice-uuid"},
+        kwargs={"ref_audio": "bXktdm9pY2Utc2FtcGxl"},
+    )
+    assert voice == "1f3a8b0c-voice-uuid"
+    assert params == {"ref_audio": "bXktdm9pY2Utc2FtcGxl"}
+
+
+def test_transform_request_builds_mistral_body():
+    config: Final = MistralTextToSpeechConfig()
+    data: Final = config.transform_text_to_speech_request(
+        model="voxtral-mini-tts-2603",
+        input="hello from litellm",
+        voice="en_paul_neutral",
+        optional_params={"response_format": "wav"},
+        litellm_params={},
+        headers={},
+    )
+    assert data["dict_body"] == {
+        "model": "voxtral-mini-tts-2603",
+        "input": "hello from litellm",
+        "voice_id": "en_paul_neutral",
+        "response_format": "wav",
+    }
+    assert data["headers"] == {"Content-Type": "application/json"}
+
+
+def test_transform_request_omits_voice_for_ref_audio_cloning():
+    config: Final = MistralTextToSpeechConfig()
+    data: Final = config.transform_text_to_speech_request(
+        model="voxtral-mini-tts-2603",
+        input="clone me",
+        voice=None,
+        optional_params={"ref_audio": "bXktdm9pY2Utc2FtcGxl"},
+        litellm_params={},
+        headers={},
+    )
+    assert data["dict_body"] == {
+        "model": "voxtral-mini-tts-2603",
+        "input": "clone me",
+        "ref_audio": "bXktdm9pY2Utc2FtcGxl",
+    }
+
+
+def test_get_complete_url_default_base(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("MISTRAL_API_BASE", raising=False)
+    config: Final = MistralTextToSpeechConfig()
+    url: Final = config.get_complete_url(model="voxtral-mini-tts-2603", api_base=None, litellm_params={})
+    assert url == SPEECH_URL
+
+
+def test_get_complete_url_custom_base():
+    config: Final = MistralTextToSpeechConfig()
+    url: Final = config.get_complete_url(
+        model="voxtral-mini-tts-2603",
+        api_base="https://custom.api.example.com/v1/",
+        litellm_params={},
+    )
+    assert url == "https://custom.api.example.com/v1/audio/speech"
+
+
+def test_validate_environment_sets_bearer_header():
+    config: Final = MistralTextToSpeechConfig()
+    headers: Final = config.validate_environment(
+        headers={"x-custom": "1"},
+        model="voxtral-mini-tts-2603",
+        api_key="sk-mistral-test",
+    )
+    assert headers == {
+        "x-custom": "1",
+        "Authorization": "Bearer sk-mistral-test",
+        "Content-Type": "application/json",
+    }
+
+
+def test_validate_environment_requires_key(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.delenv("MISTRAL_API_KEY", raising=False)
+    config: Final = MistralTextToSpeechConfig()
+    with pytest.raises(MistralTextToSpeechException, match="MISTRAL_API_KEY"):
+        config.validate_environment(headers={}, model="voxtral-mini-tts-2603")
+
+
+def test_transform_response_decodes_base64_audio():
+    config: Final = MistralTextToSpeechConfig()
+    audio_bytes: Final = b"RIFF-fake-wav-bytes"
+    raw_response: Final = httpx.Response(
+        200,
+        json={"audio_data": base64.b64encode(audio_bytes).decode()},
+        headers={"x-request-id": "req-123"},
+        request=httpx.Request(
+            "POST",
+            SPEECH_URL,
+            json={"model": "voxtral-mini-tts-2603", "input": "hi", "response_format": "wav"},
+        ),
+    )
+    result: Final = config.transform_text_to_speech_response(
+        model="voxtral-mini-tts-2603",
+        raw_response=raw_response,
+        logging_obj=MagicMock(),
+    )
+    assert result.content == audio_bytes
+    assert result.response.headers["content-type"] == "audio/wav"
+    assert result.response.headers["content-length"] == str(len(audio_bytes))
+    assert result.response.headers["x-request-id"] == "req-123"
+
+
+def test_transform_response_missing_audio_data_raises():
+    config: Final = MistralTextToSpeechConfig()
+    raw_response: Final = httpx.Response(
+        200,
+        json={"detail": "unexpected"},
+        request=httpx.Request("POST", SPEECH_URL, json={"model": "voxtral-mini-tts-2603", "input": "hi"}),
+    )
+    with pytest.raises(MistralTextToSpeechException, match="audio_data"):
+        config.transform_text_to_speech_response(
+            model="voxtral-mini-tts-2603",
+            raw_response=raw_response,
+            logging_obj=MagicMock(),
+        )
+
+
+def test_map_openai_params_maps_openai_voice_aliases():
+    config: Final = MistralTextToSpeechConfig()
+    alloy_voice, _ = config.map_openai_params(
+        model="voxtral-mini-tts-2603",
+        optional_params={},
+        voice="alloy",
+    )
+    nova_voice, _ = config.map_openai_params(
+        model="voxtral-mini-tts-2603",
+        optional_params={},
+        voice="Nova",
+    )
+    passthrough_voice, _ = config.map_openai_params(
+        model="voxtral-mini-tts-2603",
+        optional_params={},
+        voice="en_paul_happy",
+    )
+    assert alloy_voice == "en_paul_neutral"
+    assert nova_voice == "gb_jane_sarcasm"
+    assert passthrough_voice == "en_paul_happy"
+
+
+def test_transform_response_invalid_base64_raises():
+    config: Final = MistralTextToSpeechConfig()
+    raw_response: Final = httpx.Response(
+        status_code=200,
+        json={"audio_data": "!!!not-base64!!!"},
+        request=httpx.Request("POST", SPEECH_URL),
+    )
+    with pytest.raises(MistralTextToSpeechException, match="base64"):
+        config.transform_text_to_speech_response(
+            model="voxtral-mini-tts-2603",
+            raw_response=raw_response,
+            logging_obj=MagicMock(),
+        )
diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py
index 7c2174018e8..db08ee486bf 100644
--- a/tests/test_litellm/test_cost_calculator.py
+++ b/tests/test_litellm/test_cost_calculator.py
@@ -4473,3 +4473,15 @@ def test_explicit_pricing_precedes_private_provider_response_model(
     )
 
     assert selected == expected
+
+
+def test_cost_per_token_mistral_voxtral_tts_bills_per_input_character(_local_model_cost_map):
+    prompt_usd, completion_usd = cost_per_token(
+        model="voxtral-mini-tts-2603",
+        custom_llm_provider="mistral",
+        call_type="speech",
+        prompt_characters=1000,
+    )
+
+    assert prompt_usd == pytest.approx(1000 * 1.6e-05)
+    assert completion_usd == 0.0
diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py
index 8cf878d05d9..47293d9c413 100644
--- a/tests/test_litellm/test_main.py
+++ b/tests/test_litellm/test_main.py
@@ -3181,3 +3181,31 @@ def test_stream_chunk_builder_defers_cost_to_logging_obj_when_usage_cost_absent(
 
     assert response is not None
     assert response._hidden_params.get("response_cost") is None
+
+
+def test_speech_mistral_dispatches_and_decodes_audio(respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test")
+    audio_bytes: Final = b"ID3-fake-mp3-bytes"
+    mock_route: Final = respx_mock.post("https://api.mistral.ai/v1/audio/speech").mock(
+        return_value=httpx.Response(200, json={"audio_data": base64.b64encode(audio_bytes).decode()})
+    )
+
+    response: Final = litellm.speech(
+        model="mistral/voxtral-mini-tts-2603",
+        input="hello from litellm",
+        voice="en_paul_neutral",
+        response_format="wav",
+        speed=2,
+        instructions="sound cheerful",
+    )
+
+    assert mock_route.called
+    request_body: Final = json.loads(mock_route.calls.last.request.content)
+    assert request_body == {
+        "model": "voxtral-mini-tts-2603",
+        "input": "hello from litellm",
+        "voice_id": "en_paul_neutral",
+        "response_format": "wav",
+    }
+    assert mock_route.calls.last.request.headers["authorization"] == "Bearer sk-mistral-test"
+    assert response.content == audio_bytes
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 97286017ffe..388011b7d5d 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -11530,3 +11530,29 @@ class TestTierParamsTheTargetAccepts:
         accepted = router._tier_params_the_target_accepts("no-such-group", {"reasoning_effort": "max"}, {})
 
         assert accepted == {"reasoning_effort": "max"}
+
+
+@pytest.mark.asyncio
+async def test_router_aspeech_without_voice_dispatches_ref_audio_cloning(respx_mock, monkeypatch):
+    import base64
+
+    monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test")
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    audio_bytes = b"RIFFfake-wav-bytes"
+    respx_mock.post("https://api.mistral.ai/v1/audio/speech").respond(
+        json={"audio_data": base64.b64encode(audio_bytes).decode()}
+    )
+    router = Router(
+        model_list=[
+            {
+                "model_name": "voxtral-tts",
+                "litellm_params": {"model": "mistral/voxtral-mini-tts-2603"},
+            }
+        ]
+    )
+
+    response = await router.aspeech(model="voxtral-tts", input="clone me", ref_audio="ZmFrZQ==")
+
+    request_body = json.loads(respx_mock.calls.last.request.content)
+    assert request_body == {"model": "voxtral-mini-tts-2603", "input": "clone me", "ref_audio": "ZmFrZQ=="}
+    assert response.content == audio_bytes

From 7c4cf2dcffbbff32b087d7756d4c0c0a4c590a91 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 11:41:24 -0700
Subject: [PATCH 02/79] fix(mistral): reject malformed base64 audio_data with
 strict validation

---
 litellm/llms/mistral/audio_speech/transformation.py             | 2 +-
 .../audio_speech/test_mistral_audio_speech_transformation.py    | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/mistral/audio_speech/transformation.py b/litellm/llms/mistral/audio_speech/transformation.py
index 6d1a693268a..e7f7d510346 100644
--- a/litellm/llms/mistral/audio_speech/transformation.py
+++ b/litellm/llms/mistral/audio_speech/transformation.py
@@ -172,7 +172,7 @@ class MistralTextToSpeechConfig(BaseTextToSpeechConfig):
                 headers=raw_response.headers,
             )
         try:
-            audio_bytes: Final = base64.b64decode(audio_b64)
+            audio_bytes: Final = base64.b64decode(audio_b64, validate=True)
         except ValueError:
             raise MistralTextToSpeechException(
                 status_code=500,
diff --git a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
index 20d07699cb8..6d250901e50 100644
--- a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
+++ b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
@@ -187,7 +187,7 @@ def test_transform_response_invalid_base64_raises():
     config: Final = MistralTextToSpeechConfig()
     raw_response: Final = httpx.Response(
         status_code=200,
-        json={"audio_data": "!!!not-base64!!!"},
+        json={"audio_data": "QUJD!QUJD"},
         request=httpx.Request("POST", SPEECH_URL),
     )
     with pytest.raises(MistralTextToSpeechException, match="base64"):

From 7c91b0120fadccd2a97c5a8ae5db77d6d8f59ec8 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Wed, 2 Sep 2026 13:32:51 -0700
Subject: [PATCH 03/79] fix(mistral): ensure /v1 on the Voxtral TTS base URL

A host-only api_base or MISTRAL_API_BASE (the documented form, https://api.mistral.ai) built
https://api.mistral.ai/audio/speech and 404ed. Match the chat and OCR configs by appending /v1
when the configured base does not already end with it.
---
 .../mistral/audio_speech/transformation.py    |  5 +++--
 ...est_mistral_audio_speech_transformation.py | 19 +++++++++++++------
 2 files changed, 16 insertions(+), 8 deletions(-)

diff --git a/litellm/llms/mistral/audio_speech/transformation.py b/litellm/llms/mistral/audio_speech/transformation.py
index e7f7d510346..7f5a659bd08 100644
--- a/litellm/llms/mistral/audio_speech/transformation.py
+++ b/litellm/llms/mistral/audio_speech/transformation.py
@@ -115,8 +115,9 @@ class MistralTextToSpeechConfig(BaseTextToSpeechConfig):
         api_base: str | None,
         litellm_params: Mapping[str, object],
     ) -> str:
-        base_url: Final = api_base or get_secret_str("MISTRAL_API_BASE") or self.TTS_BASE_URL
-        return f"{base_url.rstrip('/')}/audio/speech"
+        configured_base: Final = (api_base or get_secret_str("MISTRAL_API_BASE") or self.TTS_BASE_URL).rstrip("/")
+        versioned_base: Final = configured_base if configured_base.endswith("/v1") else f"{configured_base}/v1"
+        return f"{versioned_base}/audio/speech"
 
     def transform_text_to_speech_request(
         self,
diff --git a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
index 6d250901e50..108d4107db1 100644
--- a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
+++ b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
@@ -91,16 +91,23 @@ def test_get_complete_url_default_base(monkeypatch: pytest.MonkeyPatch):
     assert url == SPEECH_URL
 
 
-def test_get_complete_url_custom_base():
+@pytest.mark.parametrize(
+    "api_base",
+    ["https://custom.api.example.com/v1/", "https://custom.api.example.com/v1", "https://custom.api.example.com"],
+)
+def test_get_complete_url_custom_base_always_versioned(api_base: str):
     config: Final = MistralTextToSpeechConfig()
-    url: Final = config.get_complete_url(
-        model="voxtral-mini-tts-2603",
-        api_base="https://custom.api.example.com/v1/",
-        litellm_params={},
-    )
+    url: Final = config.get_complete_url(model="voxtral-mini-tts-2603", api_base=api_base, litellm_params={})
     assert url == "https://custom.api.example.com/v1/audio/speech"
 
 
+def test_get_complete_url_host_only_env_base_gets_v1(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("MISTRAL_API_BASE", "https://api.mistral.ai")
+    config: Final = MistralTextToSpeechConfig()
+    url: Final = config.get_complete_url(model="voxtral-mini-tts-2603", api_base=None, litellm_params={})
+    assert url == SPEECH_URL
+
+
 def test_validate_environment_sets_bearer_header():
     config: Final = MistralTextToSpeechConfig()
     headers: Final = config.validate_environment(

From 0a2581c14caaa43f76f1db2399aa66a9812d63bf Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 00:30:07 -0700
Subject: [PATCH 04/79] fix(mistral): keep the deployment voice default and
 drop the unreachable api base fallback

Review turned up two real problems in the TTS path.

Router.aspeech forwarded voice=None whenever the caller omitted it, which overwrote a
voice set in the deployment's litellm_params, so a configured fallback voice was
ignored on voice-less requests. It now leaves the key alone when no voice is passed.

get_complete_url also fell back to MISTRAL_API_BASE, but speech() always receives a
non-null api_base from get_llm_provider, whose mistral branch only reads
MISTRAL_AZURE_API_BASE and otherwise hardcodes the public host. That branch could
never run, and its unit test asserted a behavior the real path does not have. The
working override is api_base on the deployment, now pinned by an end-to-end test
---
 .../mistral/audio_speech/transformation.py    |  2 +-
 litellm/router.py                             |  2 +-
 ...est_mistral_audio_speech_transformation.py | 10 +---
 tests/test_litellm/test_main.py               | 18 +++++++
 tests/test_litellm/test_router.py             | 50 +++++++++++++++++++
 5 files changed, 71 insertions(+), 11 deletions(-)

diff --git a/litellm/llms/mistral/audio_speech/transformation.py b/litellm/llms/mistral/audio_speech/transformation.py
index 7f5a659bd08..2b3264dc756 100644
--- a/litellm/llms/mistral/audio_speech/transformation.py
+++ b/litellm/llms/mistral/audio_speech/transformation.py
@@ -115,7 +115,7 @@ class MistralTextToSpeechConfig(BaseTextToSpeechConfig):
         api_base: str | None,
         litellm_params: Mapping[str, object],
     ) -> str:
-        configured_base: Final = (api_base or get_secret_str("MISTRAL_API_BASE") or self.TTS_BASE_URL).rstrip("/")
+        configured_base: Final = (api_base or self.TTS_BASE_URL).rstrip("/")
         versioned_base: Final = configured_base if configured_base.endswith("/v1") else f"{configured_base}/v1"
         return f"{versioned_base}/audio/speech"
 
diff --git a/litellm/router.py b/litellm/router.py
index ee3a3ced023..b1357374f5c 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -4422,7 +4422,7 @@ class Router:
                 **{
                     **data,
                     "input": input,
-                    "voice": voice,
+                    **({"voice": voice} if voice is not None else {}),
                     "client": model_client,
                     **kwargs,
                 }
diff --git a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
index 108d4107db1..d819a79cef1 100644
--- a/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
+++ b/tests/test_litellm/llms/mistral/audio_speech/test_mistral_audio_speech_transformation.py
@@ -84,8 +84,7 @@ def test_transform_request_omits_voice_for_ref_audio_cloning():
     }
 
 
-def test_get_complete_url_default_base(monkeypatch: pytest.MonkeyPatch):
-    monkeypatch.delenv("MISTRAL_API_BASE", raising=False)
+def test_get_complete_url_default_base():
     config: Final = MistralTextToSpeechConfig()
     url: Final = config.get_complete_url(model="voxtral-mini-tts-2603", api_base=None, litellm_params={})
     assert url == SPEECH_URL
@@ -101,13 +100,6 @@ def test_get_complete_url_custom_base_always_versioned(api_base: str):
     assert url == "https://custom.api.example.com/v1/audio/speech"
 
 
-def test_get_complete_url_host_only_env_base_gets_v1(monkeypatch: pytest.MonkeyPatch):
-    monkeypatch.setenv("MISTRAL_API_BASE", "https://api.mistral.ai")
-    config: Final = MistralTextToSpeechConfig()
-    url: Final = config.get_complete_url(model="voxtral-mini-tts-2603", api_base=None, litellm_params={})
-    assert url == SPEECH_URL
-
-
 def test_validate_environment_sets_bearer_header():
     config: Final = MistralTextToSpeechConfig()
     headers: Final = config.validate_environment(
diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py
index 9b5122c3f75..e01679048e6 100644
--- a/tests/test_litellm/test_main.py
+++ b/tests/test_litellm/test_main.py
@@ -3287,3 +3287,21 @@ def test_speech_mistral_dispatches_and_decodes_audio(respx_mock: respx.MockRoute
     }
     assert mock_route.calls.last.request.headers["authorization"] == "Bearer sk-mistral-test"
     assert response.content == audio_bytes
+
+
+def test_speech_mistral_routes_to_configured_api_base(respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test")
+    audio_bytes: Final = b"ID3-gateway-bytes"
+    gateway_route: Final = respx_mock.post("https://mistral.gateway.internal/v1/audio/speech").mock(
+        return_value=httpx.Response(200, json={"audio_data": base64.b64encode(audio_bytes).decode()})
+    )
+
+    response: Final = litellm.speech(
+        model="mistral/voxtral-mini-tts-2603",
+        input="hello from litellm",
+        voice="en_paul_neutral",
+        api_base="https://mistral.gateway.internal",
+    )
+
+    assert gateway_route.called
+    assert response.content == audio_bytes
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 048536d887f..014937b35cd 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -12278,6 +12278,56 @@ async def test_router_aspeech_without_voice_dispatches_ref_audio_cloning(respx_m
     assert response.content == audio_bytes
 
 
+@pytest.mark.asyncio
+async def test_router_aspeech_without_voice_keeps_deployment_default_voice(respx_mock, monkeypatch):
+    import base64
+
+    monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test")
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    audio_bytes = b"RIFFfake-wav-bytes"
+    respx_mock.post("https://api.mistral.ai/v1/audio/speech").respond(
+        json={"audio_data": base64.b64encode(audio_bytes).decode()}
+    )
+    router = Router(
+        model_list=[
+            {
+                "model_name": "voxtral-tts",
+                "litellm_params": {"model": "mistral/voxtral-mini-tts-2603", "voice": "en_paul_neutral"},
+            }
+        ]
+    )
+
+    await router.aspeech(model="voxtral-tts", input="use my default")
+
+    request_body = json.loads(respx_mock.calls.last.request.content)
+    assert request_body["voice_id"] == "en_paul_neutral"
+
+
+@pytest.mark.asyncio
+async def test_router_aspeech_request_voice_overrides_deployment_default(respx_mock, monkeypatch):
+    import base64
+
+    monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test")
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    audio_bytes = b"RIFFfake-wav-bytes"
+    respx_mock.post("https://api.mistral.ai/v1/audio/speech").respond(
+        json={"audio_data": base64.b64encode(audio_bytes).decode()}
+    )
+    router = Router(
+        model_list=[
+            {
+                "model_name": "voxtral-tts",
+                "litellm_params": {"model": "mistral/voxtral-mini-tts-2603", "voice": "en_paul_neutral"},
+            }
+        ]
+    )
+
+    await router.aspeech(model="voxtral-tts", input="override me", voice="gb_oliver_neutral")
+
+    request_body = json.loads(respx_mock.calls.last.request.content)
+    assert request_body["voice_id"] == "gb_oliver_neutral"
+
+
 class TestPreRoutingTierDrivesFallbacks:
     """#38832: a complexity/auto router picks a tier behind the router name, but fallback
     lookup stayed on the router name, so the tier's configured chain never ran and a

From c79d1d12ae75a8eaec1de15fe4f0de01330cc553 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 02:15:57 -0700
Subject: [PATCH 05/79] fix(ai-gateway): install a rustls crypto provider
 before dialing upstream WebSockets

The gateway's dependency graph turns on two rustls crypto backends at once:
reqwest's rustls-tls pulls in ring, and litellm-core's bedrock-auth pulls in
aws-lc-rs through aws-config. rustls 0.23 refuses to guess between them, so
ClientConfig::builder panics, and that is exactly how tokio-tungstenite builds
its TLS config. Every outbound WebSocket dial killed its tokio worker and the
client saw the socket vanish with no close frame.

reqwest and the AWS SDK both pick a provider explicitly, so only the tungstenite
path was affected. Route all three dial sites through one helper that installs
ring once per process before connecting.
---
 litellm-rust/Cargo.lock                       |  1 +
 litellm-rust/Cargo.toml                       |  1 +
 litellm-rust/crates/ai-gateway/Cargo.toml     |  3 +
 litellm-rust/crates/ai-gateway/src/io/mod.rs  |  1 +
 .../crates/ai-gateway/src/io/realtime.rs      |  6 +-
 .../crates/ai-gateway/src/io/responses_ws.rs  | 12 ++--
 litellm-rust/crates/ai-gateway/src/io/tls.rs  | 67 +++++++++++++++++++
 7 files changed, 84 insertions(+), 7 deletions(-)
 create mode 100644 litellm-rust/crates/ai-gateway/src/io/tls.rs

diff --git a/litellm-rust/Cargo.lock b/litellm-rust/Cargo.lock
index b3dac5ca935..2ed998174e4 100644
--- a/litellm-rust/Cargo.lock
+++ b/litellm-rust/Cargo.lock
@@ -1415,6 +1415,7 @@ dependencies = [
  "litellm-core",
  "pyo3",
  "reqwest",
+ "rustls 0.23.42",
  "serde",
  "serde_json",
  "sha2 0.10.9",
diff --git a/litellm-rust/Cargo.toml b/litellm-rust/Cargo.toml
index a13dd4c04b0..15074082981 100644
--- a/litellm-rust/Cargo.toml
+++ b/litellm-rust/Cargo.toml
@@ -26,6 +26,7 @@ pythonize = "0.29.0"
 rand = "0.8"
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "json", "rustls-tls", "http2", "stream"] }
 rstest = "0.26.1"
+rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 sha2 = "0.10"
diff --git a/litellm-rust/crates/ai-gateway/Cargo.toml b/litellm-rust/crates/ai-gateway/Cargo.toml
index e3dbdf24ce6..414abc2356d 100644
--- a/litellm-rust/crates/ai-gateway/Cargo.toml
+++ b/litellm-rust/crates/ai-gateway/Cargo.toml
@@ -19,6 +19,9 @@ litellm-core = { workspace = true, features = ["bedrock-auth"] }
 # reqwest (rustls + json) is used by io/ocr and ships realtime logs to the
 # Python proxy callbacks API.
 reqwest.workspace = true
+# rustls is a direct dependency so `io::tls` can install a process-level
+# crypto provider; see that module for why the graph needs one.
+rustls.workspace = true
 # `sync` powers the bounded mpsc channel the realtime logger drains.
 tokio = { workspace = true, features = ["rt-multi-thread", "macros", "net", "time", "sync"] }
 tokio-tungstenite.workspace = true
diff --git a/litellm-rust/crates/ai-gateway/src/io/mod.rs b/litellm-rust/crates/ai-gateway/src/io/mod.rs
index cce56dd2121..7098d67993f 100644
--- a/litellm-rust/crates/ai-gateway/src/io/mod.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/mod.rs
@@ -3,3 +3,4 @@ pub mod ocr;
 pub mod realtime;
 pub mod realtime_pool;
 pub mod responses_ws;
+pub(crate) mod tls;
diff --git a/litellm-rust/crates/ai-gateway/src/io/realtime.rs b/litellm-rust/crates/ai-gateway/src/io/realtime.rs
index 662f7328982..53d87848342 100644
--- a/litellm-rust/crates/ai-gateway/src/io/realtime.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/realtime.rs
@@ -23,10 +23,12 @@ use tokio_tungstenite::tungstenite::Message;
 use tokio_tungstenite::tungstenite::client::IntoClientRequest;
 use tokio_tungstenite::tungstenite::http::HeaderValue;
 use tokio_tungstenite::tungstenite::http::header::AUTHORIZATION;
-use tokio_tungstenite::{MaybeTlsStream, WebSocketStream, connect_async};
+use tokio_tungstenite::{MaybeTlsStream, WebSocketStream};
 
 use litellm_core::providers::openai::realtime::transformation::OPENAI_REALTIME_CONFIG;
 
+use crate::io::tls::connect_upstream;
+
 /// Environment variable holding the OpenAI API key (last-resort fallback).
 const OPENAI_API_KEY_ENV: &str = "OPENAI_API_KEY";
 
@@ -84,7 +86,7 @@ pub(crate) async fn dial_upstream(
             .map_err(|err| Error::Auth(err.to_string()))?,
     );
 
-    let (upstream, _response) = connect_async(request)
+    let (upstream, _response) = connect_upstream(request)
         .await
         .map_err(|err| Error::Network(err.to_string()))?;
     Ok(upstream)
diff --git a/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs b/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
index 0b01747b1a5..ee181791413 100644
--- a/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
@@ -14,7 +14,9 @@ use tokio_tungstenite::tungstenite::Message;
 use tokio_tungstenite::tungstenite::client::IntoClientRequest;
 use tokio_tungstenite::tungstenite::http::HeaderValue;
 use tokio_tungstenite::tungstenite::http::header::{AUTHORIZATION, HeaderName};
-use tokio_tungstenite::{MaybeTlsStream, WebSocketStream, connect_async};
+use tokio_tungstenite::{MaybeTlsStream, WebSocketStream};
+
+use crate::io::tls::connect_upstream;
 
 use crate::constants::{
     DEFAULT_RESPONSES_WS_CONNECT_TIMEOUT_SECS, DEFAULT_RESPONSES_WS_IDLE_TIMEOUT_SECS,
@@ -49,14 +51,14 @@ impl ResponsesWebSocketConnection {
                 .map_err(|error| Error::InvalidRequest(error.to_string()))?;
             request.headers_mut().insert(header_name, header_value);
         }
-        let connect = connect_async(request);
+        let connect = connect_upstream(request);
         let result = match timeout {
             Some(timeout) => tokio::time::timeout(timeout, connect).await.map_err(|_| {
                 Error::Network("Responses WebSocket connection timed out".to_string())
             })?,
             None => connect.await,
         };
-        let (socket, _) = result.map_err(|error| match error {
+        let (socket, _) = result.map_err(|error| match *error {
             tokio_tungstenite::tungstenite::Error::Http(response) => Error::Http {
                 status: response.status().as_u16(),
                 body: String::new(),
@@ -138,13 +140,13 @@ async fn dial_upstream(
     );
     let result = tokio::time::timeout(
         Duration::from_secs(DEFAULT_RESPONSES_WS_CONNECT_TIMEOUT_SECS),
-        connect_async(request),
+        connect_upstream(request),
     )
     .await
     .map_err(|_| Error::Network("Responses WebSocket connection timed out".to_string()))?;
     result
         .map(|(socket, _)| socket)
-        .map_err(|error| match error {
+        .map_err(|error| match *error {
             tokio_tungstenite::tungstenite::Error::Http(response) => Error::Http {
                 status: response.status().as_u16(),
                 body: String::new(),
diff --git a/litellm-rust/crates/ai-gateway/src/io/tls.rs b/litellm-rust/crates/ai-gateway/src/io/tls.rs
new file mode 100644
index 00000000000..e3754e44002
--- /dev/null
+++ b/litellm-rust/crates/ai-gateway/src/io/tls.rs
@@ -0,0 +1,67 @@
+//! Outbound WebSocket dials, with the rustls crypto provider settled first.
+//!
+//! `reqwest/rustls-tls` enables `rustls/ring` and `litellm-core`'s `bedrock-auth`
+//! enables `rustls/aws-lc-rs`, so `ClientConfig::builder()` — which is how
+//! `tokio-tungstenite` builds its TLS config — panics rather than guess between
+//! them. reqwest and the AWS SDK pick a provider explicitly and never panic.
+//!
+//! Installing from the dial rather than from a `main` also covers the `cdylib`
+//! the Python bridge loads, the tests, and the benches, none of which have one.
+//! ring is what reqwest already falls back to, so installing it changes no
+//! working path, and an embedder that installed its own provider first keeps it.
+
+use std::sync::Once;
+
+use tokio::net::TcpStream;
+use tokio_tungstenite::tungstenite::Error;
+use tokio_tungstenite::tungstenite::client::IntoClientRequest;
+use tokio_tungstenite::tungstenite::handshake::client::Response;
+use tokio_tungstenite::{MaybeTlsStream, WebSocketStream, connect_async};
+
+static INSTALL_CRYPTO_PROVIDER: Once = Once::new();
+
+pub(crate) fn ensure_crypto_provider() {
+    INSTALL_CRYPTO_PROVIDER.call_once(|| {
+        let _ = rustls::crypto::ring::default_provider().install_default();
+    });
+}
+
+pub(crate) async fn connect_upstream<R>(
+    request: R,
+) -> Result<(WebSocketStream<MaybeTlsStream<TcpStream>>, Response), Box<Error>>
+where
+    R: IntoClientRequest + Unpin,
+{
+    ensure_crypto_provider();
+    connect_async(request).await.map_err(Box::new)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::ensure_crypto_provider;
+
+    #[test]
+    fn client_config_builder_works_with_both_provider_features_enabled() {
+        ensure_crypto_provider();
+
+        assert!(rustls::crypto::CryptoProvider::get_default().is_some());
+
+        let config = rustls::ClientConfig::builder()
+            .with_root_certificates(rustls::RootCertStore::empty())
+            .with_no_client_auth();
+
+        assert!(!config.crypto_provider().cipher_suites.is_empty());
+    }
+
+    #[test]
+    fn ensure_crypto_provider_is_idempotent() {
+        ensure_crypto_provider();
+        let first = rustls::crypto::CryptoProvider::get_default().cloned();
+
+        ensure_crypto_provider();
+        let second = rustls::crypto::CryptoProvider::get_default().cloned();
+
+        assert!(first.is_some());
+        assert!(std::sync::Arc::ptr_eq(&first.unwrap(), &second.unwrap()));
+    }
+}

From 2c08e7abf8bbf1d32aab48a330f4aadd494b5dd0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 02:27:00 -0700
Subject: [PATCH 06/79] test(ai-gateway): pin the crypto provider to ring and
 prove the dial installs it

The two tests that shipped with the fix both called ensure_crypto_provider
themselves, so deleting the call from connect_upstream left the whole suite
green, and swapping ring for aws-lc-rs did too.

Adds an integration test, which gets its own process, that dials wss:// at a
local plain-TCP listener through the public Responses WebSocket entrypoint and
asserts an Err plus an installed provider. Without the install in the dial it
panics with the original CryptoProvider message. A unit test now compares the
installed provider's cipher suites and key-exchange groups against ring's, so
the choice of backend is pinned rather than assumed.

Also names tls12 in the workspace rustls features: it already arrives through
reqwest and tokio-rustls, so the graph is unchanged, but a direct dependency
should say it needs TLS 1.2 rather than inherit it.
---
 litellm-rust/Cargo.toml                       |  2 +-
 litellm-rust/crates/ai-gateway/src/io/tls.rs  | 43 +++++++++++++++++--
 .../tests/crypto_provider_wiring.rs           | 41 ++++++++++++++++++
 3 files changed, 81 insertions(+), 5 deletions(-)
 create mode 100644 litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs

diff --git a/litellm-rust/Cargo.toml b/litellm-rust/Cargo.toml
index 15074082981..2e2e8809b7f 100644
--- a/litellm-rust/Cargo.toml
+++ b/litellm-rust/Cargo.toml
@@ -26,7 +26,7 @@ pythonize = "0.29.0"
 rand = "0.8"
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "json", "rustls-tls", "http2", "stream"] }
 rstest = "0.26.1"
-rustls = { version = "0.23", default-features = false, features = ["ring", "std"] }
+rustls = { version = "0.23", default-features = false, features = ["ring", "std", "tls12"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 sha2 = "0.10"
diff --git a/litellm-rust/crates/ai-gateway/src/io/tls.rs b/litellm-rust/crates/ai-gateway/src/io/tls.rs
index e3754e44002..96544adffb7 100644
--- a/litellm-rust/crates/ai-gateway/src/io/tls.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/tls.rs
@@ -8,7 +8,7 @@
 //! Installing from the dial rather than from a `main` also covers the `cdylib`
 //! the Python bridge loads, the tests, and the benches, none of which have one.
 //! ring is what reqwest already falls back to, so installing it changes no
-//! working path, and an embedder that installed its own provider first keeps it.
+//! working path, and whoever installs into this rustls build first still wins.
 
 use std::sync::Once;
 
@@ -38,13 +38,32 @@ where
 
 #[cfg(test)]
 mod tests {
+    use rustls::crypto::CryptoProvider;
+
     use super::ensure_crypto_provider;
 
+    fn fingerprint(
+        provider: &CryptoProvider,
+    ) -> (Vec<rustls::CipherSuite>, Vec<rustls::NamedGroup>) {
+        (
+            provider
+                .cipher_suites
+                .iter()
+                .map(|suite| suite.suite())
+                .collect(),
+            provider
+                .kx_groups
+                .iter()
+                .map(|group| group.name())
+                .collect(),
+        )
+    }
+
     #[test]
     fn client_config_builder_works_with_both_provider_features_enabled() {
         ensure_crypto_provider();
 
-        assert!(rustls::crypto::CryptoProvider::get_default().is_some());
+        assert!(CryptoProvider::get_default().is_some());
 
         let config = rustls::ClientConfig::builder()
             .with_root_certificates(rustls::RootCertStore::empty())
@@ -53,13 +72,29 @@ mod tests {
         assert!(!config.crypto_provider().cipher_suites.is_empty());
     }
 
+    #[test]
+    fn installs_ring_rather_than_aws_lc_rs() {
+        ensure_crypto_provider();
+
+        let installed = CryptoProvider::get_default().expect("a provider is installed");
+
+        assert_eq!(
+            fingerprint(installed),
+            fingerprint(&rustls::crypto::ring::default_provider())
+        );
+        assert_ne!(
+            fingerprint(installed),
+            fingerprint(&rustls::crypto::aws_lc_rs::default_provider())
+        );
+    }
+
     #[test]
     fn ensure_crypto_provider_is_idempotent() {
         ensure_crypto_provider();
-        let first = rustls::crypto::CryptoProvider::get_default().cloned();
+        let first = CryptoProvider::get_default().cloned();
 
         ensure_crypto_provider();
-        let second = rustls::crypto::CryptoProvider::get_default().cloned();
+        let second = CryptoProvider::get_default().cloned();
 
         assert!(first.is_some());
         assert!(std::sync::Arc::ptr_eq(&first.unwrap(), &second.unwrap()));
diff --git a/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs b/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs
new file mode 100644
index 00000000000..db5698f8460
--- /dev/null
+++ b/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs
@@ -0,0 +1,41 @@
+//! Guards the wiring, not just the helper: the dial itself has to install the
+//! rustls provider, in a test binary where nothing else has installed one.
+
+use std::collections::HashMap;
+use std::time::Duration;
+
+use litellm_ai_gateway::io::responses_ws::ResponsesWebSocketConnection;
+use tokio::net::TcpListener;
+
+#[tokio::test]
+async fn dialing_wss_returns_an_error_instead_of_panicking() {
+    let listener = TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind a loopback port");
+    let port = listener
+        .local_addr()
+        .expect("read the bound address")
+        .port();
+
+    tokio::spawn(async move {
+        while let Ok((stream, _peer)) = listener.accept().await {
+            drop(stream);
+        }
+    });
+
+    let result = ResponsesWebSocketConnection::connect_url(
+        &format!("wss://127.0.0.1:{port}/"),
+        &HashMap::new(),
+        Some(Duration::from_secs(10)),
+    )
+    .await;
+
+    assert!(
+        result.is_err(),
+        "a plain TCP server cannot finish a TLS handshake"
+    );
+    assert!(
+        rustls::crypto::CryptoProvider::get_default().is_some(),
+        "the dial is what installs the process-wide provider"
+    );
+}

From 534ab1628dc462912179660774af395fdce50839 Mon Sep 17 00:00:00 2001
From: Riddhi04 <riddhimahajan04@gmail.com>
Date: Wed, 22 Jul 2026 15:58:37 +0400
Subject: [PATCH 07/79] fix(proxy): enforce model access checks on Bedrock
 passthrough routes

get_model_from_request could not resolve a model for /bedrock/... routes since it only checked the JSON body's model field and a small set of URL regexes, none matching Bedrock's passthrough path. This let common_checks skip the key/project model allowlist entirely for any Bedrock passthrough action (invoke, converse, and their streaming variants), while the same model was correctly blocked on /v1/chat/completions

Extract the model from the Bedrock endpoint path using the same helper the passthrough handler itself relies on, so the existing allowlist check applies uniformly across auth methods and call paths
---
 litellm/proxy/auth/auth_utils.py              | 11 ++++
 .../proxy/auth/test_auth_utils.py             | 50 +++++++++++++++++++
 2 files changed, 61 insertions(+)

diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py
index d6007a2d56e..43e60e49e3b 100644
--- a/litellm/proxy/auth/auth_utils.py
+++ b/litellm/proxy/auth/auth_utils.py
@@ -1981,6 +1981,17 @@ def get_model_from_request(
         if vertex_match:
             model = vertex_match.group(1)
 
+    if model is None and route.lower().startswith("/bedrock"):
+        from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
+            _extract_model_from_bedrock_endpoint,
+        )
+
+        bedrock_endpoint = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
+        try:
+            model = _extract_model_from_bedrock_endpoint(bedrock_endpoint)
+        except ValueError:
+            model = None
+
     return model
 
 
diff --git a/tests/test_litellm/proxy/auth/test_auth_utils.py b/tests/test_litellm/proxy/auth/test_auth_utils.py
index f513f397b64..f3426534c28 100644
--- a/tests/test_litellm/proxy/auth/test_auth_utils.py
+++ b/tests/test_litellm/proxy/auth/test_auth_utils.py
@@ -428,6 +428,56 @@ def test_get_model_from_request_openai_deployment_route_still_works():
     )
 
 
+def test_get_model_from_request_bedrock_converse_passthrough():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/bedrock/model/us.anthropic.claude-sonnet-4-6/converse",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
+def test_get_model_from_request_bedrock_invoke_passthrough():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/bedrock/model/us.anthropic.claude-sonnet-4-6/invoke",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
+def test_get_model_from_request_bedrock_v2_converse_stream_passthrough():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/bedrock/v2/model/us.anthropic.claude-sonnet-4-6/converse-stream",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
+def test_get_model_from_request_bedrock_model_id_with_slashes():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/bedrock/model/aws/anthropic/model-name/invoke",
+        )
+        == "aws/anthropic/model-name"
+    )
+
+
+def test_get_model_from_request_bedrock_unparseable_endpoint_returns_none():
+    assert (
+        get_model_from_request(
+            request_data={},
+            route="/bedrock/agents/some-agent-route",
+        )
+        is None
+    )
+
+
 def test_get_model_from_request_includes_file_endpoint_header_model():
     assert (
         get_model_from_request(

From 6981ccf0ca1c3e1937f71c3fce287c898a15feba Mon Sep 17 00:00:00 2001
From: Riddhi04 <riddhimahajan04@gmail.com>
Date: Fri, 24 Jul 2026 16:38:34 +0400
Subject: [PATCH 08/79] fix(proxy): make URL model authoritative for Bedrock
 path-routed passthrough actions

The allowlist check read model from the request body first, while
bedrock_llm_proxy_route dispatches purely on the path model for invoke,
converse, and their streaming variants. A caller could put an allowed
model in the JSON body while targeting a disallowed model in the URL and
slip past the check. count_tokens keeps reading from the body since its
route has no model segment in the path.
---
 litellm/proxy/auth/auth_utils.py              | 22 +++++-----
 .../proxy/auth/test_auth_utils.py             | 40 +++++++++++++++++++
 2 files changed, 53 insertions(+), 9 deletions(-)

diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py
index 43e60e49e3b..fd746d62e76 100644
--- a/litellm/proxy/auth/auth_utils.py
+++ b/litellm/proxy/auth/auth_utils.py
@@ -1981,16 +1981,20 @@ def get_model_from_request(
         if vertex_match:
             model = vertex_match.group(1)
 
-    if model is None and route.lower().startswith("/bedrock"):
-        from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
-            _extract_model_from_bedrock_endpoint,
-        )
-
+    if route.lower().startswith("/bedrock"):
         bedrock_endpoint = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
-        try:
-            model = _extract_model_from_bedrock_endpoint(bedrock_endpoint)
-        except ValueError:
-            model = None
+        is_bedrock_count_tokens_route = (
+            "count_tokens" in bedrock_endpoint.lower() or "count-tokens" in bedrock_endpoint.lower()
+        )
+        if not is_bedrock_count_tokens_route:
+            from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
+                _extract_model_from_bedrock_endpoint,
+            )
+
+            try:
+                model = _extract_model_from_bedrock_endpoint(bedrock_endpoint)
+            except ValueError:
+                pass
 
     return model
 
diff --git a/tests/test_litellm/proxy/auth/test_auth_utils.py b/tests/test_litellm/proxy/auth/test_auth_utils.py
index f3426534c28..61d1a308b9d 100644
--- a/tests/test_litellm/proxy/auth/test_auth_utils.py
+++ b/tests/test_litellm/proxy/auth/test_auth_utils.py
@@ -478,6 +478,46 @@ def test_get_model_from_request_bedrock_unparseable_endpoint_returns_none():
     )
 
 
+def test_get_model_from_request_bedrock_url_model_overrides_body_model():
+    assert (
+        get_model_from_request(
+            request_data={"model": "us.anthropic.claude-sonnet-4-6"},
+            route="/bedrock/model/us.anthropic.claude-opus-4-6-v1/converse",
+        )
+        == "us.anthropic.claude-opus-4-6-v1"
+    )
+
+
+def test_get_model_from_request_bedrock_invoke_url_model_overrides_body_model():
+    assert (
+        get_model_from_request(
+            request_data={"model": "us.anthropic.claude-sonnet-4-6"},
+            route="/bedrock/model/us.anthropic.claude-opus-4-6-v1/invoke",
+        )
+        == "us.anthropic.claude-opus-4-6-v1"
+    )
+
+
+def test_get_model_from_request_bedrock_count_tokens_uses_body_model():
+    assert (
+        get_model_from_request(
+            request_data={"model": "us.anthropic.claude-sonnet-4-6"},
+            route="/bedrock/v1/messages/count_tokens",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
+def test_get_model_from_request_bedrock_unparseable_endpoint_keeps_body_model():
+    assert (
+        get_model_from_request(
+            request_data={"model": "us.anthropic.claude-sonnet-4-6"},
+            route="/bedrock/agents/some-agent-route",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
 def test_get_model_from_request_includes_file_endpoint_header_model():
     assert (
         get_model_from_request(

From 32293295f8fe9811a18179b556e12cdb6c4a11bc Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 15:40:35 -0700
Subject: [PATCH 09/79] refactor(proxy): resolve the Bedrock route model
 through an early return

The Bedrock branch of get_model_from_request reassigned the already resolved
model binding. Move the route parsing into a helper that returns the URL model
or None so the caller picks between it and the body model without rebinding.
---
 litellm/proxy/auth/auth_utils.py | 29 ++++++++++++++++-------------
 1 file changed, 16 insertions(+), 13 deletions(-)

diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py
index fd746d62e76..12e6f75889e 100644
--- a/litellm/proxy/auth/auth_utils.py
+++ b/litellm/proxy/auth/auth_utils.py
@@ -1982,23 +1982,26 @@ def get_model_from_request(
             model = vertex_match.group(1)
 
     if route.lower().startswith("/bedrock"):
-        bedrock_endpoint = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
-        is_bedrock_count_tokens_route = (
-            "count_tokens" in bedrock_endpoint.lower() or "count-tokens" in bedrock_endpoint.lower()
-        )
-        if not is_bedrock_count_tokens_route:
-            from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
-                _extract_model_from_bedrock_endpoint,
-            )
-
-            try:
-                model = _extract_model_from_bedrock_endpoint(bedrock_endpoint)
-            except ValueError:
-                pass
+        bedrock_model: Final = _model_from_bedrock_route(route)
+        return model if bedrock_model is None else bedrock_model
 
     return model
 
 
+def _model_from_bedrock_route(route: str) -> str | None:
+    bedrock_endpoint: Final = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
+    if "count_tokens" in bedrock_endpoint.lower() or "count-tokens" in bedrock_endpoint.lower():
+        return None
+    from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
+        _extract_model_from_bedrock_endpoint,
+    )
+
+    try:
+        return _extract_model_from_bedrock_endpoint(bedrock_endpoint)
+    except ValueError:
+        return None
+
+
 def abbreviate_api_key(api_key: str) -> str:
     if len(api_key) < MINIMUM_CUSTOM_KEY_LENGTH:
         return "sk-..."

From bd9593b74d4a05e265225a43afa0fd6c8d837484 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 16:41:31 -0700
Subject: [PATCH 10/79] fix(proxy): share the Bedrock count-tokens predicate
 between auth and the passthrough handler

---
 litellm/proxy/auth/auth_utils.py                       |  7 ++++---
 .../llm_passthrough_endpoints.py                       |  7 +++++--
 tests/test_litellm/proxy/auth/test_auth_utils.py       | 10 ++++++++++
 3 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/litellm/proxy/auth/auth_utils.py b/litellm/proxy/auth/auth_utils.py
index 12e6f75889e..1e4836654a1 100644
--- a/litellm/proxy/auth/auth_utils.py
+++ b/litellm/proxy/auth/auth_utils.py
@@ -1989,13 +1989,14 @@ def get_model_from_request(
 
 
 def _model_from_bedrock_route(route: str) -> str | None:
-    bedrock_endpoint: Final = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
-    if "count_tokens" in bedrock_endpoint.lower() or "count-tokens" in bedrock_endpoint.lower():
-        return None
     from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import (
         _extract_model_from_bedrock_endpoint,
+        is_bedrock_count_tokens_endpoint,
     )
 
+    bedrock_endpoint: Final = re.sub(r"^/bedrock/", "", route, flags=re.IGNORECASE)
+    if is_bedrock_count_tokens_endpoint(bedrock_endpoint):
+        return None
     try:
         return _extract_model_from_bedrock_endpoint(bedrock_endpoint)
     except ValueError:
diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py
index 29f216fd450..820312ac0fd 100644
--- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py
+++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py
@@ -703,6 +703,10 @@ BEDROCK_ENDPOINT_ACTIONS: Final = {
 BEDROCK_STREAMING_ACTIONS: Final = {"invoke-with-response-stream", "converse-stream"}
 
 
+def is_bedrock_count_tokens_endpoint(endpoint: str) -> bool:
+    return "count_tokens" in endpoint or "count-tokens" in endpoint
+
+
 def _extract_model_from_bedrock_endpoint(endpoint: str) -> str:
     """
     Extract model name from Bedrock endpoint path.
@@ -977,8 +981,7 @@ async def bedrock_llm_proxy_route(
 
     request_body: Final = await _read_request_body(request=request)
 
-    # Special handling for count_tokens endpoints
-    if "count_tokens" in endpoint or "count-tokens" in endpoint:
+    if is_bedrock_count_tokens_endpoint(endpoint):
         return await handle_bedrock_count_tokens(
             endpoint=endpoint,
             request=request,
diff --git a/tests/test_litellm/proxy/auth/test_auth_utils.py b/tests/test_litellm/proxy/auth/test_auth_utils.py
index 61d1a308b9d..a996de4d40c 100644
--- a/tests/test_litellm/proxy/auth/test_auth_utils.py
+++ b/tests/test_litellm/proxy/auth/test_auth_utils.py
@@ -508,6 +508,16 @@ def test_get_model_from_request_bedrock_count_tokens_uses_body_model():
     )
 
 
+def test_get_model_from_request_bedrock_uppercase_count_tokens_segment_is_not_count_tokens():
+    assert (
+        get_model_from_request(
+            request_data={"model": "us.anthropic.claude-haiku-4-5-20251001-v1:0"},
+            route="/bedrock/model/us.anthropic.claude-sonnet-4-6/invoke/COUNT_TOKENS",
+        )
+        == "us.anthropic.claude-sonnet-4-6"
+    )
+
+
 def test_get_model_from_request_bedrock_unparseable_endpoint_keeps_body_model():
     assert (
         get_model_from_request(

From 533a1c959c21f659a78fb130b17c700a3d6d1a06 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Thu, 3 Sep 2026 17:42:46 -0700
Subject: [PATCH 11/79] fix(proxy): reserve budget for the Bedrock Converse
 prompt, not the context window

Resolving the model from the /bedrock path sends passthrough calls through
optimistic budget reservation, whose tokenizer cannot walk Converse content
blocks and so fell back to the model's max_input_tokens. Count those messages
as text and read inferenceConfig.maxTokens so a budgeted key reserves the
request's cost.
---
 .../spend_tracking/budget_reservation.py      | 37 ++++++++++++-------
 .../spend_tracking/test_budget_reservation.py | 31 +++++++++++++++-
 2 files changed, 53 insertions(+), 15 deletions(-)

diff --git a/litellm/proxy/spend_tracking/budget_reservation.py b/litellm/proxy/spend_tracking/budget_reservation.py
index 91d2ece7a51..5ed52a327d4 100644
--- a/litellm/proxy/spend_tracking/budget_reservation.py
+++ b/litellm/proxy/spend_tracking/budget_reservation.py
@@ -1362,12 +1362,15 @@ def _approximate_input_size(request_body: Mapping[str, object]) -> int:
 def _count_input_tokens(request_body: dict, model: str) -> int | None:
     try:
         if "messages" in request_body:
-            return litellm.token_counter(
-                model=model,
-                messages=request_body.get("messages") or [],
-                tools=request_body.get("tools"),
-                tool_choice=request_body.get("tool_choice"),
-            )
+            try:
+                return litellm.token_counter(
+                    model=model,
+                    messages=request_body.get("messages") or (),
+                    tools=request_body.get("tools"),
+                    tool_choice=request_body.get("tool_choice"),
+                )
+            except ValueError:
+                return _count_text_tokens(model=model, text=request_body.get("messages"))
         if "prompt" in request_body:
             return _count_text_tokens(model=model, text=request_body.get("prompt"))
         if "input" in request_body:
@@ -1415,11 +1418,7 @@ def _estimate_output_tokens(
     if _is_input_only_route(route=route):
         return 0
 
-    requested: int | None = None
-    for key in ("max_completion_tokens", "max_tokens", "max_output_tokens"):
-        requested = _to_int(request_body.get(key))
-        if requested is not None:
-            break
+    requested: Final = _requested_output_tokens(request_body)
 
     # Clamp at min(requested-or-default, model_max-or-default). Two purposes:
     # (1) Without an explicit cap we still need a finite reservation so the
@@ -1430,9 +1429,19 @@ def _estimate_output_tokens(
     #     at the cap — the model can only physically emit max_output_tokens
     #     anyway, so reserving more is both wasteful and a DoS surface.
     model_ceiling: Final = _to_int(model_info.get("max_output_tokens")) or DEFAULT_MAX_OUTPUT_TOKENS_FALLBACK
-    if requested is None:
-        requested = DEFAULT_MAX_OUTPUT_TOKENS_FALLBACK
-    return min(requested, model_ceiling)
+    return min(DEFAULT_MAX_OUTPUT_TOKENS_FALLBACK if requested is None else requested, model_ceiling)
+
+
+_OUTPUT_TOKEN_FIELDS: Final = ("max_completion_tokens", "max_tokens", "max_output_tokens")
+
+
+def _requested_output_tokens(request_body: Mapping[str, object]) -> int | None:
+    inference_config: Final = request_body.get("inferenceConfig")
+    candidates: Final = (
+        *(request_body.get(field) for field in _OUTPUT_TOKEN_FIELDS),
+        inference_config.get("maxTokens") if isinstance(inference_config, Mapping) else None,
+    )
+    return next((tokens for tokens in map(_to_int, candidates) if tokens is not None), None)
 
 
 def _count_text_tokens(model: str, text: object) -> int:
diff --git a/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py b/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py
index f65f68812a2..c7de8c943ad 100644
--- a/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py
+++ b/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py
@@ -5,7 +5,7 @@ import pytest
 from litellm.caching import DualCache
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
-from litellm.proxy.spend_tracking.budget_reservation import reserve_budget_for_request
+from litellm.proxy.spend_tracking.budget_reservation import estimate_request_max_cost, reserve_budget_for_request
 from litellm.proxy.utils import ProxyLogging
 
 TOKEN_COUNTING_ROUTES: Final = (
@@ -46,3 +46,32 @@ async def test_non_exempt_llm_route_still_reserves_budget():
 
     assert reservation is not None
     assert reservation["reserved_cost"] > 0
+
+
+BEDROCK_SONNET: Final = "us.anthropic.claude-sonnet-4-6"
+CONVERSE_BODY: Final = {
+    "messages": [{"role": "user", "content": [{"text": "Reply with one word: pong"}]}],
+    "inferenceConfig": {"maxTokens": 5},
+}
+INVOKE_BODY: Final = {
+    "anthropic_version": "bedrock-2023-05-31",
+    "max_tokens": 5,
+    "messages": [{"role": "user", "content": "Reply with one word: pong"}],
+}
+
+
+def test_bedrock_converse_body_reserves_the_prompt_not_the_context_window():
+    converse_cost: Final = estimate_request_max_cost(
+        request_body=CONVERSE_BODY,
+        route=f"/bedrock/model/{BEDROCK_SONNET}/converse",
+        llm_router=None,
+        input_token_counts={},
+    )
+    invoke_cost: Final = estimate_request_max_cost(
+        request_body=INVOKE_BODY,
+        route=f"/bedrock/model/{BEDROCK_SONNET}/invoke",
+        llm_router=None,
+        input_token_counts={},
+    )
+    assert converse_cost is not None and invoke_cost is not None
+    assert invoke_cost < converse_cost < 2 * invoke_cost

From f9a5f676a866012c24427fb8afdfa240031f70d9 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 13:16:39 -0700
Subject: [PATCH 12/79] docs(claude): have runs embed their own QA screenshots
 on visual changes

---
 CLAUDE.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index d9e9e8f1586..9c525ffe420 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -37,7 +37,7 @@ Same applies for filing bug reports and feature requests, with .github/ISSUE_TEM
 
 If you're resolving a linear ticket, in the "## Linear ticket" section of the PR, say "Resolves LIT-1234", replacing "LIT-1234" with the actual ticket id that you're resolving. If you don't have the ticket id, don't make one up or search for it. Just leave the section blank
 
-Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We prefer curl'ing a live proxy instance running on localhost:4000 (I like to run it with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`; the Admin UI dev server is `npm run dev` in `ui/litellm-dashboard`, served on port 3000) and showing both the command run and the output. Also, it should hit real LLM provider APIs, not mocks, and cost real $$$ because that is the most realistic test. The proof of fix should be exactly what the end user / customer would see / do. The run logs in PR #27703 is a prime example of how to do it (not a huge fan of using a python test script that future me and the team will have no visibility into; I prefer just curl commands or a short list of bash commands (e.g., using `for`)). If it's a UI thing, just tell me which URLs to go to (e.g., http://localhost:4000/ui/?page=logs), where to click, what fields to fill out, etc. along with the other commands to run in an ordered list, and I'll do it myself and post the screenshots after you make the PR
+Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We prefer curl'ing a live proxy instance running on localhost:4000 (I like to run it with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`; the Admin UI dev server is `npm run dev` in `ui/litellm-dashboard`, served on port 3000) and showing both the command run and the output. Also, it should hit real LLM provider APIs, not mocks, and cost real $$$ because that is the most realistic test. The proof of fix should be exactly what the end user / customer would see / do. The run logs in PR #27703 is a prime example of how to do it (not a huge fan of using a python test script that future me and the team will have no visibility into; I prefer just curl commands or a short list of bash commands (e.g., using `for`)). If it's a UI thing, or the main use case runs through a headful agentic coding tool like Claude Code or Codex, drive that surface yourself and embed your own before and after screenshots of it in the PR (the Admin UI page, or what the coding tool shows), next to an ordered list of the URLs to go to (e.g., http://localhost:4000/ui/?page=logs), where to click, and what fields to fill out so a reviewer can reproduce it
 
 If you ever write any human-facing text (pull requests, issues, commit messages, discussion posts, github comments, release notes, docs, etc.), always follow these guidelines to sound less AI-y:
 - don't use emojis

From 61bed7956689959c940fbb75d610c0dc7d5ffb44 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 17:18:17 -0700
Subject: [PATCH 13/79] feat(ci): add the cost map guard check

Replace test-model-map.yml with a pull_request_target guard that validates the
cost map, its backup, and its generated schema on every PR, and additionally
enforces the sync bot contract on litellm_cost_map_sync_* branches: only the
three cost map files may change, no model or field is removed, and the special
root keys stay untouched.
---
 .github/workflows/cost-map-guard.yml      |  45 +++++
 .github/workflows/test-model-map.yml      |  37 ----
 ci_cd/cost_map_guard.py                   | 146 ++++++++++++++++
 tests/test_litellm/test_cost_map_guard.py | 195 ++++++++++++++++++++++
 4 files changed, 386 insertions(+), 37 deletions(-)
 create mode 100644 .github/workflows/cost-map-guard.yml
 delete mode 100644 .github/workflows/test-model-map.yml
 create mode 100644 ci_cd/cost_map_guard.py
 create mode 100644 tests/test_litellm/test_cost_map_guard.py

diff --git a/.github/workflows/cost-map-guard.yml b/.github/workflows/cost-map-guard.yml
new file mode 100644
index 00000000000..61a56f1f47e
--- /dev/null
+++ b/.github/workflows/cost-map-guard.yml
@@ -0,0 +1,45 @@
+name: Cost map guard
+
+on: # zizmor: ignore[dangerous-triggers] runs the base branch's code only; the PR's cost map files are read as data and never executed
+  pull_request_target:
+    branches:
+      - main
+      - litellm_internal_staging
+      - litellm_oss_staging
+      - "litellm_**"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  cost-map-guard:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+      - name: Fetch the pull request head and its merge base
+        id: revisions
+        env:
+          GH_TOKEN: ${{ github.token }}
+          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          merge_base="$(gh api "repos/${GITHUB_REPOSITORY}/compare/${BASE_SHA}...${HEAD_SHA}" --jq '.merge_base_commit.sha')"
+          git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA"
+          echo "merge_base=$merge_base" >> "$GITHUB_OUTPUT"
+      - name: Set up uv
+        uses: ./.github/actions/setup-uv-with-retries
+        with:
+          version: "0.10.9"
+      - name: Run the guard
+        env:
+          MERGE_BASE: ${{ steps.revisions.outputs.merge_base }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          HEAD_REF: ${{ github.event.pull_request.head.ref }}
+        run: |
+          uv run --frozen python ci_cd/cost_map_guard.py --base "$MERGE_BASE" --head "$HEAD_SHA" --head-ref "$HEAD_REF"
diff --git a/.github/workflows/test-model-map.yml b/.github/workflows/test-model-map.yml
deleted file mode 100644
index c2770e5da4c..00000000000
--- a/.github/workflows/test-model-map.yml
+++ /dev/null
@@ -1,37 +0,0 @@
-name: Validate model_prices_and_context_window.json
-
-on:
-  pull_request:
-    branches:
-      - main
-      - litellm_internal_staging
-      - litellm_oss_staging
-      - "litellm_**"
-
-permissions:
-  contents: read
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
-
-jobs:
-  validate-model-prices-json:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          persist-credentials: false
-
-      - name: Validate model_prices_and_context_window.json
-        run: |
-          jq empty model_prices_and_context_window.json
-
-      - name: Set up uv
-        uses: ./.github/actions/setup-uv-with-retries
-        with:
-          version: "0.10.9"
-
-      - name: Check model_prices_and_context_window.schema.json is in sync
-        run: |
-          uv run --frozen python ci_cd/generate_model_prices_schema.py --check
diff --git a/ci_cd/cost_map_guard.py b/ci_cd/cost_map_guard.py
new file mode 100644
index 00000000000..50aa40ba220
--- /dev/null
+++ b/ci_cd/cost_map_guard.py
@@ -0,0 +1,146 @@
+"""Guard the cost map on pull requests.
+
+Every pull request gets the file checks: the three cost map files parse, the backup copy matches the root file,
+and the JSON schema is in sync and validates the map. Pull requests from the cost map sync bot (branches named
+litellm_cost_map_sync_*) additionally may only touch those three files and may only add or update models.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import subprocess
+import sys
+from collections.abc import Sequence
+from dataclasses import dataclass
+from typing import Final
+
+from generate_model_prices_schema import SPECIAL_ROOT_KEYS, build_schema, render, validation_errors
+
+COST_MAP_PATH: Final = "model_prices_and_context_window.json"
+BACKUP_PATH: Final = "litellm/model_prices_and_context_window_backup.json"
+SCHEMA_PATH: Final = "model_prices_and_context_window.schema.json"
+GUARDED_PATHS: Final = (COST_MAP_PATH, BACKUP_PATH, SCHEMA_PATH)
+BOT_BRANCH_PREFIX: Final = "litellm_cost_map_sync_"
+
+CostMap = dict[str, object]
+
+
+@dataclass(frozen=True, slots=True)
+class Snapshot:
+    cost_map: str
+    backup: str
+    schema: str
+
+
+def _parse_object(text: str, path: str) -> CostMap | str:
+    try:
+        parsed: Final = json.loads(text)
+    except json.JSONDecodeError as error:
+        return f"{path} is not valid JSON: {error}"
+    return parsed if isinstance(parsed, dict) else f"{path} must be a JSON object at the root"
+
+
+def _rendered_schema(cost_map: CostMap) -> str:
+    try:
+        return render(build_schema(cost_map))
+    except SystemExit as error:
+        return str(error)
+
+
+def _file_failures(head: Snapshot, head_map: CostMap) -> tuple[str, ...]:
+    schema_text: Final = _rendered_schema(head_map)
+    if not schema_text.startswith("{"):
+        return (schema_text,)
+    backup_failure: Final = (
+        ()
+        if head.backup == head.cost_map
+        else (f"{BACKUP_PATH} differs from {COST_MAP_PATH}; copy the root file over it",)
+    )
+    schema_failure: Final = (
+        ()
+        if head.schema == schema_text
+        else (
+            f"{SCHEMA_PATH} is out of sync with {COST_MAP_PATH}; "
+            "run `python ci_cd/generate_model_prices_schema.py` and commit the result",
+        )
+    )
+    return (
+        *backup_failure,
+        *schema_failure,
+        *(
+            f"{COST_MAP_PATH} does not validate against its schema: {error}"
+            for error in validation_errors(head_map, json.loads(schema_text))[:20]
+        ),
+    )
+
+
+def _entries(cost_map: CostMap) -> dict[str, dict[str, object]]:
+    return {key: entry for key, entry in cost_map.items() if isinstance(entry, dict)}
+
+
+def _bot_failures(base: Snapshot, head_map: CostMap, changed_files: Sequence[str]) -> tuple[str, ...]:
+    base_map: Final = _parse_object(base.cost_map, COST_MAP_PATH)
+    if isinstance(base_map, str):
+        return (f"merge base: {base_map}",)
+    base_entries: Final = _entries(base_map)
+    head_entries: Final = _entries(head_map)
+    removed_fields: Final = tuple(
+        f"{key}.{field}"
+        for key, entry in base_entries.items()
+        if key in head_entries
+        for field in entry
+        if field not in head_entries[key]
+    )
+    return (
+        *(
+            f"bot PRs may only change the cost map files, not {path}"
+            for path in changed_files
+            if path not in GUARDED_PATHS
+        ),
+        *(f"bot PRs may not remove models: {key}" for key in base_map if key not in head_map),
+        *(f"bot PRs may not remove fields: {ref}" for ref in removed_fields),
+        *(
+            f"bot PRs may not change {key}"
+            for key in sorted(SPECIAL_ROOT_KEYS)
+            if base_map.get(key) != head_map.get(key)
+        ),
+    )
+
+
+def guard_failures(base: Snapshot, head: Snapshot, changed_files: Sequence[str], bot: bool) -> tuple[str, ...]:
+    head_map: Final = _parse_object(head.cost_map, COST_MAP_PATH)
+    if isinstance(head_map, str):
+        return (head_map,)
+    return (*_file_failures(head, head_map), *(_bot_failures(base, head_map, changed_files) if bot else ()))
+
+
+def _git(*args: str) -> str:
+    result: Final = subprocess.run(("git", *args), check=False, capture_output=True, text=True)
+    return result.stdout if result.returncode == 0 else ""
+
+
+def snapshot(revision: str) -> Snapshot:
+    return Snapshot(*(_git("show", f"{revision}:{path}") for path in GUARDED_PATHS))
+
+
+def main(argv: Sequence[str]) -> int:
+    parser: Final = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--base", required=True, help="merge base of the pull request")
+    parser.add_argument("--head", required=True, help="head commit of the pull request")
+    parser.add_argument("--head-ref", required=True, help="head branch name of the pull request")
+    args: Final = parser.parse_args(argv)
+    bot: Final = args.head_ref.startswith(BOT_BRANCH_PREFIX)
+    changed_files: Final = tuple(_git("diff", "--name-only", args.base, args.head).splitlines())
+    failures: Final = guard_failures(snapshot(args.base), snapshot(args.head), changed_files, bot)
+    contract: Final = "bot contract enforced" if bot else "human PR, file checks only"
+    if failures:
+        print(f"cost map guard failed ({contract}):")
+        print("\n".join(f"- {failure}" for failure in failures))
+        return 1
+    print(f"cost map guard passed ({contract})")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))
diff --git a/tests/test_litellm/test_cost_map_guard.py b/tests/test_litellm/test_cost_map_guard.py
new file mode 100644
index 00000000000..1b4330ed62c
--- /dev/null
+++ b/tests/test_litellm/test_cost_map_guard.py
@@ -0,0 +1,195 @@
+import importlib.util
+import json
+import subprocess
+import sys
+from pathlib import Path
+from types import ModuleType
+from typing import Final
+
+import pytest
+
+ROOT: Final = Path(__file__).resolve().parents[2]
+CI_CD: Final = ROOT / "ci_cd"
+
+
+def _load(name: str) -> ModuleType:
+    spec = importlib.util.spec_from_file_location(name, CI_CD / f"{name}.py")
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+schema_module: Final = _load("generate_model_prices_schema")
+guard: Final = _load("cost_map_guard")
+
+MAP_FILES: Final = (guard.COST_MAP_PATH,)
+BOT_REF: Final = "litellm_cost_map_sync_2026-09-04T12-00Z"
+
+
+def _entry(price: float = 1e-06, **extra: object) -> dict[str, object]:
+    return {
+        "input_cost_per_token": price,
+        "output_cost_per_token": price * 2,
+        "litellm_provider": "openrouter",
+        "mode": "chat",
+        "max_tokens": 4096,
+        **extra,
+    }
+
+
+BASE_MAP: Final = {
+    "sample_spec": {"input_cost_per_token": "USD per prompt token"},
+    "fallback_generalizations": {"rules": [{"name": "r", "pattern": "^x"}]},
+    "openrouter/a": _entry(supports_vision=True),
+    "openrouter/b": _entry(2e-06),
+}
+
+
+def _serialize(cost_map: dict[str, object]) -> str:
+    return json.dumps(cost_map, indent=4, ensure_ascii=False) + "\n"
+
+
+def _snapshot(cost_map: dict[str, object], backup: str | None = None, schema: str | None = None) -> object:
+    text = _serialize(cost_map)
+    rendered = schema_module.render(schema_module.build_schema(cost_map))
+    return guard.Snapshot(
+        cost_map=text, backup=text if backup is None else backup, schema=rendered if schema is None else schema
+    )
+
+
+BASE: Final = _snapshot(BASE_MAP)
+
+
+def _failures(head: object, changed_files: tuple[str, ...] = MAP_FILES, bot: bool = True) -> tuple[str, ...]:
+    return guard.guard_failures(BASE, head, changed_files, bot)
+
+
+def test_in_sync_files_pass_for_humans_and_bots() -> None:
+    assert _failures(BASE, bot=False) == ()
+    assert _failures(BASE, bot=True) == ()
+
+
+def test_bot_may_add_and_reprice_models() -> None:
+    head = _snapshot({**BASE_MAP, "openrouter/a": _entry(9e-06, supports_vision=True), "openrouter/c": _entry()})
+    assert _failures(head) == ()
+
+
+def test_broken_json_is_reported() -> None:
+    head = guard.Snapshot(cost_map="{not json", backup="{not json", schema="{}")
+    assert _failures(head, bot=False) == (
+        f"{guard.COST_MAP_PATH} is not valid JSON: Expecting property name enclosed in double quotes: line 1 column 2 (char 1)",
+    )
+
+
+def test_non_object_root_is_reported() -> None:
+    head = guard.Snapshot(cost_map="[]", backup="[]", schema="{}")
+    assert _failures(head, bot=False) == (f"{guard.COST_MAP_PATH} must be a JSON object at the root",)
+
+
+def test_backup_drift_is_reported() -> None:
+    head = _snapshot(BASE_MAP, backup=_serialize({**BASE_MAP, "openrouter/b": _entry(3e-06)}))
+    assert [failure for failure in _failures(head, bot=False) if failure.startswith(guard.BACKUP_PATH)]
+
+
+def test_schema_out_of_sync_is_reported() -> None:
+    head = _snapshot({**BASE_MAP, "openrouter/c": _entry(supports_audio_input=True)}, schema=BASE.schema)
+    assert [failure for failure in _failures(head, bot=False) if failure.startswith(guard.SCHEMA_PATH)]
+
+
+def test_schema_validation_errors_are_reported() -> None:
+    head = _snapshot({**BASE_MAP, "openrouter/c": _entry(-1e-06)})
+    prefix = f"{guard.COST_MAP_PATH} does not validate against its schema: openrouter/c."
+    assert [failure.removeprefix(prefix).split(":")[0] for failure in _failures(head, bot=False)] == [
+        "input_cost_per_token",
+        "output_cost_per_token",
+    ]
+
+
+def test_unclassified_entry_key_is_reported() -> None:
+    text = _serialize({**BASE_MAP, "openrouter/c": _entry(weird_thing=1)})
+    head = guard.Snapshot(cost_map=text, backup=text, schema=BASE.schema)
+    (failure,) = _failures(head, bot=False)
+    assert "Unclassified keys" in failure and "weird_thing" in failure
+
+
+def test_bot_may_only_touch_the_cost_map_files() -> None:
+    changed = (*guard.GUARDED_PATHS, "litellm/utils.py", ".github/workflows/cost-map-guard.yml")
+    assert _failures(BASE, changed_files=changed, bot=False) == ()
+    assert _failures(BASE, changed_files=changed) == (
+        "bot PRs may only change the cost map files, not litellm/utils.py",
+        "bot PRs may only change the cost map files, not .github/workflows/cost-map-guard.yml",
+    )
+
+
+def test_bot_may_not_remove_models() -> None:
+    head = _snapshot({key: value for key, value in BASE_MAP.items() if key != "openrouter/b"})
+    assert _failures(head, bot=False) == ()
+    assert _failures(head) == ("bot PRs may not remove models: openrouter/b",)
+
+
+def test_bot_may_not_remove_fields() -> None:
+    head = _snapshot({**BASE_MAP, "openrouter/a": _entry()})
+    assert _failures(head, bot=False) == ()
+    assert _failures(head) == ("bot PRs may not remove fields: openrouter/a.supports_vision",)
+
+
+def test_bot_may_not_change_special_root_keys() -> None:
+    head = _snapshot({**BASE_MAP, "fallback_generalizations": {"rules": []}})
+    assert _failures(head, bot=False) == ()
+    assert _failures(head) == ("bot PRs may not change fallback_generalizations",)
+
+
+def _commit(repo: Path, cost_map: dict[str, object], message: str) -> str:
+    text = _serialize(cost_map)
+    (repo / guard.COST_MAP_PATH).write_text(text)
+    (repo / guard.BACKUP_PATH).parent.mkdir(exist_ok=True)
+    (repo / guard.BACKUP_PATH).write_text(text)
+    (repo / guard.SCHEMA_PATH).write_text(schema_module.render(schema_module.build_schema(cost_map)))
+    subprocess.run(("git", "add", "-A"), cwd=repo, check=True)
+    subprocess.run(
+        ("git", "-c", "user.name=t", "-c", "user.email=t@example.com", "commit", "-q", "-m", message),
+        cwd=repo,
+        check=True,
+    )
+    return subprocess.run(
+        ("git", "rev-parse", "HEAD"), cwd=repo, check=True, capture_output=True, text=True
+    ).stdout.strip()
+
+
+def _run_guard(repo: Path, base: str, head: str, head_ref: str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        (sys.executable, str(CI_CD / "cost_map_guard.py"), "--base", base, "--head", head, "--head-ref", head_ref),
+        cwd=repo,
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+
+
+@pytest.mark.parametrize(
+    ("head_ref", "expected_code", "expected_line"),
+    [
+        (BOT_REF, 1, "- bot PRs may not remove models: openrouter/b"),
+        ("litellm_fix_pricing", 0, "cost map guard passed (human PR, file checks only)"),
+    ],
+)
+def test_main_reads_both_revisions_from_git(
+    tmp_path: Path, head_ref: str, expected_code: int, expected_line: str
+) -> None:
+    subprocess.run(("git", "init", "-q", str(tmp_path)), check=True)
+    base = _commit(tmp_path, BASE_MAP, "base")
+    head = _commit(tmp_path, {key: value for key, value in BASE_MAP.items() if key != "openrouter/b"}, "head")
+    result = _run_guard(tmp_path, base, head, head_ref)
+    assert result.returncode == expected_code, result.stdout + result.stderr
+    assert expected_line in result.stdout.splitlines()
+
+
+def test_main_rejects_a_bot_pr_that_edits_code(tmp_path: Path) -> None:
+    subprocess.run(("git", "init", "-q", str(tmp_path)), check=True)
+    base = _commit(tmp_path, BASE_MAP, "base")
+    (tmp_path / "litellm" / "utils.py").write_text("print('hi')\n")
+    head = _commit(tmp_path, {**BASE_MAP, "openrouter/c": _entry()}, "head")
+    assert _run_guard(tmp_path, base, head, BOT_REF).returncode == 1
+    assert _run_guard(tmp_path, base, head, "litellm_fix_pricing").returncode == 0

From 6463994f78161ee4448380a5e9fb2ad28729b98f Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 17:29:10 -0700
Subject: [PATCH 14/79] fix(ai-gateway): dial upstream WebSockets over an
 explicit rustls provider

Build one ClientConfig that names ring and loads the native roots once, and hand it to every tokio-tungstenite dial as its connector instead of installing a process-wide default from the dial path. Each of the three dial sites gets a wss:// test that reproduces the panic if its connector is dropped.
---
 litellm-rust/Cargo.lock                       |   1 +
 litellm-rust/Cargo.toml                       |   1 +
 litellm-rust/crates/ai-gateway/Cargo.toml     |   5 +-
 .../crates/ai-gateway/src/io/realtime.rs      |  27 +++++
 .../crates/ai-gateway/src/io/responses_ws.rs  |  23 ++++
 litellm-rust/crates/ai-gateway/src/io/tls.rs  | 113 ++++++++++--------
 .../tests/crypto_provider_wiring.rs           |  19 ++-
 7 files changed, 134 insertions(+), 55 deletions(-)

diff --git a/litellm-rust/Cargo.lock b/litellm-rust/Cargo.lock
index 2ed998174e4..d214e80d818 100644
--- a/litellm-rust/Cargo.lock
+++ b/litellm-rust/Cargo.lock
@@ -1416,6 +1416,7 @@ dependencies = [
  "pyo3",
  "reqwest",
  "rustls 0.23.42",
+ "rustls-native-certs",
  "serde",
  "serde_json",
  "sha2 0.10.9",
diff --git a/litellm-rust/Cargo.toml b/litellm-rust/Cargo.toml
index 2e2e8809b7f..d3d25cbda8d 100644
--- a/litellm-rust/Cargo.toml
+++ b/litellm-rust/Cargo.toml
@@ -27,6 +27,7 @@ rand = "0.8"
 reqwest = { version = "0.12", default-features = false, features = ["blocking", "json", "rustls-tls", "http2", "stream"] }
 rstest = "0.26.1"
 rustls = { version = "0.23", default-features = false, features = ["ring", "std", "tls12"] }
+rustls-native-certs = "0.8"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 sha2 = "0.10"
diff --git a/litellm-rust/crates/ai-gateway/Cargo.toml b/litellm-rust/crates/ai-gateway/Cargo.toml
index 414abc2356d..82bedd0c8a0 100644
--- a/litellm-rust/crates/ai-gateway/Cargo.toml
+++ b/litellm-rust/crates/ai-gateway/Cargo.toml
@@ -19,9 +19,10 @@ litellm-core = { workspace = true, features = ["bedrock-auth"] }
 # reqwest (rustls + json) is used by io/ocr and ships realtime logs to the
 # Python proxy callbacks API.
 reqwest.workspace = true
-# rustls is a direct dependency so `io::tls` can install a process-level
-# crypto provider; see that module for why the graph needs one.
+# rustls and its root store are direct dependencies so `io::tls` can build the
+# one TLS config the outbound dials use; see that module for why it has to.
 rustls.workspace = true
+rustls-native-certs.workspace = true
 # `sync` powers the bounded mpsc channel the realtime logger drains.
 tokio = { workspace = true, features = ["rt-multi-thread", "macros", "net", "time", "sync"] }
 tokio-tungstenite.workspace = true
diff --git a/litellm-rust/crates/ai-gateway/src/io/realtime.rs b/litellm-rust/crates/ai-gateway/src/io/realtime.rs
index 53d87848342..207c31dffa0 100644
--- a/litellm-rust/crates/ai-gateway/src/io/realtime.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/realtime.rs
@@ -286,6 +286,33 @@ mod tests {
         serde_json::from_str(raw).expect("valid event json")
     }
 
+    /// The realtime dial has to reach a `wss://` upstream without a process-wide
+    /// crypto provider installed, which is what dialing through `io::tls` buys.
+    #[tokio::test]
+    async fn dial_upstream_over_wss_reports_an_error_instead_of_panicking() {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
+            .await
+            .expect("bind a loopback port");
+        let port = listener
+            .local_addr()
+            .expect("read the bound address")
+            .port();
+        tokio::spawn(async move {
+            while let Ok((stream, _peer)) = listener.accept().await {
+                drop(stream);
+            }
+        });
+
+        let result = dial_upstream(
+            "gpt-realtime",
+            "sk-test",
+            Some(&format!("wss://127.0.0.1:{port}")),
+        )
+        .await;
+
+        assert!(matches!(result, Err(Error::Network(_))));
+    }
+
     #[test]
     fn resolve_api_key_prefers_param_then_blank_falls_through() {
         assert_eq!(resolve_api_key(Some("sk-test")).unwrap(), "sk-test");
diff --git a/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs b/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
index ee181791413..9df3d0c6cc5 100644
--- a/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/responses_ws.rs
@@ -326,6 +326,29 @@ mod tests {
     use tokio::net::TcpListener;
     use tokio_tungstenite::accept_async;
 
+    /// The Responses dial has to reach a `wss://` upstream without a process-wide
+    /// crypto provider installed, which is what dialing through `io::tls` buys.
+    #[tokio::test]
+    async fn dial_upstream_over_wss_reports_an_error_instead_of_panicking() {
+        let listener = TcpListener::bind("127.0.0.1:0")
+            .await
+            .expect("bind a loopback port");
+        let port = listener
+            .local_addr()
+            .expect("read the bound address")
+            .port();
+        tokio::spawn(async move {
+            while let Ok((stream, _peer)) = listener.accept().await {
+                drop(stream);
+            }
+        });
+
+        let result =
+            dial_upstream("gpt-5", "sk-test", Some(&format!("wss://127.0.0.1:{port}"))).await;
+
+        assert!(matches!(result, Err(Error::Network(_))));
+    }
+
     async fn websocket_base() -> (String, tokio::task::JoinHandle<()>) {
         let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind");
         let address = listener.local_addr().expect("local address");
diff --git a/litellm-rust/crates/ai-gateway/src/io/tls.rs b/litellm-rust/crates/ai-gateway/src/io/tls.rs
index 96544adffb7..16fd11e2e79 100644
--- a/litellm-rust/crates/ai-gateway/src/io/tls.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/tls.rs
@@ -1,29 +1,54 @@
-//! Outbound WebSocket dials, with the rustls crypto provider settled first.
+//! Outbound WebSocket dials over a TLS config this crate builds once and owns.
 //!
 //! `reqwest/rustls-tls` enables `rustls/ring` and `litellm-core`'s `bedrock-auth`
-//! enables `rustls/aws-lc-rs`, so `ClientConfig::builder()` — which is how
-//! `tokio-tungstenite` builds its TLS config — panics rather than guess between
-//! them. reqwest and the AWS SDK pick a provider explicitly and never panic.
-//!
-//! Installing from the dial rather than from a `main` also covers the `cdylib`
-//! the Python bridge loads, the tests, and the benches, none of which have one.
-//! ring is what reqwest already falls back to, so installing it changes no
-//! working path, and whoever installs into this rustls build first still wins.
+//! enables `rustls/aws-lc-rs`, so the bare `ClientConfig::builder()` that
+//! `tokio-tungstenite` uses when handed no connector panics rather than guess
+//! between them. Naming ring on a connector of our own settles that for these
+//! dials without touching the process-wide default, and building the config
+//! once keeps the platform trust store, which `tokio-tungstenite` would
+//! otherwise re-read on every dial, off the dial path.
 
-use std::sync::Once;
+use std::io;
+use std::sync::{Arc, OnceLock};
 
+use rustls::{ClientConfig, RootCertStore};
 use tokio::net::TcpStream;
 use tokio_tungstenite::tungstenite::Error;
 use tokio_tungstenite::tungstenite::client::IntoClientRequest;
+use tokio_tungstenite::tungstenite::error::TlsError;
 use tokio_tungstenite::tungstenite::handshake::client::Response;
-use tokio_tungstenite::{MaybeTlsStream, WebSocketStream, connect_async};
+use tokio_tungstenite::{
+    Connector, MaybeTlsStream, WebSocketStream, connect_async_tls_with_config,
+};
 
-static INSTALL_CRYPTO_PROVIDER: Once = Once::new();
+static TLS_CONFIG: OnceLock<Arc<ClientConfig>> = OnceLock::new();
 
-pub(crate) fn ensure_crypto_provider() {
-    INSTALL_CRYPTO_PROVIDER.call_once(|| {
-        let _ = rustls::crypto::ring::default_provider().install_default();
-    });
+fn build_config() -> Result<ClientConfig, Box<Error>> {
+    let native = rustls_native_certs::load_native_certs();
+    let roots = {
+        let mut store = RootCertStore::empty();
+        let (added, _ignored) = store.add_parsable_certificates(native.certs);
+        if added == 0 {
+            return Err(Box::new(Error::Io(io::Error::other(format!(
+                "no usable native root certificates: {:?}",
+                native.errors
+            )))));
+        }
+        store
+    };
+
+    ClientConfig::builder_with_provider(Arc::new(rustls::crypto::ring::default_provider()))
+        .with_safe_default_protocol_versions()
+        .map(|builder| builder.with_root_certificates(roots).with_no_client_auth())
+        .map_err(|error| Box::new(Error::Tls(TlsError::Rustls(error))))
+}
+
+fn tls_config() -> Result<Arc<ClientConfig>, Box<Error>> {
+    if let Some(config) = TLS_CONFIG.get() {
+        return Ok(Arc::clone(config));
+    }
+    let built = Arc::new(build_config()?);
+    Ok(Arc::clone(TLS_CONFIG.get_or_init(|| built)))
 }
 
 pub(crate) async fn connect_upstream<R>(
@@ -32,19 +57,25 @@ pub(crate) async fn connect_upstream<R>(
 where
     R: IntoClientRequest + Unpin,
 {
-    ensure_crypto_provider();
-    connect_async(request).await.map_err(Box::new)
+    let request = request.into_client_request().map_err(Box::new)?;
+    let connector = match request.uri().scheme_str() {
+        Some("wss") => Some(Connector::Rustls(tls_config()?)),
+        _ => None,
+    };
+    connect_async_tls_with_config(request, None, false, connector)
+        .await
+        .map_err(Box::new)
 }
 
 #[cfg(test)]
 mod tests {
-    use rustls::crypto::CryptoProvider;
+    use rustls::CipherSuite;
+    use rustls::NamedGroup;
+    use rustls::crypto::{CryptoProvider, aws_lc_rs, ring};
 
-    use super::ensure_crypto_provider;
+    use super::{Arc, build_config, tls_config};
 
-    fn fingerprint(
-        provider: &CryptoProvider,
-    ) -> (Vec<rustls::CipherSuite>, Vec<rustls::NamedGroup>) {
+    fn fingerprint(provider: &CryptoProvider) -> (Vec<CipherSuite>, Vec<NamedGroup>) {
         (
             provider
                 .cipher_suites
@@ -60,43 +91,31 @@ mod tests {
     }
 
     #[test]
-    fn client_config_builder_works_with_both_provider_features_enabled() {
-        ensure_crypto_provider();
-
-        assert!(CryptoProvider::get_default().is_some());
-
-        let config = rustls::ClientConfig::builder()
-            .with_root_certificates(rustls::RootCertStore::empty())
-            .with_no_client_auth();
+    fn builds_a_usable_config_with_both_provider_features_enabled() {
+        let config = build_config().expect("a client config");
 
         assert!(!config.crypto_provider().cipher_suites.is_empty());
     }
 
     #[test]
-    fn installs_ring_rather_than_aws_lc_rs() {
-        ensure_crypto_provider();
-
-        let installed = CryptoProvider::get_default().expect("a provider is installed");
+    fn dials_with_ring_rather_than_aws_lc_rs() {
+        let config = build_config().expect("a client config");
 
         assert_eq!(
-            fingerprint(installed),
-            fingerprint(&rustls::crypto::ring::default_provider())
+            fingerprint(config.crypto_provider()),
+            fingerprint(&ring::default_provider())
         );
         assert_ne!(
-            fingerprint(installed),
-            fingerprint(&rustls::crypto::aws_lc_rs::default_provider())
+            fingerprint(config.crypto_provider()),
+            fingerprint(&aws_lc_rs::default_provider())
         );
     }
 
     #[test]
-    fn ensure_crypto_provider_is_idempotent() {
-        ensure_crypto_provider();
-        let first = CryptoProvider::get_default().cloned();
+    fn the_trust_store_is_loaded_once_and_shared() {
+        let first = tls_config().expect("a client config");
+        let second = tls_config().expect("a client config");
 
-        ensure_crypto_provider();
-        let second = CryptoProvider::get_default().cloned();
-
-        assert!(first.is_some());
-        assert!(std::sync::Arc::ptr_eq(&first.unwrap(), &second.unwrap()));
+        assert!(Arc::ptr_eq(&first, &second));
     }
 }
diff --git a/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs b/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs
index db5698f8460..05f7d9610d5 100644
--- a/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs
+++ b/litellm-rust/crates/ai-gateway/tests/crypto_provider_wiring.rs
@@ -1,5 +1,6 @@
-//! Guards the wiring, not just the helper: the dial itself has to install the
-//! rustls provider, in a test binary where nothing else has installed one.
+//! Guards the wiring, not just the helper: a `wss://` dial through the public
+//! API has to resolve its own crypto provider, in a test binary where nothing
+//! has installed a process-wide one, and has to leave it uninstalled.
 
 use std::collections::HashMap;
 use std::time::Duration;
@@ -7,8 +8,7 @@ use std::time::Duration;
 use litellm_ai_gateway::io::responses_ws::ResponsesWebSocketConnection;
 use tokio::net::TcpListener;
 
-#[tokio::test]
-async fn dialing_wss_returns_an_error_instead_of_panicking() {
+async fn dead_tls_server() -> u16 {
     let listener = TcpListener::bind("127.0.0.1:0")
         .await
         .expect("bind a loopback port");
@@ -23,6 +23,13 @@ async fn dialing_wss_returns_an_error_instead_of_panicking() {
         }
     });
 
+    port
+}
+
+#[tokio::test]
+async fn dialing_wss_returns_an_error_instead_of_panicking() {
+    let port = dead_tls_server().await;
+
     let result = ResponsesWebSocketConnection::connect_url(
         &format!("wss://127.0.0.1:{port}/"),
         &HashMap::new(),
@@ -35,7 +42,7 @@ async fn dialing_wss_returns_an_error_instead_of_panicking() {
         "a plain TCP server cannot finish a TLS handshake"
     );
     assert!(
-        rustls::crypto::CryptoProvider::get_default().is_some(),
-        "the dial is what installs the process-wide provider"
+        rustls::crypto::CryptoProvider::get_default().is_none(),
+        "the dial settles its provider on its own connector, not process-wide"
     );
 }

From 199b44a475719500915dbd9059c0ea5dfeeae782 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 17:48:56 -0700
Subject: [PATCH 15/79] fix(async): move remote image fetches off the event
 loop for Snowflake, Bedrock invoke Claude, Mantle and Gemini

---
 .../prompt_templates/image_handling.py        | 103 ++++++++++
 litellm/llms/base_llm/chat/transformation.py  |   4 +
 .../anthropic_claude3_transformation.py       |  61 +-----
 .../bedrock/chat/mantle/transformation.py     |  15 +-
 litellm/llms/custom_httpx/llm_http_handler.py | 194 ++++++++++--------
 litellm/llms/snowflake/chat/transformation.py |  16 ++
 .../llms/vertex_ai/gemini/transformation.py   |   9 +-
 litellm/main.py                               |   8 +-
 tests/test_litellm/conftest.py                |  35 ++++
 .../litellm_core_utils/test_image_handling.py |  62 ++++++
 ...ations_anthropic_claude3_transformation.py |  49 +++++
 ...test_bedrock_chat_mantle_transformation.py |  51 +++++
 .../custom_httpx/test_llm_http_handler.py     | 106 +++++++++-
 .../test_snowflake_chat_transformation.py     |  46 +++++
 .../vertex_ai/gemini/test_transformation.py   |  42 ++++
 15 files changed, 644 insertions(+), 157 deletions(-)
 create mode 100644 tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 81132fa89a5..1890d4eb682 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -2,7 +2,11 @@
 Helper functions to handle images passed in messages
 """
 
+import asyncio
 import base64
+from collections.abc import Mapping
+from dataclasses import dataclass
+from types import MappingProxyType
 from typing import Final
 
 from httpx import Response
@@ -12,6 +16,7 @@ from litellm import verbose_logger
 from litellm.caching.caching import InMemoryCache
 from litellm.constants import MAX_IMAGE_URL_DOWNLOAD_SIZE_MB
 from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get
+from litellm.types.llms.openai import AllMessageValues
 
 MAX_IMGS_IN_MEMORY: Final = 10
 
@@ -124,3 +129,101 @@ def convert_url_to_base64(url: str) -> str:
     raise litellm.ImageFetchError(
         f"Error: Unable to fetch image from URL after 3 attempts. url={url}",
     )
+
+
+_REMOTE_URL_PREFIXES: Final = ("http://", "https://")
+
+
+@dataclass(frozen=True, slots=True)
+class _RemoteImage:
+    part: Mapping[str, object]
+    image_url: Mapping[str, object] | None
+    url: str
+
+
+@dataclass(frozen=True, slots=True)
+class _RemoteFile:
+    part: Mapping[str, object]
+    file: Mapping[str, object]
+    url: str
+
+
+def _as_mapping(value: object) -> Mapping[str, object] | None:
+    return value if isinstance(value, Mapping) else None  # pyright: ignore[reportUnknownVariableType]  # fields are parsed one by one
+
+
+def _remote_url(candidate: object) -> str | None:
+    return candidate if isinstance(candidate, str) and candidate.startswith(_REMOTE_URL_PREFIXES) else None
+
+
+def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | None:
+    fields: Final = _as_mapping(part)
+    if fields is None:
+        return None
+    if fields.get("type") == "image_url":
+        image_url: Final = fields.get("image_url")
+        image_url_fields: Final = _as_mapping(image_url)
+        url: Final = _remote_url(image_url_fields.get("url") if image_url_fields is not None else image_url)
+        return _RemoteImage(fields, image_url_fields, url) if url is not None else None
+    file: Final = _as_mapping(fields.get("file")) if fields.get("type") == "file" else None
+    file_url: Final = _remote_url(file.get("file_id")) if file is not None else None
+    return _RemoteFile(fields, file, file_url) if file is not None and file_url is not None else None
+
+
+_PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"})
+
+
+def _inferred_format(file: Mapping[str, object], url: str) -> Mapping[str, str]:
+    return _PDF_FORMAT if "format" not in file and url.lower().endswith(".pdf") else MappingProxyType({})
+
+
+def _inlined_image_url(image_url: Mapping[str, object] | None, data_url: str) -> Mapping[str, object] | str:
+    return {**image_url, "url": data_url} if image_url is not None else data_url  # mutable-ok: json-serialized part
+
+
+def _inlined_file(file: Mapping[str, object], url: str, data_url: str) -> Mapping[str, object]:
+    kept: Final = {k: v for k, v in file.items() if k != "file_id"}  # mutable-ok: json-serialized message part
+    return {**kept, **_inferred_format(file, url), "file_data": data_url}  # mutable-ok: json-serialized part
+
+
+def _inline(remote: _RemoteImage | _RemoteFile, data_url: str) -> Mapping[str, object]:
+    match remote:
+        case _RemoteImage(part, image_url, _):
+            return {**part, "image_url": _inlined_image_url(image_url, data_url)}  # mutable-ok: json-serialized part
+        case _RemoteFile(part, file, url):
+            return {**part, "file": _inlined_file(file, url, data_url)}  # mutable-ok: json-serialized message part
+
+
+def _content_parts(message: Mapping[str, object]) -> tuple[object, ...]:
+    content: Final = message.get("content")
+    return tuple(content) if isinstance(content, list) else ()  # pyright: ignore[reportUnknownVariableType, reportUnknownArgumentType]  # parts are parsed one by one
+
+
+def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) -> AllMessageValues:
+    parts: Final = _content_parts(message)
+    if not parts:
+        return message
+    inlined_parts: Final = [  # mutable-ok: content must stay a list for the transforms' isinstance checks
+        _inline(remote, data_urls[remote.url]) if (remote := _parse_remote_part(part)) is not None else part
+        for part in parts
+    ]
+    inlined_message: Final = {**message, "content": inlined_parts}  # mutable-ok: json-serialized message
+    return inlined_message  # pyright: ignore[reportReturnType]  # the same message with its remote parts inlined
+
+
+async def async_inline_remote_media(
+    messages: list[AllMessageValues],  # mutable-ok: every transform_request takes list[AllMessageValues]
+) -> list[AllMessageValues]:  # mutable-ok: every transform_request takes list[AllMessageValues]
+    remote_urls: Final = tuple(
+        dict.fromkeys(
+            remote.url
+            for message in messages
+            for part in _content_parts(message)
+            if (remote := _parse_remote_part(part)) is not None
+        )
+    )
+    if not remote_urls:
+        return messages
+    data_urls: Final = await asyncio.gather(*(async_convert_url_to_base64(url) for url in remote_urls))
+    inlined: Final = MappingProxyType(dict(zip(remote_urls, data_urls, strict=True)))
+    return [_inline_message(message, inlined) for message in messages]  # mutable-ok: transform_request takes a list
diff --git a/litellm/llms/base_llm/chat/transformation.py b/litellm/llms/base_llm/chat/transformation.py
index bbe1cc85df1..7bfc87a30d6 100644
--- a/litellm/llms/base_llm/chat/transformation.py
+++ b/litellm/llms/base_llm/chat/transformation.py
@@ -411,6 +411,10 @@ class BaseConfig(ABC):
     def has_custom_stream_wrapper(self) -> bool:
         return False
 
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return False
+
     @property
     def supports_stream_param_in_request_body(self) -> bool:
         """
diff --git a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py
index 67720451c00..38f280eef03 100644
--- a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py
+++ b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py
@@ -8,7 +8,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     convert_to_anthropic_image_obj,
 )
 from litellm.litellm_core_utils.prompt_templates.image_handling import (
-    async_convert_url_to_base64,
+    async_inline_remote_media,
     convert_url_to_base64,
 )
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
@@ -172,6 +172,10 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
 
         return _anthropic_request
 
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return True
+
     async def async_transform_request(
         self,
         model: str,
@@ -180,26 +184,14 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
         litellm_params: dict,
         headers: dict,
     ) -> dict:
-        _anthropic_request: Final = self._build_bedrock_anthropic_request_base(
+        return self.transform_request(
             model=model,
-            messages=messages,
+            messages=await async_inline_remote_media(messages),
             optional_params=optional_params,
             litellm_params=litellm_params,
             headers=headers,
         )
 
-        await self._async_convert_document_url_sources_to_base64(_anthropic_request)
-        beta_list: Final = self._compute_bedrock_invoke_beta_headers(
-            model=model,
-            messages=messages,
-            optional_params=optional_params,
-            headers=headers,
-        )
-        if beta_list:
-            _anthropic_request["anthropic_beta"] = beta_list
-
-        return _anthropic_request
-
     def _build_bedrock_anthropic_request_base(
         self,
         model: str,
@@ -321,45 +313,6 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
                     "data": image_chunk["data"],
                 }
 
-    async def _async_convert_document_url_sources_to_base64(self, anthropic_request: dict) -> None:
-        """
-        Async version of document URL conversion for async completion paths.
-        """
-        messages: Final = anthropic_request.get("messages")
-        if not isinstance(messages, list):
-            return
-
-        for message in messages:
-            if not isinstance(message, dict):
-                continue
-            content = message.get("content")
-            if not isinstance(content, list):
-                continue
-
-            for block in content:
-                if not isinstance(block, dict) or block.get("type") != "document":
-                    continue
-                source = block.get("source")
-                if not isinstance(source, dict) or source.get("type") != "url":
-                    continue
-                source_url = source.get("url")
-                if not isinstance(source_url, str):
-                    continue
-
-                inferred_format: str | None = None
-                if source_url.lower().endswith(".pdf"):
-                    inferred_format = "application/pdf"
-                base64_url = await async_convert_url_to_base64(url=source_url)
-                image_chunk = convert_to_anthropic_image_obj(
-                    openai_image_url=base64_url,
-                    format=inferred_format,
-                )
-                block["source"] = {
-                    "type": "base64",
-                    "media_type": image_chunk["media_type"],
-                    "data": image_chunk["data"],
-                }
-
     def _normalize_bedrock_tool_search_tools(self, optional_params: dict) -> dict:
         """
         Convert tool search entries to the format supported by the Bedrock Invoke API.
diff --git a/litellm/llms/bedrock/chat/mantle/transformation.py b/litellm/llms/bedrock/chat/mantle/transformation.py
index 31fc079c0c9..7e2037c33f1 100644
--- a/litellm/llms/bedrock/chat/mantle/transformation.py
+++ b/litellm/llms/bedrock/chat/mantle/transformation.py
@@ -10,6 +10,7 @@ at a different endpoint (bedrock-mantle.{region}.api.aws) with AWS SigV4 auth.
 from collections.abc import AsyncIterator, Iterator
 from typing import TYPE_CHECKING, Any, Final
 
+from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media
 from litellm.llms.bedrock.chat.invoke_transformations.anthropic_claude3_transformation import (
     AmazonAnthropicClaudeConfig,
 )
@@ -110,21 +111,13 @@ class AmazonMantleConfig(AmazonAnthropicClaudeConfig):
         litellm_params: dict,
         headers: dict,
     ) -> dict:
-        model_id: Final = model.replace("mantle/", "", 1)
-
-        request: Final = self._build_bedrock_anthropic_request_base(
-            model=model_id,
-            messages=messages,
+        return self.transform_request(
+            model=model,
+            messages=await async_inline_remote_media(messages),
             optional_params=optional_params,
             litellm_params=litellm_params,
             headers=headers,
         )
-        await self._async_convert_document_url_sources_to_base64(request)
-        return self._restore_mantle_body_fields(
-            request=request,
-            model_id=model_id,
-            optional_params=optional_params,
-        )
 
     @staticmethod
     def _restore_mantle_body_fields(request: dict, model_id: str, optional_params: dict) -> dict:
diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py
index f281c249c72..54c21f8d5b3 100644
--- a/litellm/llms/custom_httpx/llm_http_handler.py
+++ b/litellm/llms/custom_httpx/llm_http_handler.py
@@ -116,6 +116,7 @@ from litellm.types.llms.anthropic_skills import (
     Skill,
 )
 from litellm.types.llms.openai import (
+    AllMessageValues,
     CreateBatchRequest,
     CreateFileRequest,
     FileContentRequest,
@@ -488,7 +489,7 @@ class BaseLLMHTTPHandler:
     def completion(
         self,
         model: str,
-        messages: list,
+        messages: list[AllMessageValues],
         api_base: str | None,
         custom_llm_provider: str,
         model_response: ModelResponse,
@@ -507,7 +508,7 @@ class BaseLLMHTTPHandler:
         shared_session: Optional["ClientSession"] = None,
     ):
         json_mode: Final[bool] = optional_params.pop("json_mode", False)
-        extra_body: Final[dict | None] = optional_params.pop("extra_body", None)
+        extra_body: Final[Mapping[str, object] | None] = optional_params.pop("extra_body", None)
 
         provider_config = provider_config or ProviderConfigManager.get_provider_chat_config(
             model=model, provider=litellm.LlmProviders(custom_llm_provider)
@@ -522,14 +523,17 @@ class BaseLLMHTTPHandler:
         )
 
         # get config from model, custom llm provider
-        headers = provider_config.validate_environment(
-            api_key=api_key,
-            headers=headers or {},
-            model=model,
-            messages=messages,
-            optional_params=optional_params,
-            api_base=api_base,
-            litellm_params=litellm_params,
+        request_headers: Final = cast(  # cast-ok: validate_environment is declared as a bare dict
+            "dict[str, object]",
+            provider_config.validate_environment(
+                api_key=api_key,
+                headers=headers or {},
+                model=model,
+                messages=messages,
+                optional_params=optional_params,
+                api_base=api_base,
+                litellm_params=litellm_params,
+            ),
         )
 
         api_base = provider_config.get_complete_url(
@@ -541,93 +545,117 @@ class BaseLLMHTTPHandler:
             litellm_params=litellm_params,
         )
 
-        data: dict[str, object] = provider_config.transform_request(
-            model=model,
-            messages=messages,
-            optional_params=optional_params,
-            litellm_params=litellm_params,
-            headers=headers,
-        )
-
-        if extra_body is not None:
-            data = {**data, **extra_body}
-
-        headers, signed_json_body = provider_config.sign_request(
-            headers=headers,
-            optional_params={
-                **optional_params,
-                **_aws_signing_overrides(optional_params, litellm_params),
-            },
-            request_data=data,
-            api_base=api_base,
-            api_key=api_key,
-            stream=stream,
-            fake_stream=fake_stream,
-            model=model,
-        )
-
-        ## LOGGING
-        logging_obj.pre_call(
-            input=messages,
-            api_key=api_key,
-            additional_args={
-                "complete_input_dict": data,
-                "api_base": api_base,
-                "headers": headers,
-            },
-        )
-
-        # Check if stream was converted for WebSearch interception
-        # This is set by the async_pre_request_hook in WebSearchInterceptionLogger
-        if litellm_params.get("_websearch_interception_converted_stream", False):
-            logging_obj.model_call_details["websearch_interception_converted_stream"] = True
-
-        if acompletion is True:
-            if stream is True:
-                data = self._add_stream_param_to_request_body(
-                    data=data,
-                    provider_config=provider_config,
+        def sign_and_log(
+            transformed: dict[str, object],  # mutable-ok: async_completion takes dict
+        ) -> tuple[dict[str, object], dict[str, object], bytes | None]:  # mutable-ok: async_completion takes dict
+            data: Final = {**transformed, **extra_body} if extra_body is not None else transformed
+            signed: Final = cast(  # cast-ok: sign_request is declared as a bare dict
+                "tuple[dict[str, object], bytes | None]",
+                provider_config.sign_request(
+                    headers=request_headers,
+                    optional_params={
+                        **optional_params,
+                        **_aws_signing_overrides(optional_params, litellm_params),
+                    },
+                    request_data=data,
+                    api_base=api_base,
+                    api_key=api_key,
+                    stream=stream,
                     fake_stream=fake_stream,
-                )
+                    model=model,
+                ),
+            )
+            logging_obj.pre_call(
+                input=messages,
+                api_key=api_key,
+                additional_args={
+                    "complete_input_dict": data,
+                    "api_base": api_base,
+                    "headers": signed[0],
+                },
+            )
+            if litellm_params.get("_websearch_interception_converted_stream", False):
+                logging_obj.model_call_details["websearch_interception_converted_stream"] = True
+            return data, signed[0], signed[1]
+
+        def dispatch_async(
+            data: dict[str, object],  # mutable-ok: async_completion takes dict
+            signed_headers: dict[str, object],  # mutable-ok: async_completion takes dict
+            signed_json_body: bytes | None,
+        ):
+            async_client: Final = client if isinstance(client, AsyncHTTPHandler) else None
+            if stream is True:
                 return self.acompletion_stream_function(
                     model=model,
                     messages=messages,
                     api_base=api_base,
-                    headers=headers,
+                    headers=signed_headers,
                     custom_llm_provider=custom_llm_provider,
                     provider_config=provider_config,
                     timeout=timeout,
                     logging_obj=logging_obj,
-                    data=data,
+                    data=self._add_stream_param_to_request_body(
+                        data=data,
+                        provider_config=provider_config,
+                        fake_stream=fake_stream,
+                    ),
                     fake_stream=fake_stream,
-                    client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
+                    client=async_client,
                     litellm_params=litellm_params,
                     json_mode=json_mode,
                     optional_params=optional_params,
                     signed_json_body=signed_json_body,
                 )
+            return self.async_completion(
+                custom_llm_provider=custom_llm_provider,
+                provider_config=provider_config,
+                api_base=api_base,
+                headers=signed_headers,
+                data=data,
+                timeout=timeout,
+                model=model,
+                model_response=model_response,
+                logging_obj=logging_obj,
+                api_key=api_key,
+                messages=messages,
+                optional_params=optional_params,
+                litellm_params=litellm_params,
+                encoding=encoding,
+                client=async_client,
+                json_mode=json_mode,
+                signed_json_body=signed_json_body,
+                shared_session=shared_session,
+            )
 
-            else:
-                return self.async_completion(
-                    custom_llm_provider=custom_llm_provider,
-                    provider_config=provider_config,
-                    api_base=api_base,
-                    headers=headers,
-                    data=data,
-                    timeout=timeout,
-                    model=model,
-                    model_response=model_response,
-                    logging_obj=logging_obj,
-                    api_key=api_key,
-                    messages=messages,
-                    optional_params=optional_params,
-                    litellm_params=litellm_params,
-                    encoding=encoding,
-                    client=(client if client is not None and isinstance(client, AsyncHTTPHandler) else None),
-                    json_mode=json_mode,
-                    signed_json_body=signed_json_body,
-                    shared_session=shared_session,
+        if acompletion is True and provider_config.uses_async_transform_request:
+
+            async def transform_then_dispatch():
+                transformed: Final = cast(  # cast-ok: async_transform_request is declared as a bare dict
+                    "dict[str, object]",
+                    await provider_config.async_transform_request(
+                        model=model,
+                        messages=messages,
+                        optional_params=optional_params,
+                        litellm_params=litellm_params,
+                        headers=request_headers,
+                    ),
                 )
+                return await dispatch_async(*sign_and_log(transformed))
+
+            return transform_then_dispatch()
+
+        data, signed_headers, signed_json_body = sign_and_log(
+            provider_config.transform_request(
+                model=model,
+                messages=messages,
+                optional_params=optional_params,
+                litellm_params=litellm_params,
+                headers=request_headers,
+            )
+        )
+
+        if acompletion is True:
+            return dispatch_async(data, signed_headers, signed_json_body)
 
         if stream is True:
             data = self._add_stream_param_to_request_body(
@@ -641,7 +669,7 @@ class BaseLLMHTTPHandler:
                     custom_llm_provider=custom_llm_provider,
                     logging_obj=logging_obj,
                     api_base=api_base,
-                    headers=headers,
+                    headers=signed_headers,
                     data=data,
                     signed_json_body=signed_json_body,
                     messages=messages,
@@ -651,7 +679,7 @@ class BaseLLMHTTPHandler:
             completion_stream, headers = self.make_sync_call(
                 provider_config=provider_config,
                 api_base=api_base,
-                headers=headers,
+                headers=signed_headers,
                 data=data,
                 signed_json_body=signed_json_body,
                 original_data=data,
@@ -684,7 +712,7 @@ class BaseLLMHTTPHandler:
             sync_httpx_client=sync_httpx_client,
             provider_config=provider_config,
             api_base=api_base,
-            headers=headers,
+            headers=signed_headers,
             data=data,
             signed_json_body=signed_json_body,
             timeout=timeout,
diff --git a/litellm/llms/snowflake/chat/transformation.py b/litellm/llms/snowflake/chat/transformation.py
index c64fc583edc..f65b0876202 100644
--- a/litellm/llms/snowflake/chat/transformation.py
+++ b/litellm/llms/snowflake/chat/transformation.py
@@ -23,6 +23,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     create_anthropic_image_param,
     select_anthropic_content_block_type_for_file,
 )
+from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media
 from litellm.llms.anthropic.chat.handler import ModelResponseIterator as AnthropicStreamParser
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 from litellm.llms.anthropic.common_utils import normalize_cache_control_in_anthropic_payload
@@ -421,6 +422,21 @@ class SnowflakeConfig(SnowflakeBaseConfig, OpenAIGPTConfig):
             return self._transform_request_anthropic(model, messages, optional_params, stream, extra_body)
         return self._transform_request_openai(model, messages, optional_params, stream, extra_body)
 
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return True
+
+    async def async_transform_request(
+        self,
+        model: str,
+        messages: list[AllMessageValues],  # mutable-ok: BaseConfig signature
+        optional_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        litellm_params: dict[str, object],  # mutable-ok: BaseConfig signature
+        headers: dict[str, object],  # mutable-ok: BaseConfig signature
+    ) -> dict[str, object]:  # mutable-ok: BaseConfig signature
+        inlined_messages: Final = await async_inline_remote_media(messages) if _is_claude_model(model) else messages
+        return self.transform_request(model, inlined_messages, optional_params, litellm_params, headers)
+
     def _transform_request_openai(
         self,
         model: str,
diff --git a/litellm/llms/vertex_ai/gemini/transformation.py b/litellm/llms/vertex_ai/gemini/transformation.py
index e2d62be6a69..c9480f07150 100644
--- a/litellm/llms/vertex_ai/gemini/transformation.py
+++ b/litellm/llms/vertex_ai/gemini/transformation.py
@@ -27,6 +27,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     convert_to_gemini_tool_call_result,
     response_schema_prompt,
 )
+from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.vertex_ai.common_utils import pop_vertex_request_labels
 from litellm.types.files import (
@@ -1348,13 +1349,15 @@ async def async_transform_request_body(
         vertex_auth_header=vertex_auth_header,
     )
 
-    if _openai_messages_may_need_sync_gcs_metadata_fetch(messages):
+    inlined_messages: Final = await async_inline_remote_media(messages) if custom_llm_provider == "gemini" else messages
+
+    if _openai_messages_may_need_sync_gcs_metadata_fetch(inlined_messages):
         # _transform_request_body may issue a sync httpx.get (up to 5s timeout)
         # via _get_gcs_object_content_type to fetch GCS object metadata. Run the
         # whole sync transformation on a worker thread so it does not block the
         # async event loop.
         return await asyncify(_transform_request_body)(
-            messages=messages,
+            messages=inlined_messages,
             model=model,
             custom_llm_provider=custom_llm_provider,
             litellm_params=litellm_params,
@@ -1363,7 +1366,7 @@ async def async_transform_request_body(
         )
 
     return _transform_request_body(
-        messages=messages,
+        messages=inlined_messages,
         model=model,
         custom_llm_provider=custom_llm_provider,
         litellm_params=litellm_params,
diff --git a/litellm/main.py b/litellm/main.py
index 2929790f2bd..9970c203e03 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -4474,7 +4474,7 @@ def _complete_snowflake(ctx: _CompletionDispatchContext) -> _CompletionDispatchR
     acompletion: Final = ctx.acompletion
     api_base: Final = ctx.api_base
     api_key: Final = ctx.api_key
-    client = _dispatch_client_http(ctx)
+    injected_client: Final = _dispatch_client_http(ctx)
     custom_llm_provider: Final = ctx.custom_llm_provider
     headers: Final = ctx.headers
     litellm_params: Final = ctx.litellm_params
@@ -4486,11 +4486,11 @@ def _complete_snowflake(ctx: _CompletionDispatchContext) -> _CompletionDispatchR
     shared_session: Final = ctx.shared_session
     stream: Final = ctx.stream
     timeout: Final = ctx.timeout
+    client: Final = (
+        injected_client if injected_client is not None else (HTTPHandler(timeout=timeout) if stream is False else None)
+    )  # Keep this here, otherwise, the httpx.client closes and streaming is impossible
 
     try:
-        client = (
-            HTTPHandler(timeout=timeout) if stream is False else None
-        )  # Keep this here, otherwise, the httpx.client closes and streaming is impossible
         response: Final = base_llm_http_handler.completion(
             model=model,
             messages=messages,
diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py
index 62c95cb100b..aa8ba168ecc 100644
--- a/tests/test_litellm/conftest.py
+++ b/tests/test_litellm/conftest.py
@@ -7,9 +7,12 @@
 # 4. Added proper cleanup in fixtures
 # 5. Added worker-specific isolation for parallel execution
 
+import base64
 import importlib
 import os
 from pathlib import Path
+from types import SimpleNamespace
+import httpx
 import pytest
 
 import asyncio
@@ -595,3 +598,35 @@ def pytest_sessionfinish(session, exitstatus):
     _close_handler_if_needed(getattr(litellm, "aclient", None))
     _close_handler_if_needed(getattr(litellm, "client", None))
     _run_coroutine_if_needed(close_litellm_async_clients())
+
+
+ONE_PIXEL_PNG = base64.b64decode(
+    "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNkYPhfDwAChwGA60e6kgAAAABJRU5ErkJggg=="
+)
+
+
+@pytest.fixture
+def async_only_image_fetch(monkeypatch):
+    from litellm.litellm_core_utils.prompt_templates import image_handling
+
+    fetch = SimpleNamespace(
+        fetched=[],
+        base64_png=base64.b64encode(ONE_PIXEL_PNG).decode(),
+        data_url="data:image/png;base64," + base64.b64encode(ONE_PIXEL_PNG).decode(),
+    )
+
+    def forbid_sync_fetch(client, url, **kwargs):
+        raise litellm.ImageFetchError(f"sync image fetch ran on the event loop: {url}")
+
+    async def serve_png(client, url, **kwargs):
+        fetch.fetched.append(url)
+        return httpx.Response(
+            200,
+            content=ONE_PIXEL_PNG,
+            headers={"content-type": "image/png"},
+            request=httpx.Request("GET", url),
+        )
+
+    monkeypatch.setattr(image_handling, "safe_get", forbid_sync_fetch)
+    monkeypatch.setattr(image_handling, "async_safe_get", serve_png)
+    return fetch
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index 893472d63ae..5b6d403fa21 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -1,3 +1,5 @@
+import copy
+import uuid
 from unittest.mock import patch
 
 import pytest
@@ -8,6 +10,7 @@ from litellm import constants
 from litellm.litellm_core_utils.prompt_templates import image_handling
 from litellm.litellm_core_utils.prompt_templates.image_handling import (
     async_convert_url_to_base64,
+    async_inline_remote_media,
     convert_url_to_base64,
 )
 
@@ -268,3 +271,62 @@ def test_image_size_limit_disabled(monkeypatch):
 
     assert "Image URL download is disabled" in str(excinfo.value)
     assert "MAX_IMAGE_URL_DOWNLOAD_SIZE_MB=0" in str(excinfo.value)
+
+
+async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_only_image_fetch):
+    image_url = f"http://img.example/{uuid.uuid4()}.png"
+    pdf_url = f"http://docs.example/{uuid.uuid4()}.pdf"
+    messages = [
+        {"role": "system", "content": "be terse"},
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "what is this?"},
+                {"type": "image_url", "image_url": {"url": image_url, "detail": "low"}},
+                {"type": "image_url", "image_url": image_url},
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}},
+                {"type": "file", "file": {"file_id": pdf_url}},
+                {"type": "file", "file": {"file_id": image_url, "format": "image/png"}},
+            ],
+        },
+    ]
+    snapshot = copy.deepcopy(messages)
+
+    inlined = await async_inline_remote_media(messages)
+
+    data_url = async_only_image_fetch.data_url
+    assert inlined[0] == {"role": "system", "content": "be terse"}
+    assert inlined[1]["content"] == [
+        {"type": "text", "text": "what is this?"},
+        {"type": "image_url", "image_url": {"url": data_url, "detail": "low"}},
+        {"type": "image_url", "image_url": data_url},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}},
+        {"type": "file", "file": {"format": "application/pdf", "file_data": data_url}},
+        {"type": "file", "file": {"format": "image/png", "file_data": data_url}},
+    ]
+    assert sorted(async_only_image_fetch.fetched) == sorted([image_url, pdf_url])
+    assert messages == snapshot
+
+
+async def test_async_inline_remote_media_leaves_messages_without_remote_parts_alone(async_only_image_fetch):
+    messages = [
+        {"role": "user", "content": [{"type": "text", "text": "hi"}]},
+        {
+            "role": "user",
+            "content": [{"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}}],
+        },
+    ]
+
+    assert await async_inline_remote_media(messages) is messages
+    assert async_only_image_fetch.fetched == []
+
+
+async def test_async_inline_remote_media_raises_image_fetch_error_when_the_fetch_fails(monkeypatch):
+    async def serve_404(client, url, **kwargs):
+        return Response(404, request=Request("GET", url))
+
+    monkeypatch.setattr(image_handling, "async_safe_get", serve_404)
+    url = f"http://img.example/{uuid.uuid4()}.png"
+
+    with pytest.raises(litellm.ImageFetchError, match="Status code: 404"):
+        await async_inline_remote_media([{"role": "user", "content": [{"type": "image_url", "image_url": url}]}])
diff --git a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
index 0d7573a2536..e808a087e3d 100644
--- a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
+++ b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
@@ -1,15 +1,19 @@
 import asyncio
 import json
+import uuid
 from unittest.mock import patch
 
+import httpx
 import pytest
 
 # Ensure the project root is on the import path so `litellm` can be imported when
 # tests are executed from any working directory.
 
+import litellm
 from litellm.llms.bedrock.chat.invoke_transformations.anthropic_claude3_transformation import (
     AmazonAnthropicClaudeConfig,
 )
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 
 
 def test_get_supported_params_thinking():
@@ -714,3 +718,48 @@ def test_bedrock_chat_invoke_response_format_stub_still_upgrades_legacy_thinking
 
     assert result["thinking"] == {"type": "adaptive"}
     assert result["output_config"] == {"effort": "high"}
+
+
+async def test_bedrock_invoke_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch):
+    image_url = f"http://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = request.content.decode()
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "us.anthropic.claude-sonnet-5",
+                "content": [{"type": "text", "text": "Green"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="bedrock/invoke/us.anthropic.claude-sonnet-5",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        aws_access_key_id="AKIAEXAMPLE",
+        aws_secret_access_key="fake-secret",
+        aws_region_name="us-east-1",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    assert image_url not in captured["body"]
+    assert async_only_image_fetch.base64_png in captured["body"]
diff --git a/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py
new file mode 100644
index 00000000000..9b491d305c7
--- /dev/null
+++ b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py
@@ -0,0 +1,51 @@
+import uuid
+
+import httpx
+
+import litellm
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
+
+
+async def test_bedrock_mantle_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch):
+    image_url = f"http://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = request.content.decode()
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "us.anthropic.claude-sonnet-5",
+                "content": [{"type": "text", "text": "Green"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="bedrock/mantle/us.anthropic.claude-sonnet-5",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        aws_access_key_id="AKIAEXAMPLE",
+        aws_secret_access_key="fake-secret",
+        aws_region_name="us-east-1",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    assert image_url not in captured["body"]
+    assert async_only_image_fetch.base64_png in captured["body"]
diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
index 023e2d8843f..e86e671b939 100644
--- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
+++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
@@ -17,7 +17,7 @@ from litellm.llms.base_llm.audio_transcription.transformation import (
     AudioTranscriptionRequestData,
     BaseAudioTranscriptionConfig,
 )
-from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.custom_httpx.llm_http_handler import (
     BaseLLMHTTPHandler,
@@ -30,7 +30,7 @@ from litellm.llms.azure.videos.transformation import AzureVideoConfig
 from litellm.llms.openai.videos.transformation import OpenAIVideoConfig
 from litellm.types.llms.openai import ResponsesAPIResponse
 from litellm.types.router import GenericLiteLLMParams
-from litellm.types.utils import TranscriptionResponse
+from litellm.types.utils import ModelResponse, TranscriptionResponse
 
 _ACTIVE_KEY = "_code_interpreter_interception_active"
 _SANDBOX_KEY = "_code_interpreter_interception_sandbox_key"
@@ -3186,3 +3186,105 @@ async def test_async_container_list_handler_transforms_success_response():
 
     assert [container.id for container in response.data] == ["cntr_a"]
     assert response.has_more is True
+
+
+class _TransformRecordingConfig(BaseConfig):
+    def __init__(self, transform_async: bool):
+        self.transform_async = transform_async
+        self.transform_calls = []
+
+    @property
+    def uses_async_transform_request(self) -> bool:
+        return self.transform_async
+
+    def get_supported_openai_params(self, model):
+        return []
+
+    def map_openai_params(self, non_default_params, optional_params, model, drop_params):
+        return optional_params
+
+    def validate_environment(
+        self, headers, model, messages, optional_params, litellm_params, api_key=None, api_base=None
+    ):
+        return {}
+
+    def transform_request(self, model, messages, optional_params, litellm_params, headers):
+        self.transform_calls.append("sync")
+        return {"transformed_by": "sync"}
+
+    async def async_transform_request(self, model, messages, optional_params, litellm_params, headers):
+        self.transform_calls.append("async")
+        return {"transformed_by": "async"}
+
+    def transform_response(
+        self,
+        model,
+        raw_response,
+        model_response,
+        logging_obj,
+        request_data,
+        messages,
+        optional_params,
+        litellm_params,
+        encoding,
+        api_key=None,
+        json_mode=None,
+    ):
+        model_response.choices[0].message.content = raw_response.json()["transformed_by"]
+        return model_response
+
+    def get_error_class(self, error_message, status_code, headers):
+        return BaseLLMException(status_code=status_code, message=error_message, headers=headers)
+
+
+def _start_async_completion(config):
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(200, json=captured["body"])
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+    pending = BaseLLMHTTPHandler().completion(
+        model="stub-model",
+        messages=[{"role": "user", "content": "hi"}],
+        api_base="https://llm.example/v1/chat",
+        custom_llm_provider="openai",
+        model_response=ModelResponse(),
+        encoding=None,
+        logging_obj=Mock(dynamic_success_callbacks=None, model_call_details={}),
+        optional_params={},
+        timeout=10.0,
+        litellm_params={},
+        acompletion=True,
+        client=client,
+        provider_config=config,
+    )
+    return pending, captured
+
+
+async def test_completion_awaits_async_transform_request_when_config_opts_in():
+    config = _TransformRecordingConfig(transform_async=True)
+
+    pending, captured = _start_async_completion(config)
+    assert config.transform_calls == []
+
+    response = await pending
+
+    assert config.transform_calls == ["async"]
+    assert captured["body"] == {"transformed_by": "async"}
+    assert response.choices[0].message.content == "async"
+
+
+async def test_completion_keeps_sync_transform_request_before_returning_by_default():
+    config = _TransformRecordingConfig(transform_async=False)
+
+    pending, captured = _start_async_completion(config)
+    assert config.transform_calls == ["sync"]
+
+    response = await pending
+
+    assert config.transform_calls == ["sync"]
+    assert captured["body"] == {"transformed_by": "sync"}
+    assert response.choices[0].message.content == "sync"
diff --git a/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py b/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py
index 25a961c3413..5687a319f06 100644
--- a/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py
+++ b/tests/test_litellm/llms/snowflake/chat/test_snowflake_chat_transformation.py
@@ -6,6 +6,7 @@ Tests tool calling request/response transformations and chat completions
 import asyncio
 import os
 import copy
+import uuid
 import json
 from typing import Any, Dict, List
 
@@ -945,3 +946,48 @@ class TestSnowflakeChatCompletion:
 
         assert len(chunks_received) > 0
         content = "".join(c.choices[0].delta.content for c in chunks_received if c.choices[0].delta.content)
+
+
+async def test_snowflake_claude_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch):
+    image_url = f"http://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = request.content.decode()
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-sonnet-4-6",
+                "content": [{"type": "text", "text": "Green"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="snowflake/claude-sonnet-4-6",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        api_key="fake-jwt",
+        account_id="FAKE-ACCOUNT",
+        api_base=FAKE_API_BASE,
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    assert image_url not in captured["body"]
+    assert async_only_image_fetch.base64_png in captured["body"]
diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
index fad310fc5c0..ec445342523 100644
--- a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
+++ b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
@@ -1,6 +1,10 @@
 
+import uuid
+import httpx
 import pytest
 
+import litellm
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.llms.vertex_ai.gemini import transformation
 from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
     VertexGeminiConfig,
@@ -338,3 +342,41 @@ def test_map_function_enterprise_web_search_snake_case():
 
     assert len(result) == 1
     assert "enterpriseWebSearch" in result[0]
+
+
+async def test_gemini_ai_studio_async_completion_inlines_remote_images_off_the_event_loop(async_only_image_fetch):
+    image_url = f"http://img.example/{uuid.uuid4()}.png"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = request.content.decode()
+        return httpx.Response(
+            200,
+            json={
+                "candidates": [{"content": {"parts": [{"text": "Green"}], "role": "model"}, "finishReason": "STOP"}],
+                "usageMetadata": {"promptTokenCount": 1, "candidatesTokenCount": 1, "totalTokenCount": 2},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="gemini/gemini-3.8-flash",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What colour is this?"},
+                    {"type": "image_url", "image_url": {"url": image_url}},
+                ],
+            }
+        ],
+        api_key="fake-gemini-key",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "Green"
+    assert async_only_image_fetch.fetched == [image_url]
+    assert image_url not in captured["body"]
+    assert async_only_image_fetch.base64_png in captured["body"]

From e355203014326462c3e70e6b6c8d37fcc0abe656 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 18:11:41 -0700
Subject: [PATCH 16/79] fix(gemini): keep Files API refs out of the async
 remote media walker

Gemini AI Studio file URIs under generativelanguage.googleapis.com/v1beta/files/
answer 403 when fetched and must pass through as file_data.file_uri, which the
sync transform already did. Give the shared walker a skip_url_prefixes parameter,
pass that prefix from the Gemini body builder, and skip it in the AI Studio
message transform for both image_url and file parts
---
 .../prompt_templates/image_handling.py        | 12 +++--
 litellm/llms/gemini/chat/transformation.py    | 14 ++++--
 .../llms/vertex_ai/gemini/transformation.py   |  9 +++-
 .../litellm_core_utils/test_image_handling.py | 28 ++++++++++++
 .../vertex_ai/gemini/test_transformation.py   | 44 +++++++++++++++++++
 5 files changed, 99 insertions(+), 8 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 1890d4eb682..5ae8d224556 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -199,13 +199,18 @@ def _content_parts(message: Mapping[str, object]) -> tuple[object, ...]:
     return tuple(content) if isinstance(content, list) else ()  # pyright: ignore[reportUnknownVariableType, reportUnknownArgumentType]  # parts are parsed one by one
 
 
+def _inline_part(part: object, data_urls: Mapping[str, str]) -> object:
+    remote: Final = _parse_remote_part(part)
+    data_url: Final = data_urls.get(remote.url) if remote is not None else None
+    return _inline(remote, data_url) if remote is not None and data_url is not None else part
+
+
 def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) -> AllMessageValues:
     parts: Final = _content_parts(message)
     if not parts:
         return message
     inlined_parts: Final = [  # mutable-ok: content must stay a list for the transforms' isinstance checks
-        _inline(remote, data_urls[remote.url]) if (remote := _parse_remote_part(part)) is not None else part
-        for part in parts
+        _inline_part(part, data_urls) for part in parts
     ]
     inlined_message: Final = {**message, "content": inlined_parts}  # mutable-ok: json-serialized message
     return inlined_message  # pyright: ignore[reportReturnType]  # the same message with its remote parts inlined
@@ -213,13 +218,14 @@ def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) ->
 
 async def async_inline_remote_media(
     messages: list[AllMessageValues],  # mutable-ok: every transform_request takes list[AllMessageValues]
+    skip_url_prefixes: tuple[str, ...] = (),
 ) -> list[AllMessageValues]:  # mutable-ok: every transform_request takes list[AllMessageValues]
     remote_urls: Final = tuple(
         dict.fromkeys(
             remote.url
             for message in messages
             for part in _content_parts(message)
-            if (remote := _parse_remote_part(part)) is not None
+            if (remote := _parse_remote_part(part)) is not None and not remote.url.startswith(skip_url_prefixes)
         )
     )
     if not remote_urls:
diff --git a/litellm/llms/gemini/chat/transformation.py b/litellm/llms/gemini/chat/transformation.py
index 1a67b33665b..42c9ef13730 100644
--- a/litellm/llms/gemini/chat/transformation.py
+++ b/litellm/llms/gemini/chat/transformation.py
@@ -12,7 +12,7 @@ from litellm.types.llms.openai import AllMessageValues, ChatCompletionFileObject
 from litellm.types.llms.vertex_ai import ContentType, PartType
 from litellm.utils import supports_reasoning
 
-from ...vertex_ai.gemini.transformation import _gemini_convert_messages_with_history
+from ...vertex_ai.gemini.transformation import GEMINI_FILES_API_URI_PREFIX, _gemini_convert_messages_with_history
 from ...vertex_ai.gemini.vertex_and_google_ai_studio_gemini import VertexGeminiConfig
 
 
@@ -127,7 +127,11 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig):
                     if element.get("type") == "image_url":
                         img_element = cast(ChatCompletionImageObject, element)  # cast-ok: runtime type tag checked
                         _image_url, format, detail = _image_url_fields(img_element)
-                        if _image_url and "https://" in _image_url:
+                        if (
+                            _image_url
+                            and "https://" in _image_url
+                            and not _image_url.startswith(GEMINI_FILES_API_URI_PREFIX)
+                        ):
                             image_obj = convert_to_anthropic_image_obj(_image_url, format=format)
                             converted_image_url = convert_generic_image_chunk_to_openai_image_obj(image_obj)
                             if detail is not None:
@@ -147,7 +151,11 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig):
                                 llm_provider="gemini",
                             )
                         file_id = _file_field.get("file_id")
-                        if file_id and ("http://" in file_id or "https://" in file_id):
+                        if (
+                            file_id
+                            and ("http://" in file_id or "https://" in file_id)
+                            and not file_id.startswith(GEMINI_FILES_API_URI_PREFIX)
+                        ):
                             # Convert HTTP/HTTPS file URL to base64 data
                             try:
                                 base64_data = convert_url_to_base64(file_id)
diff --git a/litellm/llms/vertex_ai/gemini/transformation.py b/litellm/llms/vertex_ai/gemini/transformation.py
index c9480f07150..6d100143e52 100644
--- a/litellm/llms/vertex_ai/gemini/transformation.py
+++ b/litellm/llms/vertex_ai/gemini/transformation.py
@@ -69,6 +69,7 @@ _GCS_METADATA_VERTEX_BASE: Any | None = None
 # Shared sync client for GCS JSON API metadata reads so proxy/SSL settings
 # from litellm's HTTP stack apply (see Greptile review on PR #27278).
 _GCS_METADATA_HTTP_HANDLER: HTTPHandler | None = None
+GEMINI_FILES_API_URI_PREFIX: Final = "https://generativelanguage.googleapis.com/v1beta/files/"
 _GEMINI_MIME_TYPE_ALIASES: Final[dict[str, str]] = {
     "image/jpg": "image/jpeg",
 }
@@ -557,7 +558,7 @@ def _process_gemini_media(
             file_data = FileDataType(mime_type=mime_type, file_uri=image_url)
             part: PartType = {"file_data": file_data}
             return _apply_gemini_metadata(part, model, media_resolution_enum, video_metadata)
-        elif image_url.startswith("https://generativelanguage.googleapis.com/v1beta/files/"):
+        elif image_url.startswith(GEMINI_FILES_API_URI_PREFIX):
             # Gemini Files API URIs — the file is already uploaded to Google's
             # servers; pass the URI through as file_data without fetching it.
             # These URLs return 403 when accessed directly, so we must not try
@@ -1349,7 +1350,11 @@ async def async_transform_request_body(
         vertex_auth_header=vertex_auth_header,
     )
 
-    inlined_messages: Final = await async_inline_remote_media(messages) if custom_llm_provider == "gemini" else messages
+    inlined_messages: Final = (
+        await async_inline_remote_media(messages, skip_url_prefixes=(GEMINI_FILES_API_URI_PREFIX,))
+        if custom_llm_provider == "gemini"
+        else messages
+    )
 
     if _openai_messages_may_need_sync_gcs_metadata_fetch(inlined_messages):
         # _transform_request_body may issue a sync httpx.get (up to 5s timeout)
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index 5b6d403fa21..bcf4c7cb6ff 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -308,6 +308,34 @@ async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_o
     assert messages == snapshot
 
 
+async def test_async_inline_remote_media_leaves_skipped_url_prefixes_untouched(async_only_image_fetch):
+    skipped_prefix = "https://generativelanguage.googleapis.com/v1beta/files/"
+    skipped_file = f"{skipped_prefix}{uuid.uuid4().hex}"
+    skipped_image = f"{skipped_prefix}{uuid.uuid4().hex}"
+    fetched_image = f"https://img.example/{uuid.uuid4()}.png"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "file", "file": {"file_id": skipped_file, "format": "application/pdf"}},
+                {"type": "image_url", "image_url": {"url": skipped_image}},
+                {"type": "image_url", "image_url": fetched_image},
+            ],
+        }
+    ]
+    snapshot = copy.deepcopy(messages)
+
+    inlined = await async_inline_remote_media(messages, skip_url_prefixes=(skipped_prefix,))
+
+    assert inlined[0]["content"] == [
+        {"type": "file", "file": {"file_id": skipped_file, "format": "application/pdf"}},
+        {"type": "image_url", "image_url": {"url": skipped_image}},
+        {"type": "image_url", "image_url": async_only_image_fetch.data_url},
+    ]
+    assert async_only_image_fetch.fetched == [fetched_image]
+    assert messages == snapshot
+
+
 async def test_async_inline_remote_media_leaves_messages_without_remote_parts_alone(async_only_image_fetch):
     messages = [
         {"role": "user", "content": [{"type": "text", "text": "hi"}]},
diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
index ec445342523..d31254746d4 100644
--- a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
+++ b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
@@ -1,4 +1,5 @@
 
+import json
 import uuid
 import httpx
 import pytest
@@ -380,3 +381,46 @@ async def test_gemini_ai_studio_async_completion_inlines_remote_images_off_the_e
     assert async_only_image_fetch.fetched == [image_url]
     assert image_url not in captured["body"]
     assert async_only_image_fetch.base64_png in captured["body"]
+
+
+async def test_gemini_ai_studio_async_completion_passes_files_api_uris_through_unfetched(async_only_image_fetch):
+    files_api_pdf = f"https://generativelanguage.googleapis.com/v1beta/files/{uuid.uuid4().hex}"
+    files_api_image = f"https://generativelanguage.googleapis.com/v1beta/files/{uuid.uuid4().hex}"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "candidates": [{"content": {"parts": [{"text": "A report"}], "role": "model"}, "finishReason": "STOP"}],
+                "usageMetadata": {"promptTokenCount": 1, "candidatesTokenCount": 1, "totalTokenCount": 2},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="gemini/gemini-3.8-flash",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Summarize these"},
+                    {"type": "file", "file": {"file_id": files_api_pdf, "format": "application/pdf"}},
+                    {"type": "image_url", "image_url": {"url": files_api_image, "format": "image/png"}},
+                ],
+            }
+        ],
+        api_key="fake-gemini-key",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "A report"
+    assert async_only_image_fetch.fetched == []
+    file_parts = [part["file_data"] for part in captured["body"]["contents"][0]["parts"] if "file_data" in part]
+    assert file_parts == [
+        {"mime_type": "application/pdf", "file_uri": files_api_pdf},
+        {"mime_type": "image/png", "file_uri": files_api_image},
+    ]

From 08bb7de868f0d740c937585835333ac226faccad Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 18:28:17 -0700
Subject: [PATCH 17/79] fix(image_handling): cap in-flight remote media fetches
 per request

---
 .../prompt_templates/image_handling.py        |  9 +++++++-
 .../litellm_core_utils/test_image_handling.py | 22 +++++++++++++++++++
 2 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 5ae8d224556..4beaabc8b24 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -19,6 +19,7 @@ from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get
 from litellm.types.llms.openai import AllMessageValues
 
 MAX_IMGS_IN_MEMORY: Final = 10
+MAX_CONCURRENT_REMOTE_MEDIA_FETCHES: Final = 20
 
 in_memory_cache: Final = InMemoryCache(max_size_in_memory=MAX_IMGS_IN_MEMORY)
 
@@ -216,6 +217,11 @@ def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) ->
     return inlined_message  # pyright: ignore[reportReturnType]  # the same message with its remote parts inlined
 
 
+async def _fetch_data_url(url: str, in_flight: asyncio.Semaphore) -> str:
+    async with in_flight:
+        return await async_convert_url_to_base64(url)
+
+
 async def async_inline_remote_media(
     messages: list[AllMessageValues],  # mutable-ok: every transform_request takes list[AllMessageValues]
     skip_url_prefixes: tuple[str, ...] = (),
@@ -230,6 +236,7 @@ async def async_inline_remote_media(
     )
     if not remote_urls:
         return messages
-    data_urls: Final = await asyncio.gather(*(async_convert_url_to_base64(url) for url in remote_urls))
+    in_flight: Final = asyncio.Semaphore(MAX_CONCURRENT_REMOTE_MEDIA_FETCHES)
+    data_urls: Final = await asyncio.gather(*(_fetch_data_url(url, in_flight) for url in remote_urls))
     inlined: Final = MappingProxyType(dict(zip(remote_urls, data_urls, strict=True)))
     return [_inline_message(message, inlined) for message in messages]  # mutable-ok: transform_request takes a list
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index bcf4c7cb6ff..ae9016f2f9e 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -1,3 +1,4 @@
+import asyncio
 import copy
 import uuid
 from unittest.mock import patch
@@ -9,6 +10,7 @@ import litellm
 from litellm import constants
 from litellm.litellm_core_utils.prompt_templates import image_handling
 from litellm.litellm_core_utils.prompt_templates.image_handling import (
+    MAX_CONCURRENT_REMOTE_MEDIA_FETCHES,
     async_convert_url_to_base64,
     async_inline_remote_media,
     convert_url_to_base64,
@@ -336,6 +338,26 @@ async def test_async_inline_remote_media_leaves_skipped_url_prefixes_untouched(a
     assert messages == snapshot
 
 
+async def test_async_inline_remote_media_caps_in_flight_fetches_per_request(monkeypatch):
+    in_flight = {"now": 0, "peak": 0}
+
+    async def serve_png_slowly(client, url, **kwargs):
+        in_flight["now"] += 1
+        in_flight["peak"] = max(in_flight["peak"], in_flight["now"])
+        await asyncio.sleep(0.01)
+        in_flight["now"] -= 1
+        return Response(200, content=b"\x89PNG", headers={"content-type": "image/png"}, request=Request("GET", url))
+
+    monkeypatch.setattr(image_handling, "async_safe_get", serve_png_slowly)
+    urls = [f"https://img.example/{uuid.uuid4()}.png" for _ in range(MAX_CONCURRENT_REMOTE_MEDIA_FETCHES + 5)]
+    messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": url}} for url in urls]}]
+
+    inlined = await async_inline_remote_media(messages)
+
+    assert in_flight["peak"] == MAX_CONCURRENT_REMOTE_MEDIA_FETCHES
+    assert all(part["image_url"]["url"].startswith("data:image/png;base64,") for part in inlined[0]["content"])
+
+
 async def test_async_inline_remote_media_leaves_messages_without_remote_parts_alone(async_only_image_fetch):
     messages = [
         {"role": "user", "content": [{"type": "text", "text": "hi"}]},

From 567915aeee82953c4127456654f3a63ec9862970 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 18:51:14 -0700
Subject: [PATCH 18/79] fix(image_handling): inline url-sourced Anthropic
 document and image blocks in the async walker

---
 .../prompt_templates/image_handling.py        | 51 +++++++++++++++----
 .../litellm_core_utils/test_image_handling.py | 11 ++++
 ...ations_anthropic_claude3_transformation.py | 47 +++++++++++++++++
 ...test_bedrock_chat_mantle_transformation.py | 48 +++++++++++++++++
 4 files changed, 147 insertions(+), 10 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 4beaabc8b24..55c01b3b1cb 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -157,18 +157,41 @@ def _remote_url(candidate: object) -> str | None:
     return candidate if isinstance(candidate, str) and candidate.startswith(_REMOTE_URL_PREFIXES) else None
 
 
-def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | None:
+_ANTHROPIC_MEDIA_BLOCK_TYPES: Final = frozenset({"document", "image"})
+
+
+@dataclass(frozen=True, slots=True)
+class _RemoteSource:
+    part: Mapping[str, object]
+    url: str
+
+
+def _parse_remote_image(fields: Mapping[str, object]) -> _RemoteImage | None:
+    if fields.get("type") != "image_url":
+        return None
+    image_url: Final = fields.get("image_url")
+    image_url_fields: Final = _as_mapping(image_url)
+    url: Final = _remote_url(image_url_fields.get("url") if image_url_fields is not None else image_url)
+    return _RemoteImage(fields, image_url_fields, url) if url is not None else None
+
+
+def _parse_remote_file(fields: Mapping[str, object]) -> _RemoteFile | None:
+    file: Final = _as_mapping(fields.get("file")) if fields.get("type") == "file" else None
+    url: Final = _remote_url(file.get("file_id")) if file is not None else None
+    return _RemoteFile(fields, file, url) if file is not None and url is not None else None
+
+
+def _parse_remote_source(fields: Mapping[str, object]) -> _RemoteSource | None:
+    source: Final = _as_mapping(fields.get("source")) if fields.get("type") in _ANTHROPIC_MEDIA_BLOCK_TYPES else None
+    url: Final = _remote_url(source.get("url")) if source is not None and source.get("type") == "url" else None
+    return _RemoteSource(fields, url) if url is not None else None
+
+
+def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | _RemoteSource | None:
     fields: Final = _as_mapping(part)
     if fields is None:
         return None
-    if fields.get("type") == "image_url":
-        image_url: Final = fields.get("image_url")
-        image_url_fields: Final = _as_mapping(image_url)
-        url: Final = _remote_url(image_url_fields.get("url") if image_url_fields is not None else image_url)
-        return _RemoteImage(fields, image_url_fields, url) if url is not None else None
-    file: Final = _as_mapping(fields.get("file")) if fields.get("type") == "file" else None
-    file_url: Final = _remote_url(file.get("file_id")) if file is not None else None
-    return _RemoteFile(fields, file, file_url) if file is not None and file_url is not None else None
+    return _parse_remote_image(fields) or _parse_remote_file(fields) or _parse_remote_source(fields)
 
 
 _PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"})
@@ -187,12 +210,20 @@ def _inlined_file(file: Mapping[str, object], url: str, data_url: str) -> Mappin
     return {**kept, **_inferred_format(file, url), "file_data": data_url}  # mutable-ok: json-serialized part
 
 
-def _inline(remote: _RemoteImage | _RemoteFile, data_url: str) -> Mapping[str, object]:
+def _base64_source(url: str, data_url: str) -> Mapping[str, str]:
+    fetched_media_type, data = data_url.removeprefix("data:").split(";base64,", 1)
+    media_type: Final = "application/pdf" if url.lower().endswith(".pdf") else fetched_media_type
+    return {"type": "base64", "media_type": media_type, "data": data}  # mutable-ok: json-serialized message part
+
+
+def _inline(remote: _RemoteImage | _RemoteFile | _RemoteSource, data_url: str) -> Mapping[str, object]:
     match remote:
         case _RemoteImage(part, image_url, _):
             return {**part, "image_url": _inlined_image_url(image_url, data_url)}  # mutable-ok: json-serialized part
         case _RemoteFile(part, file, url):
             return {**part, "file": _inlined_file(file, url, data_url)}  # mutable-ok: json-serialized message part
+        case _RemoteSource(part, url):
+            return {**part, "source": _base64_source(url, data_url)}  # mutable-ok: json-serialized message part
 
 
 def _content_parts(message: Mapping[str, object]) -> tuple[object, ...]:
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index ae9016f2f9e..106bf90213b 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -289,6 +289,9 @@ async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_o
                 {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}},
                 {"type": "file", "file": {"file_id": pdf_url}},
                 {"type": "file", "file": {"file_id": image_url, "format": "image/png"}},
+                {"type": "document", "source": {"type": "url", "url": pdf_url}, "title": "the doc"},
+                {"type": "image", "source": {"type": "url", "url": image_url}},
+                {"type": "document", "source": {"type": "file", "file_id": "file_abc"}},
             ],
         },
     ]
@@ -297,6 +300,7 @@ async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_o
     inlined = await async_inline_remote_media(messages)
 
     data_url = async_only_image_fetch.data_url
+    base64_png = async_only_image_fetch.base64_png
     assert inlined[0] == {"role": "system", "content": "be terse"}
     assert inlined[1]["content"] == [
         {"type": "text", "text": "what is this?"},
@@ -305,6 +309,13 @@ async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_o
         {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo="}},
         {"type": "file", "file": {"format": "application/pdf", "file_data": data_url}},
         {"type": "file", "file": {"format": "image/png", "file_data": data_url}},
+        {
+            "type": "document",
+            "source": {"type": "base64", "media_type": "application/pdf", "data": base64_png},
+            "title": "the doc",
+        },
+        {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": base64_png}},
+        {"type": "document", "source": {"type": "file", "file_id": "file_abc"}},
     ]
     assert sorted(async_only_image_fetch.fetched) == sorted([image_url, pdf_url])
     assert messages == snapshot
diff --git a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
index e808a087e3d..cbf160c451f 100644
--- a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
+++ b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py
@@ -763,3 +763,50 @@ async def test_bedrock_invoke_claude_async_completion_inlines_remote_images_off_
     assert async_only_image_fetch.fetched == [image_url]
     assert image_url not in captured["body"]
     assert async_only_image_fetch.base64_png in captured["body"]
+
+
+async def test_bedrock_invoke_claude_async_completion_inlines_document_url_sources_off_the_event_loop(async_only_image_fetch):
+    pdf_url = f"http://docs.example/{uuid.uuid4()}.pdf"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "us.anthropic.claude-sonnet-5",
+                "content": [{"type": "text", "text": "A lease"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="bedrock/invoke/us.anthropic.claude-sonnet-5",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is this document?"},
+                    {"type": "document", "source": {"type": "url", "url": pdf_url}},
+                ],
+            }
+        ],
+        aws_access_key_id="AKIAEXAMPLE",
+        aws_secret_access_key="fake-secret",
+        aws_region_name="us-east-1",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "A lease"
+    assert async_only_image_fetch.fetched == [pdf_url]
+    assert {
+        "type": "document",
+        "source": {"type": "base64", "media_type": "application/pdf", "data": async_only_image_fetch.base64_png},
+    } in captured["body"]["messages"][0]["content"]
diff --git a/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py
index 9b491d305c7..a8448f5fa7a 100644
--- a/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py
+++ b/tests/test_litellm/llms/bedrock/chat/mantle/test_bedrock_chat_mantle_transformation.py
@@ -1,3 +1,4 @@
+import json
 import uuid
 
 import httpx
@@ -49,3 +50,50 @@ async def test_bedrock_mantle_claude_async_completion_inlines_remote_images_off_
     assert async_only_image_fetch.fetched == [image_url]
     assert image_url not in captured["body"]
     assert async_only_image_fetch.base64_png in captured["body"]
+
+
+async def test_bedrock_mantle_claude_async_completion_inlines_document_url_sources_off_the_event_loop(async_only_image_fetch):
+    pdf_url = f"http://docs.example/{uuid.uuid4()}.pdf"
+    captured = {}
+
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_1",
+                "type": "message",
+                "role": "assistant",
+                "model": "us.anthropic.claude-sonnet-5",
+                "content": [{"type": "text", "text": "A lease"}],
+                "stop_reason": "end_turn",
+                "usage": {"input_tokens": 1, "output_tokens": 1},
+            },
+        )
+
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
+
+    response = await litellm.acompletion(
+        model="bedrock/mantle/us.anthropic.claude-sonnet-5",
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is this document?"},
+                    {"type": "document", "source": {"type": "url", "url": pdf_url}},
+                ],
+            }
+        ],
+        aws_access_key_id="AKIAEXAMPLE",
+        aws_secret_access_key="fake-secret",
+        aws_region_name="us-east-1",
+        client=client,
+    )
+
+    assert response.choices[0].message.content == "A lease"
+    assert async_only_image_fetch.fetched == [pdf_url]
+    assert {
+        "type": "document",
+        "source": {"type": "base64", "media_type": "application/pdf", "data": async_only_image_fetch.base64_png},
+    } in captured["body"]["messages"][0]["content"]

From 001b531ae960aad3fa38447508d4ce0d05c118c4 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 19:11:27 -0700
Subject: [PATCH 19/79] fix(cost): bill realtime reasoning tokens nested in
 text_tokens once

OpenAI and Azure realtime usage reports output_tokens == text_tokens + audio_tokens
with reasoning_tokens counted inside text_tokens, so generic_cost_per_token billed
the reasoning share twice. When the output token details sum past completion_tokens,
the nested reasoning overlap is now subtracted from text_tokens before pricing;
shapes where text_tokens already excludes reasoning are unchanged.
---
 .../litellm_core_utils/llm_cost_calc/utils.py | 19 ++++++-
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 51 +++++++++++++++++++
 tests/test_litellm/test_cost_calculator.py    | 42 +++++++++++++++
 3 files changed, 111 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py
index 8e24302b440..1b8980abd35 100644
--- a/litellm/litellm_core_utils/llm_cost_calc/utils.py
+++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py
@@ -852,6 +852,17 @@ class CompletionTokensDetailsResult(TypedDict):
     video_tokens: int
 
 
+def _text_tokens_without_nested_reasoning(
+    completion_tokens: int,
+    text_tokens: int,
+    reasoning_tokens: int,
+    other_modality_tokens: int,
+) -> int:
+    reported_total: Final = text_tokens + reasoning_tokens + other_modality_tokens
+    nested_reasoning_tokens: Final = min(reasoning_tokens, text_tokens, max(reported_total - completion_tokens, 0))
+    return text_tokens - nested_reasoning_tokens
+
+
 def parse_completion_tokens_details(usage: Usage) -> CompletionTokensDetailsResult:
     audio_tokens: Final = (
         cast(
@@ -860,7 +871,7 @@ def parse_completion_tokens_details(usage: Usage) -> CompletionTokensDetailsResu
         )
         or 0
     )
-    text_tokens: Final = (
+    reported_text_tokens: Final = (
         cast(
             int | None,
             getattr(usage.completion_tokens_details, "text_tokens", None),
@@ -882,6 +893,12 @@ def parse_completion_tokens_details(usage: Usage) -> CompletionTokensDetailsResu
         or 0
     )
     video_tokens: Final = _coerce_token_count(getattr(usage.completion_tokens_details, "video_tokens", 0))
+    text_tokens: Final = _text_tokens_without_nested_reasoning(
+        completion_tokens=usage.completion_tokens,
+        text_tokens=reported_text_tokens,
+        reasoning_tokens=reasoning_tokens,
+        other_modality_tokens=audio_tokens + image_tokens + video_tokens,
+    )
 
     return CompletionTokensDetailsResult(
         audio_tokens=audio_tokens,
diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index 0f8084643ea..f7ca2a89048 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -4723,3 +4723,54 @@ def test_route_image_generation_cost_falls_back_to_requested_size(monkeypatch, r
     )
 
     assert cost == expected_cost
+
+
+def test_generic_cost_per_token_bills_reasoning_nested_in_text_tokens_once(_local_model_cost_map):
+    """
+    Realtime usage (OpenAI and Azure) reports output_tokens == text_tokens + audio_tokens with
+    reasoning_tokens already counted inside text_tokens, so reasoning must not be billed on top.
+    """
+
+    model = "gpt-realtime-2.1-mini"
+    usage = Usage(
+        prompt_tokens=346,
+        completion_tokens=29,
+        total_tokens=375,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=152, image_tokens=194, audio_tokens=0, cached_tokens=128
+        ),
+        completion_tokens_details=CompletionTokensDetailsWrapper(
+            text_tokens=29, audio_tokens=0, reasoning_tokens=19
+        ),
+    )
+
+    prompt_cost, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="openai")
+    breakdown = get_token_type_cost_breakdown(model=model, custom_llm_provider="openai", usage=usage)
+
+    info = litellm.get_model_info(model=model, custom_llm_provider="openai")
+    assert completion_cost == pytest.approx(29 * info["output_cost_per_token"])
+    assert completion_cost - breakdown.reasoning_cost == pytest.approx(10 * info["output_cost_per_token"])
+    assert prompt_cost == pytest.approx(
+        24 * info["input_cost_per_token"]
+        + 128 * info["cache_read_input_token_cost"]
+        + 194 * info["input_cost_per_image_token"]
+    )
+
+
+def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tokens(_local_model_cost_map):
+    """Providers whose text_tokens exclude reasoning (text + reasoning == completion) stay billed in full."""
+
+    model = "gpt-realtime-2.1-mini"
+    usage = Usage(
+        prompt_tokens=100,
+        completion_tokens=44,
+        total_tokens=144,
+        completion_tokens_details=CompletionTokensDetailsWrapper(
+            text_tokens=25, audio_tokens=0, reasoning_tokens=19
+        ),
+    )
+
+    _, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="openai")
+
+    info = litellm.get_model_info(model=model, custom_llm_provider="openai")
+    assert completion_cost == pytest.approx(44 * info["output_cost_per_token"])
diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py
index 2046695f151..203f0a9e840 100644
--- a/tests/test_litellm/test_cost_calculator.py
+++ b/tests/test_litellm/test_cost_calculator.py
@@ -4492,3 +4492,45 @@ def test_batch_cost_calculator_gpt_6_astra_bills_half_the_standard_rate(_local_m
 
     assert prompt_cost == pytest.approx(1000 * 5e-6)
     assert completion_cost == pytest.approx(500 * 2.5e-5)
+
+
+def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_once(_local_model_cost_map):
+    """Realtime response.done nests reasoning_tokens inside text_tokens, so they are billed once."""
+    results: OpenAIRealtimeStreamList = [
+        {"type": "session.created", "session": {"model": "gpt-realtime-2.1-mini"}},
+        {
+            "type": "response.done",
+            "response": {
+                "usage": {
+                    "total_tokens": 260,
+                    "input_tokens": 237,
+                    "output_tokens": 23,
+                    "input_token_details": {
+                        "text_tokens": 43,
+                        "audio_tokens": 0,
+                        "image_tokens": 194,
+                        "cached_tokens": 0,
+                        "cached_tokens_details": {"text_tokens": 0, "audio_tokens": 0, "image_tokens": 0},
+                    },
+                    "output_token_details": {"text_tokens": 23, "audio_tokens": 0, "reasoning_tokens": 18},
+                }
+            },
+        },
+    ]
+    combined_usage_object = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(
+        results=results,
+    )
+
+    total_cost = handle_realtime_stream_cost_calculation(
+        results=results,
+        combined_usage_object=combined_usage_object,
+        custom_llm_provider="azure",
+        litellm_model_name="azure/gpt-realtime-2.1-mini",
+    )
+
+    info = litellm.get_model_info(model="azure/gpt-realtime-2.1-mini", custom_llm_provider="azure")
+    expected = (
+        43 * info["input_cost_per_token"] + 194 * info["input_cost_per_image_token"] + 23 * info["output_cost_per_token"]
+    )
+    assert total_cost == pytest.approx(expected)
+    assert total_cost == pytest.approx(0.0002362)

From 5a7e938f371e787d1e60af1eddea70bc10978813 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 19:40:46 -0700
Subject: [PATCH 20/79] test(cost): type the cost map fixture parameter on the
 new tests

---
 .../llm_cost_calc/test_llm_cost_calc_utils.py                 | 4 ++--
 tests/test_litellm/test_cost_calculator.py                    | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index f7ca2a89048..19464dfd2a0 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -4725,7 +4725,7 @@ def test_route_image_generation_cost_falls_back_to_requested_size(monkeypatch, r
     assert cost == expected_cost
 
 
-def test_generic_cost_per_token_bills_reasoning_nested_in_text_tokens_once(_local_model_cost_map):
+def test_generic_cost_per_token_bills_reasoning_nested_in_text_tokens_once(_local_model_cost_map: None) -> None:
     """
     Realtime usage (OpenAI and Azure) reports output_tokens == text_tokens + audio_tokens with
     reasoning_tokens already counted inside text_tokens, so reasoning must not be billed on top.
@@ -4757,7 +4757,7 @@ def test_generic_cost_per_token_bills_reasoning_nested_in_text_tokens_once(_loca
     )
 
 
-def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tokens(_local_model_cost_map):
+def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tokens(_local_model_cost_map: None) -> None:
     """Providers whose text_tokens exclude reasoning (text + reasoning == completion) stay billed in full."""
 
     model = "gpt-realtime-2.1-mini"
diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py
index 203f0a9e840..3eb051982e4 100644
--- a/tests/test_litellm/test_cost_calculator.py
+++ b/tests/test_litellm/test_cost_calculator.py
@@ -4494,7 +4494,7 @@ def test_batch_cost_calculator_gpt_6_astra_bills_half_the_standard_rate(_local_m
     assert completion_cost == pytest.approx(500 * 2.5e-5)
 
 
-def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_once(_local_model_cost_map):
+def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_once(_local_model_cost_map: None) -> None:
     """Realtime response.done nests reasoning_tokens inside text_tokens, so they are billed once."""
     results: OpenAIRealtimeStreamList = [
         {"type": "session.created", "session": {"model": "gpt-realtime-2.1-mini"}},

From 92e449d4d0c486f24c7a09bfce988a6be5b1dd70 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 20:35:21 -0700
Subject: [PATCH 21/79] test(cost): cover reasoning nested in text_tokens
 beside audio output

---
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index 19464dfd2a0..29874c5cca7 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -4774,3 +4774,22 @@ def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tok
 
     info = litellm.get_model_info(model=model, custom_llm_provider="openai")
     assert completion_cost == pytest.approx(44 * info["output_cost_per_token"])
+
+
+def test_generic_cost_per_token_bills_nested_reasoning_once_beside_audio_output(_local_model_cost_map: None) -> None:
+    """Audio-output realtime usage nests reasoning inside text_tokens next to audio_tokens; text is billed net of it."""
+
+    model = "gpt-realtime-2.1-mini"
+    usage = Usage(
+        prompt_tokens=120,
+        completion_tokens=100,
+        total_tokens=220,
+        completion_tokens_details=CompletionTokensDetailsWrapper(text_tokens=30, audio_tokens=70, reasoning_tokens=20),
+    )
+
+    _, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="openai")
+    breakdown = get_token_type_cost_breakdown(model=model, custom_llm_provider="openai", usage=usage)
+
+    info = litellm.get_model_info(model=model, custom_llm_provider="openai")
+    assert breakdown.reasoning_cost == pytest.approx(20 * info["output_cost_per_token"])
+    assert completion_cost == pytest.approx(30 * info["output_cost_per_token"] + 70 * info["output_cost_per_audio_token"])

From 93fa9892389c6206f6b7fe99b3032096abd10d88 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 20:42:41 -0700
Subject: [PATCH 22/79] fix(router): fall back to the deployment voice without
 a conditional dict spread

---
 litellm/router.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/litellm/router.py b/litellm/router.py
index 3b7f2d79823..c211e385d95 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -4518,7 +4518,7 @@ class Router:
                 **{
                     **data,
                     "input": input,
-                    **({"voice": voice} if voice is not None else {}),
+                    "voice": data.get("voice") if voice is None else voice,
                     "client": model_client,
                     **kwargs,
                 }

From f09992b73082e37d7337402954387def48fc7420 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 20:54:40 -0700
Subject: [PATCH 23/79] fix(image_handling): keep DNS, signing, and Vertex
 Gemini fetches off the event loop

The SSRF check in async_safe_get resolved DNS on the event loop and a blocked
address was retried three times; validate_url now runs in a thread and an
SSRFError fails the fetch on the first attempt in both fetchers. The shared
HTTP handler signed the request and ran pre_call logging on the loop after an
async transform; both now run in a thread. Vertex AI Gemini still fetched
http:// images and https images without an inferrable mime type with the sync
converter inside its async body builder; the walker takes a should_inline
predicate and Vertex AI inlines exactly those URLs, leaving https images with a
known mime type and Files API refs to Google. When one download fails the
other in-flight downloads for that request are now cancelled instead of
finishing in the background
---
 .../prompt_templates/image_handling.py        |  73 ++++++++--
 litellm/litellm_core_utils/url_utils.py       |   3 +-
 litellm/llms/custom_httpx/llm_http_handler.py |   2 +-
 .../llms/vertex_ai/gemini/transformation.py   |  26 +++-
 .../litellm_core_utils/test_image_handling.py | 135 ++++++++++++++++--
 .../litellm_core_utils/test_url_utils.py      |  35 +++++
 .../custom_httpx/test_llm_http_handler.py     |  28 +++-
 .../vertex_ai/gemini/test_transformation.py   |  51 +++++++
 8 files changed, 318 insertions(+), 35 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 55c01b3b1cb..dcad7776b58 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -4,7 +4,7 @@ Helper functions to handle images passed in messages
 
 import asyncio
 import base64
-from collections.abc import Mapping
+from collections.abc import Callable, Mapping
 from dataclasses import dataclass
 from types import MappingProxyType
 from typing import Final
@@ -15,7 +15,7 @@ import litellm
 from litellm import verbose_logger
 from litellm.caching.caching import InMemoryCache
 from litellm.constants import MAX_IMAGE_URL_DOWNLOAD_SIZE_MB
-from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get
+from litellm.litellm_core_utils.url_utils import SSRFError, async_safe_get, safe_get
 from litellm.types.llms.openai import AllMessageValues
 
 MAX_IMGS_IN_MEMORY: Final = 10
@@ -99,6 +99,8 @@ async def async_convert_url_to_base64(url: str) -> str:
             return _process_image_response(response, url)
         except litellm.ImageFetchError:
             raise
+        except SSRFError as e:
+            raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL. {e} url={url}") from e
         except Exception:
             pass
     raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL after 3 attempts. url={url}")
@@ -125,6 +127,8 @@ def convert_url_to_base64(url: str) -> str:
             return _process_image_response(response, url)
         except litellm.ImageFetchError:
             raise
+        except SSRFError as e:
+            raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL. {e} url={url}") from e
         except Exception as e:
             verbose_logger.exception(e)
     raise litellm.ImageFetchError(
@@ -163,9 +167,23 @@ _ANTHROPIC_MEDIA_BLOCK_TYPES: Final = frozenset({"document", "image"})
 @dataclass(frozen=True, slots=True)
 class _RemoteSource:
     part: Mapping[str, object]
+    source: Mapping[str, object]
     url: str
 
 
+@dataclass(frozen=True, slots=True)
+class RemoteMedia:
+    url: str
+    fields: Mapping[str, object]
+
+
+_NO_FIELDS: Final[Mapping[str, object]] = MappingProxyType({})
+
+
+def inline_every_remote_url(_media: RemoteMedia) -> bool:
+    return True
+
+
 def _parse_remote_image(fields: Mapping[str, object]) -> _RemoteImage | None:
     if fields.get("type") != "image_url":
         return None
@@ -184,7 +202,7 @@ def _parse_remote_file(fields: Mapping[str, object]) -> _RemoteFile | None:
 def _parse_remote_source(fields: Mapping[str, object]) -> _RemoteSource | None:
     source: Final = _as_mapping(fields.get("source")) if fields.get("type") in _ANTHROPIC_MEDIA_BLOCK_TYPES else None
     url: Final = _remote_url(source.get("url")) if source is not None and source.get("type") == "url" else None
-    return _RemoteSource(fields, url) if url is not None else None
+    return _RemoteSource(fields, source, url) if source is not None and url is not None else None
 
 
 def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | _RemoteSource | None:
@@ -194,6 +212,16 @@ def _parse_remote_part(part: object) -> _RemoteImage | _RemoteFile | _RemoteSour
     return _parse_remote_image(fields) or _parse_remote_file(fields) or _parse_remote_source(fields)
 
 
+def _remote_media(remote: _RemoteImage | _RemoteFile | _RemoteSource) -> RemoteMedia:
+    match remote:
+        case _RemoteImage(_, image_url, url):
+            return RemoteMedia(url, image_url if image_url is not None else _NO_FIELDS)
+        case _RemoteFile(_, file, url):
+            return RemoteMedia(url, file)
+        case _RemoteSource(_, source, url):
+            return RemoteMedia(url, source)
+
+
 _PDF_FORMAT: Final = MappingProxyType({"format": "application/pdf"})
 
 
@@ -222,7 +250,7 @@ def _inline(remote: _RemoteImage | _RemoteFile | _RemoteSource, data_url: str) -
             return {**part, "image_url": _inlined_image_url(image_url, data_url)}  # mutable-ok: json-serialized part
         case _RemoteFile(part, file, url):
             return {**part, "file": _inlined_file(file, url, data_url)}  # mutable-ok: json-serialized message part
-        case _RemoteSource(part, url):
+        case _RemoteSource(part, _, url):
             return {**part, "source": _base64_source(url, data_url)}  # mutable-ok: json-serialized message part
 
 
@@ -231,18 +259,22 @@ def _content_parts(message: Mapping[str, object]) -> tuple[object, ...]:
     return tuple(content) if isinstance(content, list) else ()  # pyright: ignore[reportUnknownVariableType, reportUnknownArgumentType]  # parts are parsed one by one
 
 
-def _inline_part(part: object, data_urls: Mapping[str, str]) -> object:
+def _inline_part(part: object, data_urls: Mapping[str, str], should_inline: Callable[[RemoteMedia], bool]) -> object:
     remote: Final = _parse_remote_part(part)
-    data_url: Final = data_urls.get(remote.url) if remote is not None else None
-    return _inline(remote, data_url) if remote is not None and data_url is not None else part
+    if remote is None or not should_inline(_remote_media(remote)):
+        return part
+    data_url: Final = data_urls.get(remote.url)
+    return _inline(remote, data_url) if data_url is not None else part
 
 
-def _inline_message(message: AllMessageValues, data_urls: Mapping[str, str]) -> AllMessageValues:
+def _inline_message(
+    message: AllMessageValues, data_urls: Mapping[str, str], should_inline: Callable[[RemoteMedia], bool]
+) -> AllMessageValues:
     parts: Final = _content_parts(message)
     if not parts:
         return message
     inlined_parts: Final = [  # mutable-ok: content must stay a list for the transforms' isinstance checks
-        _inline_part(part, data_urls) for part in parts
+        _inline_part(part, data_urls, should_inline) for part in parts
     ]
     inlined_message: Final = {**message, "content": inlined_parts}  # mutable-ok: json-serialized message
     return inlined_message  # pyright: ignore[reportReturnType]  # the same message with its remote parts inlined
@@ -253,21 +285,34 @@ async def _fetch_data_url(url: str, in_flight: asyncio.Semaphore) -> str:
         return await async_convert_url_to_base64(url)
 
 
+async def _fetch_data_urls(remote_urls: tuple[str, ...]) -> tuple[str, ...]:
+    in_flight: Final = asyncio.Semaphore(MAX_CONCURRENT_REMOTE_MEDIA_FETCHES)
+    fetches: Final = tuple(asyncio.create_task(_fetch_data_url(url, in_flight)) for url in remote_urls)
+    try:
+        return tuple(await asyncio.gather(*fetches))
+    except BaseException:
+        for fetch in fetches:
+            fetch.cancel()
+        await asyncio.gather(*fetches, return_exceptions=True)
+        raise
+
+
 async def async_inline_remote_media(
     messages: list[AllMessageValues],  # mutable-ok: every transform_request takes list[AllMessageValues]
-    skip_url_prefixes: tuple[str, ...] = (),
+    should_inline: Callable[[RemoteMedia], bool] = inline_every_remote_url,
 ) -> list[AllMessageValues]:  # mutable-ok: every transform_request takes list[AllMessageValues]
     remote_urls: Final = tuple(
         dict.fromkeys(
             remote.url
             for message in messages
             for part in _content_parts(message)
-            if (remote := _parse_remote_part(part)) is not None and not remote.url.startswith(skip_url_prefixes)
+            if (remote := _parse_remote_part(part)) is not None and should_inline(_remote_media(remote))
         )
     )
     if not remote_urls:
         return messages
-    in_flight: Final = asyncio.Semaphore(MAX_CONCURRENT_REMOTE_MEDIA_FETCHES)
-    data_urls: Final = await asyncio.gather(*(_fetch_data_url(url, in_flight) for url in remote_urls))
+    data_urls: Final = await _fetch_data_urls(remote_urls)
     inlined: Final = MappingProxyType(dict(zip(remote_urls, data_urls, strict=True)))
-    return [_inline_message(message, inlined) for message in messages]  # mutable-ok: transform_request takes a list
+    return [  # mutable-ok: transform_request takes a list
+        _inline_message(message, inlined, should_inline) for message in messages
+    ]
diff --git a/litellm/litellm_core_utils/url_utils.py b/litellm/litellm_core_utils/url_utils.py
index 1e43117933d..fa070a648f5 100644
--- a/litellm/litellm_core_utils/url_utils.py
+++ b/litellm/litellm_core_utils/url_utils.py
@@ -19,6 +19,7 @@ Admins can opt out via two ``litellm`` globals (wired from proxy config):
   check but still resolve DNS and still rewrite HTTP to the resolved IP.
 """
 
+import asyncio
 import socket
 from ipaddress import ip_address, ip_network
 from typing import Any, Final, Protocol
@@ -471,7 +472,7 @@ async def async_safe_get(client: Any, url: str, **kwargs: Any) -> httpx.Response
     kwargs.pop("follow_redirects", None)
     headers_view: Final[_CallerHeadersView] = {"headers": kwargs.pop("headers", {})}
     for _ in range(_MAX_REDIRECTS):
-        validated_url, original_host = validate_url(url)
+        validated_url, original_host = await asyncio.to_thread(validate_url, url)
         response = await fetcher.get(
             validated_url,
             headers={**headers_view["headers"], "Host": original_host},
diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py
index 54c21f8d5b3..d57aee025d5 100644
--- a/litellm/llms/custom_httpx/llm_http_handler.py
+++ b/litellm/llms/custom_httpx/llm_http_handler.py
@@ -640,7 +640,7 @@ class BaseLLMHTTPHandler:
                         headers=request_headers,
                     ),
                 )
-                return await dispatch_async(*sign_and_log(transformed))
+                return await dispatch_async(*await asyncio.to_thread(sign_and_log, transformed))
 
             return transform_then_dispatch()
 
diff --git a/litellm/llms/vertex_ai/gemini/transformation.py b/litellm/llms/vertex_ai/gemini/transformation.py
index 6d100143e52..13e2238fdf6 100644
--- a/litellm/llms/vertex_ai/gemini/transformation.py
+++ b/litellm/llms/vertex_ai/gemini/transformation.py
@@ -7,6 +7,7 @@ Why separate file? Make it easy to see how transformation works
 import json
 import os
 import re
+from collections.abc import Mapping
 from typing import TYPE_CHECKING, Any, Final, Literal, cast
 from urllib.parse import quote
 
@@ -27,7 +28,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
     convert_to_gemini_tool_call_result,
     response_schema_prompt,
 )
-from litellm.litellm_core_utils.prompt_templates.image_handling import async_inline_remote_media
+from litellm.litellm_core_utils.prompt_templates.image_handling import RemoteMedia, async_inline_remote_media
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.vertex_ai.common_utils import pop_vertex_request_labels
 from litellm.types.files import (
@@ -1309,6 +1310,23 @@ def sync_transform_request_body(
     )
 
 
+def _explicit_mime_type(fields: Mapping[str, object]) -> str | None:
+    hint: Final = fields.get("format") or fields.get("mime_type") or fields.get("content_type")
+    return hint if isinstance(hint, str) else None
+
+
+def _ai_studio_inlines(media: RemoteMedia) -> bool:
+    return not media.url.startswith(GEMINI_FILES_API_URI_PREFIX)
+
+
+def _vertex_inlines(media: RemoteMedia) -> bool:
+    if media.url.startswith(GEMINI_FILES_API_URI_PREFIX):
+        return False
+    return media.url.startswith("http://") or (
+        _explicit_mime_type(media.fields) is None and _get_image_mime_type_from_url(media.url) is None
+    )
+
+
 async def async_transform_request_body(
     gemini_api_key: str | None,
     messages: list[AllMessageValues],
@@ -1350,10 +1368,8 @@ async def async_transform_request_body(
         vertex_auth_header=vertex_auth_header,
     )
 
-    inlined_messages: Final = (
-        await async_inline_remote_media(messages, skip_url_prefixes=(GEMINI_FILES_API_URI_PREFIX,))
-        if custom_llm_provider == "gemini"
-        else messages
+    inlined_messages: Final = await async_inline_remote_media(
+        messages, should_inline=_ai_studio_inlines if custom_llm_provider == "gemini" else _vertex_inlines
     )
 
     if _openai_messages_may_need_sync_gcs_metadata_fetch(inlined_messages):
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index 106bf90213b..e7b28d0d3a2 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -1,5 +1,6 @@
 import asyncio
 import copy
+import time
 import uuid
 from unittest.mock import patch
 
@@ -11,10 +12,12 @@ from litellm import constants
 from litellm.litellm_core_utils.prompt_templates import image_handling
 from litellm.litellm_core_utils.prompt_templates.image_handling import (
     MAX_CONCURRENT_REMOTE_MEDIA_FETCHES,
+    RemoteMedia,
     async_convert_url_to_base64,
     async_inline_remote_media,
     convert_url_to_base64,
 )
+from litellm.litellm_core_utils.url_utils import SSRFError
 
 
 @pytest.fixture(autouse=True)
@@ -321,34 +324,142 @@ async def test_async_inline_remote_media_inlines_every_remote_part_shape(async_o
     assert messages == snapshot
 
 
-async def test_async_inline_remote_media_leaves_skipped_url_prefixes_untouched(async_only_image_fetch):
-    skipped_prefix = "https://generativelanguage.googleapis.com/v1beta/files/"
-    skipped_file = f"{skipped_prefix}{uuid.uuid4().hex}"
-    skipped_image = f"{skipped_prefix}{uuid.uuid4().hex}"
-    fetched_image = f"https://img.example/{uuid.uuid4()}.png"
+async def test_async_inline_remote_media_inlines_only_the_parts_the_predicate_accepts(async_only_image_fetch):
+    files_api_prefix = "https://generativelanguage.googleapis.com/v1beta/files/"
+    files_api_pdf = f"{files_api_prefix}{uuid.uuid4().hex}"
+    hinted_image = f"https://img.example/{uuid.uuid4()}.png"
+    plain_image = f"https://img.example/{uuid.uuid4()}.png"
+    hinted_document = f"https://docs.example/{uuid.uuid4()}.pdf"
+    seen = []
+
+    def inline_unhinted_outside_files_api(media: RemoteMedia) -> bool:
+        seen.append(media)
+        return not media.url.startswith(files_api_prefix) and "format" not in media.fields
+
     messages = [
         {
             "role": "user",
             "content": [
-                {"type": "file", "file": {"file_id": skipped_file, "format": "application/pdf"}},
-                {"type": "image_url", "image_url": {"url": skipped_image}},
-                {"type": "image_url", "image_url": fetched_image},
+                {"type": "file", "file": {"file_id": files_api_pdf}},
+                {"type": "image_url", "image_url": {"url": hinted_image, "format": "image/png"}},
+                {"type": "image_url", "image_url": {"url": plain_image}},
+                {"type": "image_url", "image_url": plain_image},
+                {"type": "document", "source": {"type": "url", "url": hinted_document, "format": "application/pdf"}},
             ],
         }
     ]
     snapshot = copy.deepcopy(messages)
 
-    inlined = await async_inline_remote_media(messages, skip_url_prefixes=(skipped_prefix,))
+    inlined = await async_inline_remote_media(messages, should_inline=inline_unhinted_outside_files_api)
 
     assert inlined[0]["content"] == [
-        {"type": "file", "file": {"file_id": skipped_file, "format": "application/pdf"}},
-        {"type": "image_url", "image_url": {"url": skipped_image}},
+        {"type": "file", "file": {"file_id": files_api_pdf}},
+        {"type": "image_url", "image_url": {"url": hinted_image, "format": "image/png"}},
+        {"type": "image_url", "image_url": {"url": async_only_image_fetch.data_url}},
         {"type": "image_url", "image_url": async_only_image_fetch.data_url},
+        {"type": "document", "source": {"type": "url", "url": hinted_document, "format": "application/pdf"}},
+    ]
+    assert async_only_image_fetch.fetched == [plain_image]
+    assert [(media.url, dict(media.fields)) for media in seen[:5]] == [
+        (files_api_pdf, {"file_id": files_api_pdf}),
+        (hinted_image, {"url": hinted_image, "format": "image/png"}),
+        (plain_image, {"url": plain_image}),
+        (plain_image, {}),
+        (hinted_document, {"type": "url", "url": hinted_document, "format": "application/pdf"}),
     ]
-    assert async_only_image_fetch.fetched == [fetched_image]
     assert messages == snapshot
 
 
+async def test_async_inline_remote_media_inlines_a_shared_url_only_where_the_predicate_accepts_it(
+    async_only_image_fetch,
+):
+    shared = f"https://img.example/{uuid.uuid4()}.png"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": shared, "format": "image/png"}},
+                {"type": "image_url", "image_url": {"url": shared}},
+            ],
+        }
+    ]
+
+    inlined = await async_inline_remote_media(messages, should_inline=lambda media: "format" not in media.fields)
+
+    assert inlined[0]["content"] == [
+        {"type": "image_url", "image_url": {"url": shared, "format": "image/png"}},
+        {"type": "image_url", "image_url": {"url": async_only_image_fetch.data_url}},
+    ]
+    assert async_only_image_fetch.fetched == [shared]
+
+
+async def test_async_inline_remote_media_cancels_the_other_fetches_when_one_fails(monkeypatch):
+    missing = f"http://img.example/{uuid.uuid4()}-missing.png"
+    slow = f"http://img.example/{uuid.uuid4()}-slow.png"
+    slow_fetch_outcomes = []
+
+    async def serve(client, url, **kwargs):
+        if url == missing:
+            return Response(404, request=Request("GET", url))
+        try:
+            await asyncio.sleep(5)
+        except asyncio.CancelledError:
+            slow_fetch_outcomes.append("cancelled")
+            raise
+        slow_fetch_outcomes.append("finished")
+        return Response(200, content=b"\x89PNG", headers={"content-type": "image/png"}, request=Request("GET", url))
+
+    monkeypatch.setattr(image_handling, "async_safe_get", serve)
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": missing}},
+                {"type": "image_url", "image_url": {"url": slow}},
+            ],
+        }
+    ]
+    started = time.perf_counter()
+
+    with pytest.raises(litellm.ImageFetchError, match="Status code: 404"):
+        await async_inline_remote_media(messages)
+
+    assert slow_fetch_outcomes == ["cancelled"]
+    assert time.perf_counter() - started < 1
+
+
+async def test_async_convert_url_to_base64_reports_a_blocked_url_without_retrying(monkeypatch):
+    attempts = []
+
+    async def block(client, url, **kwargs):
+        attempts.append(url)
+        raise SSRFError("URL targets a blocked address (10.0.0.8)")
+
+    monkeypatch.setattr(image_handling, "async_safe_get", block)
+    url = f"http://internal.example/{uuid.uuid4()}.png"
+
+    with pytest.raises(litellm.ImageFetchError, match=r"blocked address \(10\.0\.0\.8\)"):
+        await async_convert_url_to_base64(url)
+
+    assert attempts == [url]
+
+
+def test_convert_url_to_base64_reports_a_blocked_url_without_retrying(monkeypatch):
+    attempts = []
+
+    def block(client, url, **kwargs):
+        attempts.append(url)
+        raise SSRFError("URL targets a blocked address (10.0.0.8)")
+
+    monkeypatch.setattr(image_handling, "safe_get", block)
+    url = f"http://internal.example/{uuid.uuid4()}.png"
+
+    with pytest.raises(litellm.ImageFetchError, match=r"blocked address \(10\.0\.0\.8\)"):
+        convert_url_to_base64(url)
+
+    assert attempts == [url]
+
+
 async def test_async_inline_remote_media_caps_in_flight_fetches_per_request(monkeypatch):
     in_flight = {"now": 0, "peak": 0}
 
diff --git a/tests/test_litellm/litellm_core_utils/test_url_utils.py b/tests/test_litellm/litellm_core_utils/test_url_utils.py
index aaaa43a0dc4..fccdc1a2a0a 100644
--- a/tests/test_litellm/litellm_core_utils/test_url_utils.py
+++ b/tests/test_litellm/litellm_core_utils/test_url_utils.py
@@ -1,5 +1,9 @@
+import asyncio
 import socket
+import threading
+import time
 
+import httpx
 import pytest
 
 import litellm
@@ -535,3 +539,34 @@ def test_assert_same_origin_error_message_does_not_leak_hostnames():
     detail = str(exc.value)
     assert "attacker.example.com" not in detail
     assert "api.internal-corp.example" not in detail
+
+
+async def test_async_safe_get_resolves_dns_off_the_event_loop(monkeypatch):
+    loop_thread = threading.current_thread()
+    resolver_threads = []
+
+    def slow_getaddrinfo(host, port, *args, **kwargs):
+        resolver_threads.append(threading.current_thread())
+        time.sleep(0.4)
+        return [(socket.AF_INET, socket.SOCK_STREAM, 6, "", ("93.184.216.34", port or 443))]
+
+    monkeypatch.setattr(url_utils.socket, "getaddrinfo", slow_getaddrinfo)
+
+    class FakeClient:
+        async def get(self, url, **kwargs):
+            return httpx.Response(200, request=httpx.Request("GET", url))
+
+    ticks = [time.perf_counter()]
+
+    async def heartbeat():
+        while True:
+            await asyncio.sleep(0.01)
+            ticks.append(time.perf_counter())
+
+    beating = asyncio.create_task(heartbeat())
+    response = await url_utils.async_safe_get(FakeClient(), "https://img.example/a.png")
+    beating.cancel()
+
+    assert response.status_code == 200
+    assert resolver_threads and all(thread is not loop_thread for thread in resolver_threads)
+    assert max(b - a for a, b in zip(ticks, ticks[1:])) < 0.2
diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
index e86e671b939..f1614654ffb 100644
--- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
+++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
@@ -1,6 +1,7 @@
 import asyncio
 import json
 import logging
+import threading
 import time
 from unittest.mock import AsyncMock, Mock, patch
 
@@ -3192,6 +3193,7 @@ class _TransformRecordingConfig(BaseConfig):
     def __init__(self, transform_async: bool):
         self.transform_async = transform_async
         self.transform_calls = []
+        self.sign_threads = []
 
     @property
     def uses_async_transform_request(self) -> bool:
@@ -3216,6 +3218,12 @@ class _TransformRecordingConfig(BaseConfig):
         self.transform_calls.append("async")
         return {"transformed_by": "async"}
 
+    def sign_request(
+        self, headers, optional_params, request_data, api_base, api_key=None, model=None, stream=None, fake_stream=None
+    ):
+        self.sign_threads.append(threading.current_thread())
+        return headers, None
+
     def transform_response(
         self,
         model,
@@ -3237,7 +3245,7 @@ class _TransformRecordingConfig(BaseConfig):
         return BaseLLMException(status_code=status_code, message=error_message, headers=headers)
 
 
-def _start_async_completion(config):
+def _start_async_completion(config, logging_obj=None):
     captured = {}
 
     def handle(request):
@@ -3253,7 +3261,7 @@ def _start_async_completion(config):
         custom_llm_provider="openai",
         model_response=ModelResponse(),
         encoding=None,
-        logging_obj=Mock(dynamic_success_callbacks=None, model_call_details={}),
+        logging_obj=logging_obj if logging_obj is not None else Mock(dynamic_success_callbacks=None, model_call_details={}),
         optional_params={},
         timeout=10.0,
         litellm_params={},
@@ -3277,6 +3285,22 @@ async def test_completion_awaits_async_transform_request_when_config_opts_in():
     assert response.choices[0].message.content == "async"
 
 
+async def test_completion_signs_and_logs_off_the_event_loop_after_the_async_transform():
+    config = _TransformRecordingConfig(transform_async=True)
+    loop_thread = threading.current_thread()
+    pre_call_threads = []
+    logging_obj = Mock(dynamic_success_callbacks=None, model_call_details={})
+    logging_obj.pre_call.side_effect = lambda **kwargs: pre_call_threads.append(threading.current_thread())
+
+    pending, captured = _start_async_completion(config, logging_obj)
+    response = await pending
+
+    assert response.choices[0].message.content == "async"
+    assert captured["body"] == {"transformed_by": "async"}
+    assert config.sign_threads and all(thread is not loop_thread for thread in config.sign_threads)
+    assert pre_call_threads and all(thread is not loop_thread for thread in pre_call_threads)
+
+
 async def test_completion_keeps_sync_transform_request_before_returning_by_default():
     config = _TransformRecordingConfig(transform_async=False)
 
diff --git a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
index d31254746d4..f135acd094f 100644
--- a/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
+++ b/tests/test_litellm/llms/vertex_ai/gemini/test_transformation.py
@@ -1,6 +1,8 @@
 
 import json
 import uuid
+from unittest.mock import Mock
+
 import httpx
 import pytest
 
@@ -424,3 +426,52 @@ async def test_gemini_ai_studio_async_completion_passes_files_api_uris_through_u
         {"mime_type": "application/pdf", "file_uri": files_api_pdf},
         {"mime_type": "image/png", "file_uri": files_api_image},
     ]
+
+
+async def test_vertex_ai_async_transform_inlines_only_the_urls_gemini_cannot_fetch_itself(async_only_image_fetch):
+    plain_http_png = f"http://img.example/{uuid.uuid4()}.png"
+    extensionless_https = f"https://cdn.example/files/{uuid.uuid4().hex}"
+    https_png = f"https://img.example/{uuid.uuid4()}.png"
+    hinted_extensionless = f"https://cdn.example/files/{uuid.uuid4().hex}"
+    files_api_pdf = f"https://generativelanguage.googleapis.com/v1beta/files/{uuid.uuid4().hex}"
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Describe these"},
+                {"type": "image_url", "image_url": {"url": plain_http_png}},
+                {"type": "image_url", "image_url": {"url": extensionless_https}},
+                {"type": "image_url", "image_url": {"url": https_png}},
+                {"type": "image_url", "image_url": {"url": hinted_extensionless, "mime_type": "image/webp"}},
+                {"type": "file", "file": {"file_id": files_api_pdf, "format": "application/pdf"}},
+            ],
+        }
+    ]
+
+    body = await transformation.async_transform_request_body(
+        gemini_api_key=None,
+        messages=messages,
+        api_base=None,
+        model="gemini-3.8-flash",
+        client=None,
+        timeout=None,
+        extra_headers=None,
+        optional_params={},
+        logging_obj=Mock(),
+        custom_llm_provider="vertex_ai",
+        litellm_params={},
+        vertex_project="qa-project",
+        vertex_location="us-central1",
+        vertex_auth_header=None,
+    )
+
+    inlined = {"inline_data": {"mime_type": "image/png", "data": async_only_image_fetch.base64_png}}
+    assert body["contents"][0]["parts"] == [
+        {"text": "Describe these"},
+        inlined,
+        inlined,
+        {"file_data": {"mime_type": "image/png", "file_uri": https_png}},
+        {"file_data": {"mime_type": "image/webp", "file_uri": hinted_extensionless}},
+        {"file_data": {"mime_type": "application/pdf", "file_uri": files_api_pdf}},
+    ]
+    assert sorted(async_only_image_fetch.fetched) == sorted([plain_http_png, extensionless_https])

From 836c20a4b6dd931d119ac2cce92977c3c947e062 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 22:31:06 -0700
Subject: [PATCH 24/79] test(ai-gateway): drop provider fingerprint and cache
 identity assertions

---
 litellm-rust/crates/ai-gateway/src/io/tls.rs | 43 +-------------------
 1 file changed, 1 insertion(+), 42 deletions(-)

diff --git a/litellm-rust/crates/ai-gateway/src/io/tls.rs b/litellm-rust/crates/ai-gateway/src/io/tls.rs
index 16fd11e2e79..a2562f60345 100644
--- a/litellm-rust/crates/ai-gateway/src/io/tls.rs
+++ b/litellm-rust/crates/ai-gateway/src/io/tls.rs
@@ -69,26 +69,7 @@ where
 
 #[cfg(test)]
 mod tests {
-    use rustls::CipherSuite;
-    use rustls::NamedGroup;
-    use rustls::crypto::{CryptoProvider, aws_lc_rs, ring};
-
-    use super::{Arc, build_config, tls_config};
-
-    fn fingerprint(provider: &CryptoProvider) -> (Vec<CipherSuite>, Vec<NamedGroup>) {
-        (
-            provider
-                .cipher_suites
-                .iter()
-                .map(|suite| suite.suite())
-                .collect(),
-            provider
-                .kx_groups
-                .iter()
-                .map(|group| group.name())
-                .collect(),
-        )
-    }
+    use super::build_config;
 
     #[test]
     fn builds_a_usable_config_with_both_provider_features_enabled() {
@@ -96,26 +77,4 @@ mod tests {
 
         assert!(!config.crypto_provider().cipher_suites.is_empty());
     }
-
-    #[test]
-    fn dials_with_ring_rather_than_aws_lc_rs() {
-        let config = build_config().expect("a client config");
-
-        assert_eq!(
-            fingerprint(config.crypto_provider()),
-            fingerprint(&ring::default_provider())
-        );
-        assert_ne!(
-            fingerprint(config.crypto_provider()),
-            fingerprint(&aws_lc_rs::default_provider())
-        );
-    }
-
-    #[test]
-    fn the_trust_store_is_loaded_once_and_shared() {
-        let first = tls_config().expect("a client config");
-        let second = tls_config().expect("a client config");
-
-        assert!(Arc::ptr_eq(&first, &second));
-    }
 }

From e1d900d1c29e7f5dbd5db2ca09ab3592312bb1fe Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 22:47:09 -0700
Subject: [PATCH 25/79] fix(realtime): store text_tokens without the nested
 reasoning share

The realtime usage writer passed the provider's output_token_details through as sent, so spend logs and callbacks kept a text_tokens that still contained reasoning_tokens while every other completion_tokens_details producer stores the partitioned share. The writer now applies the same rule the cost calculator uses, moved to litellm/types/utils.py so both read one definition, and the calculator keeps it for usage objects that arrive nested from elsewhere
---
 .../litellm_core_utils/llm_cost_calc/utils.py | 14 +-----
 litellm/responses/utils.py                    | 25 ++++++++--
 litellm/types/utils.py                        | 11 ++++
 .../responses/test_responses_utils.py         | 50 +++++++++++++++++++
 tests/test_litellm/test_cost_calculator.py    | 39 +++++++++++++++
 5 files changed, 123 insertions(+), 16 deletions(-)

diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py
index 1b8980abd35..9432fefc368 100644
--- a/litellm/litellm_core_utils/llm_cost_calc/utils.py
+++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py
@@ -26,6 +26,7 @@ from litellm.types.utils import (
     PromptTokensDetailsWrapper,
     ServiceTier,
     Usage,
+    text_tokens_without_nested_reasoning,
 )
 from litellm.utils import get_model_info
 
@@ -852,17 +853,6 @@ class CompletionTokensDetailsResult(TypedDict):
     video_tokens: int
 
 
-def _text_tokens_without_nested_reasoning(
-    completion_tokens: int,
-    text_tokens: int,
-    reasoning_tokens: int,
-    other_modality_tokens: int,
-) -> int:
-    reported_total: Final = text_tokens + reasoning_tokens + other_modality_tokens
-    nested_reasoning_tokens: Final = min(reasoning_tokens, text_tokens, max(reported_total - completion_tokens, 0))
-    return text_tokens - nested_reasoning_tokens
-
-
 def parse_completion_tokens_details(usage: Usage) -> CompletionTokensDetailsResult:
     audio_tokens: Final = (
         cast(
@@ -893,7 +883,7 @@ def parse_completion_tokens_details(usage: Usage) -> CompletionTokensDetailsResu
         or 0
     )
     video_tokens: Final = _coerce_token_count(getattr(usage.completion_tokens_details, "video_tokens", 0))
-    text_tokens: Final = _text_tokens_without_nested_reasoning(
+    text_tokens: Final = text_tokens_without_nested_reasoning(
         completion_tokens=usage.completion_tokens,
         text_tokens=reported_text_tokens,
         reasoning_tokens=reasoning_tokens,
diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py
index 39675faf735..28f590dcb66 100644
--- a/litellm/responses/utils.py
+++ b/litellm/responses/utils.py
@@ -25,9 +25,15 @@ from litellm.types.utils import (
     PromptTokensDetailsWrapper,
     SpecialEnums,
     Usage,
+    text_tokens_without_nested_reasoning,
 )
 
 
+def _output_token_detail(details: object, field: str) -> int | None:
+    value: Final = getattr(details, field, None)
+    return value if isinstance(value, int) else None
+
+
 def _is_object_sequence(value: object) -> TypeIs[Sequence[object]]:  # guard-ok: a list is a Sequence of anything
     return isinstance(value, list)
 
@@ -1134,11 +1140,22 @@ class ResponseAPILoggingUtils:
             response_api_usage, "output_tokens_details", None
         )
         if output_tokens_details:
+            reasoning_tokens: Final = _output_token_detail(output_tokens_details, "reasoning_tokens")
+            image_tokens: Final = _output_token_detail(output_tokens_details, "image_tokens")
+            audio_tokens: Final = _output_token_detail(output_tokens_details, "audio_tokens")
+            reported_text_tokens: Final = _output_token_detail(output_tokens_details, "text_tokens")
             completion_tokens_details = CompletionTokensDetailsWrapper(
-                reasoning_tokens=getattr(output_tokens_details, "reasoning_tokens", None),
-                image_tokens=getattr(output_tokens_details, "image_tokens", None),
-                text_tokens=getattr(output_tokens_details, "text_tokens", None),
-                audio_tokens=getattr(output_tokens_details, "audio_tokens", None),
+                reasoning_tokens=reasoning_tokens,
+                image_tokens=image_tokens,
+                text_tokens=None
+                if reported_text_tokens is None
+                else text_tokens_without_nested_reasoning(
+                    completion_tokens=completion_tokens,
+                    text_tokens=reported_text_tokens,
+                    reasoning_tokens=reasoning_tokens or 0,
+                    other_modality_tokens=(audio_tokens or 0) + (image_tokens or 0),
+                ),
+                audio_tokens=audio_tokens,
             )
 
         extra_usage_fields: Final = {
diff --git a/litellm/types/utils.py b/litellm/types/utils.py
index 5052cd6ef48..9b24f1d837b 100644
--- a/litellm/types/utils.py
+++ b/litellm/types/utils.py
@@ -1622,6 +1622,17 @@ class Choices(SafeAttributeModel, OpenAIObject):
         setattr(self, key, value)
 
 
+def text_tokens_without_nested_reasoning(
+    completion_tokens: int,
+    text_tokens: int,
+    reasoning_tokens: int,
+    other_modality_tokens: int,
+) -> int:
+    reported_total: Final = text_tokens + reasoning_tokens + other_modality_tokens
+    nested_reasoning_tokens: Final = min(reasoning_tokens, text_tokens, max(reported_total - completion_tokens, 0))
+    return text_tokens - nested_reasoning_tokens
+
+
 class CompletionTokensDetailsWrapper(CompletionTokensDetails):  # wrapper for older openai versions
     text_tokens: int | None = None
     """Text tokens generated by the model."""
diff --git a/tests/test_litellm/responses/test_responses_utils.py b/tests/test_litellm/responses/test_responses_utils.py
index cb6efa21036..9d9eefdceb3 100644
--- a/tests/test_litellm/responses/test_responses_utils.py
+++ b/tests/test_litellm/responses/test_responses_utils.py
@@ -440,6 +440,56 @@ class TestResponseAPILoggingUtils:
         assert result.completion_tokens_details.text_tokens == 20
         assert result.completion_tokens_details.audio_tokens is None
 
+    def test_transform_realtime_usage_partitions_reasoning_out_of_text_tokens(self):
+        """Realtime nests reasoning_tokens inside text_tokens; the stored text share excludes them."""
+        usage = {
+            "input_tokens": 237,
+            "output_tokens": 70,
+            "total_tokens": 307,
+            "input_token_details": {"text_tokens": 43, "audio_tokens": 0, "image_tokens": 194, "cached_tokens": 0},
+            "output_token_details": {"text_tokens": 70, "audio_tokens": 0, "reasoning_tokens": 52},
+        }
+
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
+
+        assert result.completion_tokens == 70
+        assert result.completion_tokens_details is not None
+        assert result.completion_tokens_details.text_tokens == 18
+        assert result.completion_tokens_details.reasoning_tokens == 52
+        assert result.completion_tokens_details.audio_tokens == 0
+
+    def test_transform_realtime_usage_partitions_reasoning_beside_audio_output(self):
+        """Audio output stays as reported; only the text share sheds the nested reasoning tokens."""
+        usage = {
+            "input_tokens": 100,
+            "output_tokens": 70,
+            "total_tokens": 170,
+            "input_token_details": {"text_tokens": 100, "audio_tokens": 0, "cached_tokens": 0},
+            "output_token_details": {"text_tokens": 39, "audio_tokens": 31, "reasoning_tokens": 23},
+        }
+
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
+
+        assert result.completion_tokens_details is not None
+        assert result.completion_tokens_details.text_tokens == 16
+        assert result.completion_tokens_details.audio_tokens == 31
+        assert result.completion_tokens_details.reasoning_tokens == 23
+
+    def test_transform_response_api_usage_keeps_partitioned_text_tokens(self):
+        """A provider already reporting text_tokens beside reasoning_tokens is stored as sent."""
+        usage = {
+            "input_tokens": 10,
+            "output_tokens": 20,
+            "total_tokens": 30,
+            "output_tokens_details": {"text_tokens": 12, "reasoning_tokens": 5},
+        }
+
+        result = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(usage)
+
+        assert result.completion_tokens_details is not None
+        assert result.completion_tokens_details.text_tokens == 12
+        assert result.completion_tokens_details.reasoning_tokens == 5
+
     def test_transform_response_api_usage_carries_extra_provider_fields(self):
         """Non-standard usage fields (e.g. xAI tool details) must survive chat normalization."""
         details = {"web_search_calls": 2, "x_search_calls": 0}
diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py
index 3eb051982e4..1f14fdc2c67 100644
--- a/tests/test_litellm/test_cost_calculator.py
+++ b/tests/test_litellm/test_cost_calculator.py
@@ -4534,3 +4534,42 @@ def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_o
     )
     assert total_cost == pytest.approx(expected)
     assert total_cost == pytest.approx(0.0002362)
+
+
+def test_collect_and_combine_realtime_usage_stores_partitioned_text_tokens() -> None:
+    """The combined usage that lands in spend logs keeps reasoning out of text_tokens for every turn."""
+    results: OpenAIRealtimeStreamList = [
+        {"type": "session.created", "session": {"model": "gpt-realtime-2.1-mini"}},
+        {
+            "type": "response.done",
+            "response": {
+                "usage": {
+                    "total_tokens": 307,
+                    "input_tokens": 237,
+                    "output_tokens": 70,
+                    "input_token_details": {"text_tokens": 43, "audio_tokens": 0, "image_tokens": 194, "cached_tokens": 0},
+                    "output_token_details": {"text_tokens": 70, "audio_tokens": 0, "reasoning_tokens": 52},
+                }
+            },
+        },
+        {
+            "type": "response.done",
+            "response": {
+                "usage": {
+                    "total_tokens": 363,
+                    "input_tokens": 300,
+                    "output_tokens": 63,
+                    "input_token_details": {"text_tokens": 106, "audio_tokens": 0, "image_tokens": 194, "cached_tokens": 0},
+                    "output_token_details": {"text_tokens": 63, "audio_tokens": 0, "reasoning_tokens": 43},
+                }
+            },
+        },
+    ]
+
+    combined = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(results=results)
+
+    assert combined.completion_tokens == 133
+    assert combined.completion_tokens_details is not None
+    assert combined.completion_tokens_details.reasoning_tokens == 95
+    assert combined.completion_tokens_details.text_tokens == 38
+    assert combined.completion_tokens_details.audio_tokens == 0

From 53178486a0c919481bd40da62bfa4ff98c5c2e5a Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 22:59:22 -0700
Subject: [PATCH 26/79] style(tests): wrap realtime cost test lines to the
 120-column limit

---
 .../llm_cost_calc/test_llm_cost_calc_utils.py |  8 +++++--
 tests/test_litellm/test_cost_calculator.py    | 22 +++++++++++++++----
 2 files changed, 24 insertions(+), 6 deletions(-)

diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index e7a92a96f34..e8eae73df3f 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -4800,7 +4800,9 @@ def test_generic_cost_per_token_bills_reasoning_nested_in_text_tokens_once(_loca
     )
 
 
-def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tokens(_local_model_cost_map: None) -> None:
+def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tokens(
+    _local_model_cost_map: None,
+) -> None:
     """Providers whose text_tokens exclude reasoning (text + reasoning == completion) stay billed in full."""
 
     model = "gpt-realtime-2.1-mini"
@@ -4835,4 +4837,6 @@ def test_generic_cost_per_token_bills_nested_reasoning_once_beside_audio_output(
 
     info = litellm.get_model_info(model=model, custom_llm_provider="openai")
     assert breakdown.reasoning_cost == pytest.approx(20 * info["output_cost_per_token"])
-    assert completion_cost == pytest.approx(30 * info["output_cost_per_token"] + 70 * info["output_cost_per_audio_token"])
+    assert completion_cost == pytest.approx(
+        30 * info["output_cost_per_token"] + 70 * info["output_cost_per_audio_token"]
+    )
diff --git a/tests/test_litellm/test_cost_calculator.py b/tests/test_litellm/test_cost_calculator.py
index 1f14fdc2c67..57246ec08a5 100644
--- a/tests/test_litellm/test_cost_calculator.py
+++ b/tests/test_litellm/test_cost_calculator.py
@@ -4494,7 +4494,9 @@ def test_batch_cost_calculator_gpt_6_astra_bills_half_the_standard_rate(_local_m
     assert completion_cost == pytest.approx(500 * 2.5e-5)
 
 
-def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_once(_local_model_cost_map: None) -> None:
+def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_once(
+    _local_model_cost_map: None,
+) -> None:
     """Realtime response.done nests reasoning_tokens inside text_tokens, so they are billed once."""
     results: OpenAIRealtimeStreamList = [
         {"type": "session.created", "session": {"model": "gpt-realtime-2.1-mini"}},
@@ -4530,7 +4532,9 @@ def test_handle_realtime_stream_cost_calculation_bills_nested_reasoning_tokens_o
 
     info = litellm.get_model_info(model="azure/gpt-realtime-2.1-mini", custom_llm_provider="azure")
     expected = (
-        43 * info["input_cost_per_token"] + 194 * info["input_cost_per_image_token"] + 23 * info["output_cost_per_token"]
+        43 * info["input_cost_per_token"]
+        + 194 * info["input_cost_per_image_token"]
+        + 23 * info["output_cost_per_token"]
     )
     assert total_cost == pytest.approx(expected)
     assert total_cost == pytest.approx(0.0002362)
@@ -4547,7 +4551,12 @@ def test_collect_and_combine_realtime_usage_stores_partitioned_text_tokens() ->
                     "total_tokens": 307,
                     "input_tokens": 237,
                     "output_tokens": 70,
-                    "input_token_details": {"text_tokens": 43, "audio_tokens": 0, "image_tokens": 194, "cached_tokens": 0},
+                    "input_token_details": {
+                        "text_tokens": 43,
+                        "audio_tokens": 0,
+                        "image_tokens": 194,
+                        "cached_tokens": 0,
+                    },
                     "output_token_details": {"text_tokens": 70, "audio_tokens": 0, "reasoning_tokens": 52},
                 }
             },
@@ -4559,7 +4568,12 @@ def test_collect_and_combine_realtime_usage_stores_partitioned_text_tokens() ->
                     "total_tokens": 363,
                     "input_tokens": 300,
                     "output_tokens": 63,
-                    "input_token_details": {"text_tokens": 106, "audio_tokens": 0, "image_tokens": 194, "cached_tokens": 0},
+                    "input_token_details": {
+                        "text_tokens": 106,
+                        "audio_tokens": 0,
+                        "image_tokens": 194,
+                        "cached_tokens": 0,
+                    },
                     "output_token_details": {"text_tokens": 63, "audio_tokens": 0, "reasoning_tokens": 43},
                 }
             },

From 0ee3bec046bf0f15eebdddc1d52610d389ca220f Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 23:20:21 -0700
Subject: [PATCH 27/79] fix(image_edit): await an async transform hook and hide
 the URL policy verdict from callers

The image edit handler now awaits BaseImageEditConfig.async_transform_image_edit_request, and
Black Forest Labs overrides it so URL images and masks download through async_safe_get instead of
the blocking safe_get on the event loop. Rejected image fetches raise a fixed policy message with
the user_url_allowed_hosts hint rather than echoing the resolver's verdict (resolved IP, DNS
failure) back to the caller. The test fixture also fails any request-path call of the sync
convert_url_to_base64 so a regression cannot pass unnoticed.
---
 .../prompt_templates/image_handling.py        |  12 +-
 .../base_llm/image_edit/transformation.py     |  19 ++
 .../image_edit/transformation.py              | 102 ++++++-----
 litellm/llms/custom_httpx/llm_http_handler.py |   2 +-
 tests/test_litellm/conftest.py                |  10 +-
 .../litellm_core_utils/test_image_handling.py |  65 ++++---
 .../test_bfl_image_edit_transformation.py     |  78 ++++++++-
 .../custom_httpx/test_llm_http_handler.py     | 163 +++++++++++++++++-
 8 files changed, 379 insertions(+), 72 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index dcad7776b58..4e7dfdb4017 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -78,6 +78,14 @@ def _process_image_response(response: Response, url: str) -> str:
     return result
 
 
+def _url_policy_rejection(url: str, verdict: SSRFError) -> "litellm.ImageFetchError":
+    verbose_logger.warning("Image fetch of %s rejected by the URL policy: %s", url, verdict)
+    return litellm.ImageFetchError(
+        "Error: Unable to fetch image from URL. The proxy's URL policy rejected this host; "
+        f"an admin can allow it with `user_url_allowed_hosts` in general_settings. url={url}"
+    )
+
+
 async def async_convert_url_to_base64(url: str) -> str:
     if url.startswith("data:") and ";base64," in url:
         return url
@@ -100,7 +108,7 @@ async def async_convert_url_to_base64(url: str) -> str:
         except litellm.ImageFetchError:
             raise
         except SSRFError as e:
-            raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL. {e} url={url}") from e
+            raise _url_policy_rejection(url, e) from e
         except Exception:
             pass
     raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL after 3 attempts. url={url}")
@@ -128,7 +136,7 @@ def convert_url_to_base64(url: str) -> str:
         except litellm.ImageFetchError:
             raise
         except SSRFError as e:
-            raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL. {e} url={url}") from e
+            raise _url_policy_rejection(url, e) from e
         except Exception as e:
             verbose_logger.exception(e)
     raise litellm.ImageFetchError(
diff --git a/litellm/llms/base_llm/image_edit/transformation.py b/litellm/llms/base_llm/image_edit/transformation.py
index 9a25d3294e0..4faf0aaaf30 100644
--- a/litellm/llms/base_llm/image_edit/transformation.py
+++ b/litellm/llms/base_llm/image_edit/transformation.py
@@ -1,5 +1,6 @@
 import types
 from abc import ABC, abstractmethod
+from collections.abc import Mapping
 from typing import TYPE_CHECKING, Any
 
 import httpx
@@ -102,6 +103,24 @@ class BaseImageEditConfig(ABC):
     ) -> tuple[dict, RequestFiles]:
         pass
 
+    async def async_transform_image_edit_request(
+        self,
+        model: str,
+        prompt: str | None,
+        image: FileTypes | None,
+        image_edit_optional_request_params: Mapping[str, object],
+        litellm_params: GenericLiteLLMParams,
+        headers: Mapping[str, str],
+    ) -> tuple[dict, RequestFiles]:
+        return self.transform_image_edit_request(
+            model=model,
+            prompt=prompt,
+            image=image,
+            image_edit_optional_request_params=dict(image_edit_optional_request_params),
+            litellm_params=litellm_params,
+            headers=dict(headers),
+        )
+
     def finalize_image_edit_request_data(self, data: dict, resolved_request_url: str) -> dict:
         """
         Last pass on the request dict after ``transform_image_edit_request``, using the
diff --git a/litellm/llms/black_forest_labs/image_edit/transformation.py b/litellm/llms/black_forest_labs/image_edit/transformation.py
index 013053e5bd5..c6de9f9ba65 100644
--- a/litellm/llms/black_forest_labs/image_edit/transformation.py
+++ b/litellm/llms/black_forest_labs/image_edit/transformation.py
@@ -9,6 +9,7 @@ API Reference: https://docs.bfl.ai/
 
 import base64
 import time
+from collections.abc import Mapping
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
@@ -16,7 +17,7 @@ from httpx._types import RequestFiles
 
 import litellm
 from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH
-from litellm.litellm_core_utils.url_utils import safe_get
+from litellm.litellm_core_utils.url_utils import async_safe_get, safe_get
 from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.images.main import ImageEditOptionalRequestParams
@@ -37,6 +38,22 @@ else:
     LiteLLMLoggingObj = Any
 
 
+_BFL_REQUEST_PARAMS: Final = (
+    "seed",
+    "output_format",
+    "safety_tolerance",
+    "prompt_upsampling",
+    "aspect_ratio",
+    "steps",
+    "guidance",
+    "grow_mask",
+    "top",
+    "bottom",
+    "left",
+    "right",
+)
+
+
 class BlackForestLabsImageEditConfig(BaseImageEditConfig):
     """
     Configuration for Black Forest Labs image editing.
@@ -85,34 +102,10 @@ class BlackForestLabsImageEditConfig(BaseImageEditConfig):
         BFL-specific params are passed through directly.
         """
         optional_params: Final[dict[str, Any]] = {}
-
-        # Pass through BFL-specific params
-        bfl_params: Final = [
-            "seed",
-            "output_format",
-            "safety_tolerance",
-            "prompt_upsampling",
-            # Kontext-specific
-            "aspect_ratio",
-            # Fill/Inpaint-specific
-            "steps",
-            "guidance",
-            "grow_mask",
-            # Expand-specific
-            "top",
-            "bottom",
-            "left",
-            "right",
-        ]
-
-        # Convert TypedDict to regular dict for access
-        params_dict: Final = dict(image_edit_optional_params)
-
-        for param in bfl_params:
-            if param in params_dict:
-                value = params_dict[param]
-                if value is not None:
-                    optional_params[param] = value
+        params: Final[Mapping[str, object]] = image_edit_optional_params
+        for param in _BFL_REQUEST_PARAMS:
+            if (value := params.get(param)) is not None:
+                optional_params[param] = value
 
         # Set default output format
         if "output_format" not in optional_params:
@@ -251,23 +244,8 @@ class BlackForestLabsImageEditConfig(BaseImageEditConfig):
             "input_image": b64_image,
         }
 
-        # Add optional params (only BFL-recognized parameters)
-        bfl_request_params: Final = [
-            "seed",
-            "output_format",
-            "safety_tolerance",
-            "prompt_upsampling",
-            "aspect_ratio",
-            "steps",
-            "guidance",
-            "grow_mask",
-            "top",
-            "bottom",
-            "left",
-            "right",
-        ]
         for key, value in image_edit_optional_request_params.items():
-            if key in bfl_request_params and value is not None:
+            if key in _BFL_REQUEST_PARAMS and value is not None:
                 request_body[key] = value
 
         # Handle mask if provided (for inpainting)
@@ -277,7 +255,39 @@ class BlackForestLabsImageEditConfig(BaseImageEditConfig):
             request_body["mask"] = base64.b64encode(mask_bytes).decode("utf-8")
 
         # BFL uses JSON, not multipart - return empty files
-        return request_body, []
+        return request_body, ()
+
+    async def async_transform_image_edit_request(
+        self,
+        model: str,
+        prompt: str | None,
+        image: FileTypes | None,
+        image_edit_optional_request_params: Mapping[str, object],
+        litellm_params: GenericLiteLLMParams,
+        headers: Mapping[str, str],
+    ) -> tuple[dict, RequestFiles]:
+        downloaded_image: Final = await self._fetch_remote_image(image)
+        downloaded_mask: Final = await self._fetch_remote_image(image_edit_optional_request_params.get("mask"))
+        return self.transform_image_edit_request(
+            model=model,
+            prompt=prompt,
+            image=image if downloaded_image is None else downloaded_image,
+            image_edit_optional_request_params=(
+                dict(image_edit_optional_request_params)
+                if downloaded_mask is None
+                else {**image_edit_optional_request_params, "mask": downloaded_mask}
+            ),
+            litellm_params=litellm_params,
+            headers=dict(headers),
+        )
+
+    async def _fetch_remote_image(self, image: object) -> bytes | None:
+        candidate: Final = image[0] if isinstance(image, list) and image else image
+        if not isinstance(candidate, str) or not candidate.startswith(("http://", "https://")):
+            return None
+        response: Final = await async_safe_get(litellm.module_level_aclient, candidate, timeout=60.0)
+        response.raise_for_status()
+        return response.content
 
     def transform_image_edit_response(
         self,
diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py
index d57aee025d5..6a17841faa2 100644
--- a/litellm/llms/custom_httpx/llm_http_handler.py
+++ b/litellm/llms/custom_httpx/llm_http_handler.py
@@ -6787,7 +6787,7 @@ class BaseLLMHTTPHandler:
             litellm_params=dict(litellm_params),
         )
 
-        data, files = image_edit_provider_config.transform_image_edit_request(
+        data, files = await image_edit_provider_config.async_transform_image_edit_request(
             model=model,
             image=image,
             prompt=prompt,
diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py
index aa8ba168ecc..a4f32df46ae 100644
--- a/tests/test_litellm/conftest.py
+++ b/tests/test_litellm/conftest.py
@@ -607,7 +607,8 @@ ONE_PIXEL_PNG = base64.b64decode(
 
 @pytest.fixture
 def async_only_image_fetch(monkeypatch):
-    from litellm.litellm_core_utils.prompt_templates import image_handling
+    from litellm.litellm_core_utils.prompt_templates import factory, image_handling
+    from litellm.llms.gemini.chat import transformation as gemini_chat_transformation
 
     fetch = SimpleNamespace(
         fetched=[],
@@ -627,6 +628,13 @@ def async_only_image_fetch(monkeypatch):
             request=httpx.Request("GET", url),
         )
 
+    def forbid_sync_convert(url, *args, **kwargs):
+        if url.startswith(("http://", "https://")):
+            raise litellm.ImageFetchError(f"sync convert_url_to_base64 ran on the request path: {url}")
+        return url
+
     monkeypatch.setattr(image_handling, "safe_get", forbid_sync_fetch)
     monkeypatch.setattr(image_handling, "async_safe_get", serve_png)
+    for module in (image_handling, factory, gemini_chat_transformation):
+        monkeypatch.setattr(module, "convert_url_to_base64", forbid_sync_convert)
     return fetch
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index e7b28d0d3a2..d8ccc9251dd 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -428,36 +428,61 @@ async def test_async_inline_remote_media_cancels_the_other_fetches_when_one_fail
     assert time.perf_counter() - started < 1
 
 
-async def test_async_convert_url_to_base64_reports_a_blocked_url_without_retrying(monkeypatch):
+_SSRF_VERDICTS = (
+    "URL targets a blocked address (10.0.0.8). If this is a legitimate internal service, "
+    "add the host to `user_url_allowed_hosts` in general_settings.",
+    "DNS resolution failed for 'internal.example': [Errno 8] nodename nor servname provided, or not known",
+    "No addresses found for 'internal.example'",
+)
+
+
+def _assert_one_verdict_free_message(messages, url):
+    assert len(set(messages)) == 1
+    assert "10.0.0.8" not in messages[0]
+    assert "DNS" not in messages[0]
+    assert "No addresses" not in messages[0]
+    assert "user_url_allowed_hosts" in messages[0]
+    assert url in messages[0]
+
+
+async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeypatch):
     attempts = []
-
-    async def block(client, url, **kwargs):
-        attempts.append(url)
-        raise SSRFError("URL targets a blocked address (10.0.0.8)")
-
-    monkeypatch.setattr(image_handling, "async_safe_get", block)
+    messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    with pytest.raises(litellm.ImageFetchError, match=r"blocked address \(10\.0\.0\.8\)"):
-        await async_convert_url_to_base64(url)
+    for verdict in _SSRF_VERDICTS:
 
-    assert attempts == [url]
+        async def block(client, fetched_url, verdict=verdict, **kwargs):
+            attempts.append(fetched_url)
+            raise SSRFError(verdict)
+
+        monkeypatch.setattr(image_handling, "async_safe_get", block)
+        with pytest.raises(litellm.ImageFetchError) as raised:
+            await async_convert_url_to_base64(url)
+        messages.append(raised.value.message)
+
+    assert attempts == [url] * len(_SSRF_VERDICTS)
+    _assert_one_verdict_free_message(messages, url)
 
 
-def test_convert_url_to_base64_reports_a_blocked_url_without_retrying(monkeypatch):
+def test_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeypatch):
     attempts = []
-
-    def block(client, url, **kwargs):
-        attempts.append(url)
-        raise SSRFError("URL targets a blocked address (10.0.0.8)")
-
-    monkeypatch.setattr(image_handling, "safe_get", block)
+    messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    with pytest.raises(litellm.ImageFetchError, match=r"blocked address \(10\.0\.0\.8\)"):
-        convert_url_to_base64(url)
+    for verdict in _SSRF_VERDICTS:
 
-    assert attempts == [url]
+        def block(client, fetched_url, verdict=verdict, **kwargs):
+            attempts.append(fetched_url)
+            raise SSRFError(verdict)
+
+        monkeypatch.setattr(image_handling, "safe_get", block)
+        with pytest.raises(litellm.ImageFetchError) as raised:
+            convert_url_to_base64(url)
+        messages.append(raised.value.message)
+
+    assert attempts == [url] * len(_SSRF_VERDICTS)
+    _assert_one_verdict_free_message(messages, url)
 
 
 async def test_async_inline_remote_media_caps_in_flight_fetches_per_request(monkeypatch):
diff --git a/tests/test_litellm/llms/black_forest_labs/image_edit/test_bfl_image_edit_transformation.py b/tests/test_litellm/llms/black_forest_labs/image_edit/test_bfl_image_edit_transformation.py
index ec243b7058d..d9d6e813d86 100644
--- a/tests/test_litellm/llms/black_forest_labs/image_edit/test_bfl_image_edit_transformation.py
+++ b/tests/test_litellm/llms/black_forest_labs/image_edit/test_bfl_image_edit_transformation.py
@@ -16,6 +16,9 @@ import httpx
 import pytest
 
 
+from litellm.llms.black_forest_labs.image_edit import (
+    transformation as bfl_transformation,
+)
 from litellm.llms.black_forest_labs.image_edit.transformation import (
     BlackForestLabsImageEditConfig,
 )
@@ -186,7 +189,7 @@ class TestBlackForestLabsImageEditTransformation:
         assert data["output_format"] == "jpeg"
 
         # BFL uses JSON, not multipart - files should be empty
-        assert files == []
+        assert files == ()
 
     def test_transform_image_edit_request_with_mask(self):
         """Test request transformation with mask for inpainting."""
@@ -299,3 +302,76 @@ class TestBlackForestLabsImageEditTransformation:
     def test_use_multipart_form_data_returns_false(self):
         """Test that use_multipart_form_data returns False for BFL."""
         assert self.config.use_multipart_form_data() is False
+
+
+async def test_async_transform_image_edit_request_downloads_url_images_with_the_async_fetcher(monkeypatch):
+    served = b"png-bytes-from-cdn"
+    fetched = []
+
+    def forbid_sync_fetch(client, url, **kwargs):
+        raise AssertionError(f"sync image fetch ran on the event loop: {url}")
+
+    async def serve(client, url, **kwargs):
+        fetched.append((url, kwargs.get("timeout")))
+        return httpx.Response(200, content=served, request=httpx.Request("GET", url))
+
+    monkeypatch.setattr(bfl_transformation, "safe_get", forbid_sync_fetch)
+    monkeypatch.setattr(bfl_transformation, "async_safe_get", serve)
+
+    data, files = await BlackForestLabsImageEditConfig().async_transform_image_edit_request(
+        model="flux-kontext-pro",
+        prompt="Add a red hat",
+        image="https://cdn.example/photo.png",
+        image_edit_optional_request_params={"mask": "https://cdn.example/mask.png", "seed": 7},
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    assert base64.b64decode(data["input_image"]) == served
+    assert base64.b64decode(data["mask"]) == served
+    assert data["seed"] == 7
+    assert files == ()
+    assert fetched == [("https://cdn.example/photo.png", 60.0), ("https://cdn.example/mask.png", 60.0)]
+
+
+async def test_async_transform_image_edit_request_never_fetches_for_local_images(monkeypatch):
+    def refuse(*args, **kwargs):
+        raise AssertionError("no network fetch expected for local image bytes")
+
+    monkeypatch.setattr(bfl_transformation, "safe_get", refuse)
+    monkeypatch.setattr(bfl_transformation, "async_safe_get", refuse)
+
+    data, _ = await BlackForestLabsImageEditConfig().async_transform_image_edit_request(
+        model="flux-kontext-pro",
+        prompt="Add a red hat",
+        image=[BytesIO(b"first"), BytesIO(b"other")],
+        image_edit_optional_request_params={"mask": b"mask-bytes"},
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    assert base64.b64decode(data["input_image"]) == b"first"
+    assert base64.b64decode(data["mask"]) == b"mask-bytes"
+
+
+async def test_async_transform_image_edit_request_downloads_only_the_first_url_of_a_list(monkeypatch):
+    fetched = []
+
+    async def serve(client, url, **kwargs):
+        fetched.append(url)
+        return httpx.Response(200, content=b"first-bytes", request=httpx.Request("GET", url))
+
+    monkeypatch.setattr(bfl_transformation, "safe_get", lambda *args, **kwargs: pytest.fail("sync fetch ran"))
+    monkeypatch.setattr(bfl_transformation, "async_safe_get", serve)
+
+    data, _ = await BlackForestLabsImageEditConfig().async_transform_image_edit_request(
+        model="flux-kontext-pro",
+        prompt="Add a red hat",
+        image=["https://cdn.example/a.png", "https://cdn.example/b.png"],
+        image_edit_optional_request_params={},
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    assert fetched == ["https://cdn.example/a.png"]
+    assert base64.b64decode(data["input_image"]) == b"first-bytes"
diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
index f1614654ffb..3afd57bbf34 100644
--- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
+++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
@@ -19,6 +19,7 @@ from litellm.llms.base_llm.audio_transcription.transformation import (
     BaseAudioTranscriptionConfig,
 )
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
+from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.custom_httpx.llm_http_handler import (
     BaseLLMHTTPHandler,
@@ -31,7 +32,7 @@ from litellm.llms.azure.videos.transformation import AzureVideoConfig
 from litellm.llms.openai.videos.transformation import OpenAIVideoConfig
 from litellm.types.llms.openai import ResponsesAPIResponse
 from litellm.types.router import GenericLiteLLMParams
-from litellm.types.utils import ModelResponse, TranscriptionResponse
+from litellm.types.utils import ImageObject, ImageResponse, ModelResponse, TranscriptionResponse
 
 _ACTIVE_KEY = "_code_interpreter_interception_active"
 _SANDBOX_KEY = "_code_interpreter_interception_sandbox_key"
@@ -3244,6 +3245,11 @@ class _TransformRecordingConfig(BaseConfig):
     def get_error_class(self, error_message, status_code, headers):
         return BaseLLMException(status_code=status_code, message=error_message, headers=headers)
 
+    def get_model_response_iterator(self, streaming_response, sync_stream, json_mode=False):
+        return litellm.OpenAIGPTConfig().get_model_response_iterator(
+            streaming_response=streaming_response, sync_stream=sync_stream, json_mode=json_mode
+        )
+
 
 def _start_async_completion(config, logging_obj=None):
     captured = {}
@@ -3312,3 +3318,158 @@ async def test_completion_keeps_sync_transform_request_before_returning_by_defau
     assert config.transform_calls == ["sync"]
     assert captured["body"] == {"transformed_by": "sync"}
     assert response.choices[0].message.content == "sync"
+
+
+def _sse_echoing_transformed_by(request):
+    transformed_by = json.loads(request.content)["transformed_by"]
+    chunk = {
+        "id": "chatcmpl-1",
+        "object": "chat.completion.chunk",
+        "created": 1,
+        "model": "stub-model",
+        "choices": [{"index": 0, "delta": {"content": transformed_by}, "finish_reason": None}],
+    }
+    return httpx.Response(
+        200,
+        content=f"data: {json.dumps(chunk)}\n\ndata: [DONE]\n\n".encode(),
+        headers={"content-type": "text/event-stream"},
+        request=request,
+    )
+
+
+def _streaming_logging_obj():
+    from litellm.litellm_core_utils.litellm_logging import Logging
+
+    logging_obj = Logging(
+        model="stub-model",
+        messages=[{"role": "user", "content": "hi"}],
+        stream=True,
+        call_type="acompletion",
+        start_time=time.time(),
+        litellm_call_id="async-transform-stream",
+        function_id="f",
+    )
+    logging_obj.update_environment_variables(
+        model="stub-model", user="", optional_params={}, litellm_params={}, custom_llm_provider="openai"
+    )
+    return logging_obj
+
+
+async def test_completion_streams_after_the_async_transform_request():
+    config = _TransformRecordingConfig(transform_async=True)
+    loop_thread = threading.current_thread()
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(_sse_echoing_transformed_by))
+
+    stream = await BaseLLMHTTPHandler().completion(
+        model="stub-model",
+        messages=[{"role": "user", "content": "hi"}],
+        api_base="https://llm.example/v1/chat",
+        custom_llm_provider="openai",
+        model_response=ModelResponse(),
+        encoding=None,
+        logging_obj=_streaming_logging_obj(),
+        optional_params={},
+        timeout=10.0,
+        litellm_params={},
+        acompletion=True,
+        stream=True,
+        client=client,
+        provider_config=config,
+    )
+    collected = [chunk async for chunk in stream]
+
+    assert config.transform_calls == ["async"]
+    assert config.sign_threads and all(thread is not loop_thread for thread in config.sign_threads)
+    assert "".join(chunk.choices[0].delta.content or "" for chunk in collected) == "async"
+
+
+class _ImageEditRecordingConfig(BaseImageEditConfig):
+    def __init__(self):
+        self.transform_calls = []
+
+    def get_supported_openai_params(self, model):
+        return []
+
+    def map_openai_params(self, image_edit_optional_params, model, drop_params):
+        return dict(image_edit_optional_params)
+
+    def validate_environment(self, headers, model, api_key=None, litellm_params=None, api_base=None):
+        return {}
+
+    def get_complete_url(self, model, api_base, litellm_params):
+        return "https://images.example/v1/edits"
+
+    def use_multipart_form_data(self):
+        return False
+
+    def transform_image_edit_request(
+        self, model, prompt, image, image_edit_optional_request_params, litellm_params, headers
+    ):
+        self.transform_calls.append("sync")
+        return {"transformed_by": "sync"}, []
+
+    async def async_transform_image_edit_request(
+        self, model, prompt, image, image_edit_optional_request_params, litellm_params, headers
+    ):
+        self.transform_calls.append("async")
+        return {"transformed_by": "async"}, []
+
+    def transform_image_edit_response(self, model, raw_response, logging_obj):
+        return ImageResponse(data=[ImageObject(b64_json=raw_response.json()["transformed_by"])])
+
+
+def _echo_json_transport(captured):
+    def handle(request):
+        captured["body"] = json.loads(request.content)
+        return httpx.Response(200, json=captured["body"])
+
+    return httpx.MockTransport(handle)
+
+
+async def test_async_image_edit_handler_awaits_the_async_transform():
+    config = _ImageEditRecordingConfig()
+    captured = {}
+    client = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=_echo_json_transport(captured))
+
+    response = await BaseLLMHTTPHandler().async_image_edit_handler(
+        model="edit-model",
+        image=b"raw-image",
+        prompt="add a hat",
+        image_edit_provider_config=config,
+        image_edit_optional_request_params={},
+        custom_llm_provider="openai",
+        litellm_params=GenericLiteLLMParams(),
+        logging_obj=Mock(),
+        timeout=10.0,
+        client=client,
+    )
+
+    assert config.transform_calls == ["async"]
+    assert captured["body"] == {"transformed_by": "async"}
+    assert response.data[0].b64_json == "async"
+
+
+def test_image_edit_handler_keeps_the_sync_transform():
+    config = _ImageEditRecordingConfig()
+    captured = {}
+    client = HTTPHandler()
+    client.client = httpx.Client(transport=_echo_json_transport(captured))
+
+    response = BaseLLMHTTPHandler().image_edit_handler(
+        model="edit-model",
+        image=b"raw-image",
+        prompt="add a hat",
+        image_edit_provider_config=config,
+        image_edit_optional_request_params={},
+        custom_llm_provider="openai",
+        litellm_params=GenericLiteLLMParams(),
+        logging_obj=Mock(),
+        timeout=10.0,
+        client=client,
+    )
+
+    assert config.transform_calls == ["sync"]
+    assert captured["body"] == {"transformed_by": "sync"}
+    assert response.data[0].b64_json == "sync"

From 3920cf4dfe458b6d4982e2a64c162de21defa2a0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 23:45:37 -0700
Subject: [PATCH 28/79] fix(image_handling): tell callers when the image host
 did not resolve instead of blaming the URL policy

validate_url raises HostResolutionError, a SSRFError subclass, for the two
DNS outcomes (lookup failed, no addresses). The image fetch helper maps
that to a "host could not be resolved" message and keeps the
user_url_allowed_hosts hint for the policy verdicts it can actually fix.
---
 .../prompt_templates/image_handling.py        | 14 +++--
 litellm/litellm_core_utils/url_utils.py       |  8 ++-
 .../litellm_core_utils/test_image_handling.py | 56 +++++++++++--------
 .../litellm_core_utils/test_url_utils.py      | 19 ++++++-
 4 files changed, 64 insertions(+), 33 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 4e7dfdb4017..99efbd13320 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -15,7 +15,7 @@ import litellm
 from litellm import verbose_logger
 from litellm.caching.caching import InMemoryCache
 from litellm.constants import MAX_IMAGE_URL_DOWNLOAD_SIZE_MB
-from litellm.litellm_core_utils.url_utils import SSRFError, async_safe_get, safe_get
+from litellm.litellm_core_utils.url_utils import HostResolutionError, SSRFError, async_safe_get, safe_get
 from litellm.types.llms.openai import AllMessageValues
 
 MAX_IMGS_IN_MEMORY: Final = 10
@@ -78,8 +78,12 @@ def _process_image_response(response: Response, url: str) -> str:
     return result
 
 
-def _url_policy_rejection(url: str, verdict: SSRFError) -> "litellm.ImageFetchError":
-    verbose_logger.warning("Image fetch of %s rejected by the URL policy: %s", url, verdict)
+def _rejected_image_fetch(url: str, verdict: SSRFError) -> "litellm.ImageFetchError":
+    verbose_logger.warning("Image fetch of %s rejected before any request went out: %s", url, verdict)
+    if isinstance(verdict, HostResolutionError):
+        return litellm.ImageFetchError(
+            f"Error: Unable to fetch image from URL. The image host could not be resolved. url={url}"
+        )
     return litellm.ImageFetchError(
         "Error: Unable to fetch image from URL. The proxy's URL policy rejected this host; "
         f"an admin can allow it with `user_url_allowed_hosts` in general_settings. url={url}"
@@ -108,7 +112,7 @@ async def async_convert_url_to_base64(url: str) -> str:
         except litellm.ImageFetchError:
             raise
         except SSRFError as e:
-            raise _url_policy_rejection(url, e) from e
+            raise _rejected_image_fetch(url, e) from e
         except Exception:
             pass
     raise litellm.ImageFetchError(f"Error: Unable to fetch image from URL after 3 attempts. url={url}")
@@ -136,7 +140,7 @@ def convert_url_to_base64(url: str) -> str:
         except litellm.ImageFetchError:
             raise
         except SSRFError as e:
-            raise _url_policy_rejection(url, e) from e
+            raise _rejected_image_fetch(url, e) from e
         except Exception as e:
             verbose_logger.exception(e)
     raise litellm.ImageFetchError(
diff --git a/litellm/litellm_core_utils/url_utils.py b/litellm/litellm_core_utils/url_utils.py
index fa070a648f5..f4388d5fd12 100644
--- a/litellm/litellm_core_utils/url_utils.py
+++ b/litellm/litellm_core_utils/url_utils.py
@@ -93,6 +93,10 @@ class SSRFError(ValueError):
     """Raised when a URL targets a blocked network."""
 
 
+class HostResolutionError(SSRFError):
+    pass
+
+
 def encode_url_path_segment(value: object, *, field_name: str = "path parameter") -> str:
     """Percent-encode one user-controlled URL path segment.
 
@@ -324,10 +328,10 @@ def validate_url(url: str) -> tuple[str, str]:
     try:
         addrinfo: Final = socket.getaddrinfo(hostname, effective_port, proto=socket.IPPROTO_TCP)
     except socket.gaierror as e:
-        raise SSRFError(f"DNS resolution failed for '{hostname}': {e}")
+        raise HostResolutionError(f"DNS resolution failed for '{hostname}': {e}")
 
     if not addrinfo:
-        raise SSRFError(f"No addresses found for '{hostname}'")
+        raise HostResolutionError(f"No addresses found for '{hostname}'")
 
     if not is_allowlisted:
         for addrinfo_entry in addrinfo:
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index d8ccc9251dd..3a909298aba 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -17,7 +17,7 @@ from litellm.litellm_core_utils.prompt_templates.image_handling import (
     async_inline_remote_media,
     convert_url_to_base64,
 )
-from litellm.litellm_core_utils.url_utils import SSRFError
+from litellm.litellm_core_utils.url_utils import HostResolutionError, SSRFError
 
 
 @pytest.fixture(autouse=True)
@@ -115,9 +115,7 @@ class StreamingLargeImageClient:
             request=Request("GET", url),
         )
         # Mock the iter_bytes method to return our generator
-        response.iter_bytes = lambda chunk_size=8192: generate_chunks(
-            size_bytes, chunk_size
-        )
+        response.iter_bytes = lambda chunk_size=8192: generate_chunks(size_bytes, chunk_size)
         return response
 
 
@@ -215,9 +213,7 @@ def test_streaming_download_handles_petabyte_file(monkeypatch):
     """
     # Simulate a 1 petabyte file (1,000,000 GB)
     # Without streaming protection, this would cause OOM or hang indefinitely
-    client = StreamingLargeImageClient(
-        size_mb=1_000_000_000, include_content_length=False
-    )
+    client = StreamingLargeImageClient(size_mb=1_000_000_000, include_content_length=False)
     monkeypatch.setattr(litellm, "module_level_client", client)
 
     with pytest.raises(litellm.ImageFetchError) as excinfo:
@@ -429,20 +425,32 @@ async def test_async_inline_remote_media_cancels_the_other_fetches_when_one_fail
 
 
 _SSRF_VERDICTS = (
-    "URL targets a blocked address (10.0.0.8). If this is a legitimate internal service, "
-    "add the host to `user_url_allowed_hosts` in general_settings.",
-    "DNS resolution failed for 'internal.example': [Errno 8] nodename nor servname provided, or not known",
-    "No addresses found for 'internal.example'",
+    (
+        SSRFError(
+            "URL targets a blocked address (10.0.0.8). If this is a legitimate internal service, "
+            "add the host to `user_url_allowed_hosts` in general_settings."
+        ),
+        "The proxy's URL policy rejected this host; an admin can allow it with `user_url_allowed_hosts`",
+    ),
+    (
+        HostResolutionError(
+            "DNS resolution failed for 'internal.example': [Errno 8] nodename nor servname provided, or not known"
+        ),
+        "The image host could not be resolved",
+    ),
+    (HostResolutionError("No addresses found for 'internal.example'"), "The image host could not be resolved"),
 )
 
 
-def _assert_one_verdict_free_message(messages, url):
-    assert len(set(messages)) == 1
-    assert "10.0.0.8" not in messages[0]
-    assert "DNS" not in messages[0]
-    assert "No addresses" not in messages[0]
-    assert "user_url_allowed_hosts" in messages[0]
-    assert url in messages[0]
+def _assert_verdict_free_messages(messages, url):
+    for message, (verdict, expected_guidance) in zip(messages, _SSRF_VERDICTS, strict=True):
+        assert expected_guidance in message
+        assert url in message
+        assert "10.0.0.8" not in message
+        assert "DNS" not in message
+        assert "No addresses" not in message
+        if isinstance(verdict, HostResolutionError):
+            assert "user_url_allowed_hosts" not in message
 
 
 async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeypatch):
@@ -450,11 +458,11 @@ async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_r
     messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    for verdict in _SSRF_VERDICTS:
+    for verdict, _ in _SSRF_VERDICTS:
 
         async def block(client, fetched_url, verdict=verdict, **kwargs):
             attempts.append(fetched_url)
-            raise SSRFError(verdict)
+            raise verdict
 
         monkeypatch.setattr(image_handling, "async_safe_get", block)
         with pytest.raises(litellm.ImageFetchError) as raised:
@@ -462,7 +470,7 @@ async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_r
         messages.append(raised.value.message)
 
     assert attempts == [url] * len(_SSRF_VERDICTS)
-    _assert_one_verdict_free_message(messages, url)
+    _assert_verdict_free_messages(messages, url)
 
 
 def test_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeypatch):
@@ -470,11 +478,11 @@ def test_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeyp
     messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    for verdict in _SSRF_VERDICTS:
+    for verdict, _ in _SSRF_VERDICTS:
 
         def block(client, fetched_url, verdict=verdict, **kwargs):
             attempts.append(fetched_url)
-            raise SSRFError(verdict)
+            raise verdict
 
         monkeypatch.setattr(image_handling, "safe_get", block)
         with pytest.raises(litellm.ImageFetchError) as raised:
@@ -482,7 +490,7 @@ def test_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeyp
         messages.append(raised.value.message)
 
     assert attempts == [url] * len(_SSRF_VERDICTS)
-    _assert_one_verdict_free_message(messages, url)
+    _assert_verdict_free_messages(messages, url)
 
 
 async def test_async_inline_remote_media_caps_in_flight_fetches_per_request(monkeypatch):
diff --git a/tests/test_litellm/litellm_core_utils/test_url_utils.py b/tests/test_litellm/litellm_core_utils/test_url_utils.py
index fccdc1a2a0a..151c7ceb5bd 100644
--- a/tests/test_litellm/litellm_core_utils/test_url_utils.py
+++ b/tests/test_litellm/litellm_core_utils/test_url_utils.py
@@ -9,6 +9,7 @@ import pytest
 import litellm
 from litellm.litellm_core_utils import url_utils
 from litellm.litellm_core_utils.url_utils import (
+    HostResolutionError,
     SSRFError,
     _is_blocked_ip,
     assert_same_origin,
@@ -161,10 +162,24 @@ class TestValidateUrl:
         assert "/path" in rewritten
         assert "key=value" in rewritten
 
-    def test_dns_failure_raises(self, mock_dns_failure):
-        with pytest.raises(SSRFError, match="DNS resolution failed"):
+    def test_dns_failure_raises_a_host_resolution_error(self, mock_dns_failure):
+        with pytest.raises(HostResolutionError, match="DNS resolution failed"):
             validate_url("http://this-domain-does-not-exist-xyz123.invalid/test")
 
+    def test_empty_resolution_raises_a_host_resolution_error(self, monkeypatch):
+        monkeypatch.setattr(url_utils.socket, "getaddrinfo", lambda *args, **kwargs: [])
+        with pytest.raises(HostResolutionError, match="No addresses found"):
+            validate_url("http://this-domain-resolves-to-nothing.invalid/test")
+
+    def test_blocked_address_is_not_a_host_resolution_error(self, monkeypatch):
+        def fake(host, port, *a, **kw):
+            return [(socket.AF_INET, socket.SOCK_STREAM, 6, "", ("10.0.0.8", port or 80))]
+
+        monkeypatch.setattr(url_utils.socket, "getaddrinfo", fake)
+        with pytest.raises(SSRFError, match="blocked address") as raised:
+            validate_url("http://internal.example/test")
+        assert not isinstance(raised.value, HostResolutionError)
+
     def test_blocks_localhost_hostname(self, monkeypatch):
         def fake(host, port, *a, **kw):
             return [

From de2ba3fab1daab25ba2517ea1a93cf7120f996f5 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 4 Sep 2026 23:50:01 -0700
Subject: [PATCH 29/79] fix(make check): lint the test tree on tests-only
 changes like CI does

CI's required lint job runs ruff with ruff-tests.toml over tests/ and the
test-quality budget gate, but scripts/pre_commit_lint.sh only triggered make
lint on litellm/ files, so a tests-only commit passed make check with a no-op
note and then failed CI (a duplicate test name, ruff F811, did exactly that).

When tests/ Python files are in scope and no litellm/ files are, run ruff
with ruff-tests.toml over the changed test files and make lint-test-quality,
with the matching partial-staging warning, summary line, and no-op condition.
---
 scripts/pre_commit_lint.sh                 | 25 +++++-
 tests/test_litellm/test_pre_commit_lint.py | 98 ++++++++++++++++++++--
 2 files changed, 114 insertions(+), 9 deletions(-)

diff --git a/scripts/pre_commit_lint.sh b/scripts/pre_commit_lint.sh
index d38a0eee3de..ad1b4e793ab 100755
--- a/scripts/pre_commit_lint.sh
+++ b/scripts/pre_commit_lint.sh
@@ -12,6 +12,8 @@
 #   - litellm/ Python  -> `make lint` (test-linting.yml's lint job)
 #   - tests/e2e Python -> `make lint-e2e-basedpyright` (test-linting.yml's e2e type-check step)
 #                         + raw HTTP client ban (test-code-quality.yml's check_e2e_no_raw_requests)
+#   - tests/ Python    -> ruff over ruff-tests.toml + `make lint-test-quality` (test-linting.yml's
+#                         test-tree ruff and test-quality budget steps)
 #   - dashboard        -> prettier + eslint + lint budgets (test-litellm-ui-build.yml's frontend-lint)
 #   - proxy/types      -> regenerate the lazy OpenAPI snapshot and dashboard API types, fail on drift (check-ui-api-types.yml)
 #
@@ -88,15 +90,18 @@ existing_files() {
 
 litellm_py_pattern='^litellm/.*\.py$'
 e2e_py_pattern='^tests/e2e/.*\.py$'
+tests_py_pattern='^tests/.*\.py$'
 spec_pattern='^(litellm/(proxy|types)/.*|ui/litellm-dashboard/(scripts/gen-api-types\.mjs|package\.json|package-lock\.json|src/lib/http/schema\.d\.ts))$'
 ui_prettier_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs|json|css|scss|md|mdx|yml|yaml|html)$'
 ui_eslint_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs)$'
 
-# CI's lint job (test-linting.yml) only inspects litellm/, so a tests-only or
-# scripts-only commit can't turn it red; scope the trigger there to skip the slow
-# make lint when it couldn't catch anything.
+# CI's lint job (test-linting.yml) is make lint: litellm/ checks plus two test-tree
+# steps (ruff over ruff-tests.toml and the test-quality budget). Trigger the slow
+# make lint on litellm/ files only; a tests-only commit runs just those two steps.
 litellm_py_files=$(scope_match "$litellm_py_pattern")
 e2e_py_files=$(scope_match "$e2e_py_pattern")
+tests_py_changed=$(scope_match "$tests_py_pattern")
+tests_py_files=$(printf '%s\n' "$tests_py_changed" | existing_files)
 # ruff format (and CI's format step) skip enterprise; the rest of make lint covers it.
 fmt_files=$(printf '%s\n' "$litellm_py_files" | grep -v '^litellm/enterprise/' | existing_files)
 # check-ui-api-types.yml triggers on any file under litellm/proxy or litellm/types
@@ -136,6 +141,7 @@ if [ -n "$staged" ]; then
     }
     warn_skipped "Python lint (make lint)" "$litellm_py_pattern" "$litellm_py_files"
     warn_skipped "tests/e2e checks (basedpyright + raw HTTP client ban)" "$e2e_py_pattern" "$e2e_py_files"
+    warn_skipped "test-tree lint (ruff-tests.toml + test-quality budget)" "$tests_py_pattern" "$tests_py_changed"
     warn_skipped "dashboard lint (prettier + eslint + lint budgets)" "$ui_prettier_pattern" "$ui_prettier_changed"
     warn_skipped "dashboard API-type sync (npm run gen:api)" "$spec_pattern" "$spec_files"
 fi
@@ -225,6 +231,16 @@ if [ -n "$e2e_py_files" ]; then
         || { echo "✗ Raw HTTP client import in tests/e2e. Route the call through tests/e2e/e2e_http.py, then re-run make check." >&2; status=1; }
 fi
 
+if [ -n "$tests_py_changed" ] && [ -z "$litellm_py_files" ]; then
+    if [ -n "$tests_py_files" ]; then
+        echo "check: linting the test tree (ruff check --config ruff-tests.toml, scoped tests files)"
+        printf '%s\n' "$tests_py_files" | xargs uv run --no-sync ruff check --config ruff-tests.toml \
+            || { echo "✗ Test-tree ruff failed. Fix the errors above, then re-run make check." >&2; status=1; }
+    fi
+    echo "check: checking the test-quality budget (make lint-test-quality)"
+    make lint-test-quality || { echo "✗ Test-quality budget failed. Fix the errors above, then re-run make check." >&2; status=1; }
+fi
+
 dashboard_checks() {
     echo "check: linting dashboard (prettier + eslint + lint budgets)"
     if [ ! -d ui/litellm-dashboard/node_modules ]; then
@@ -313,10 +329,11 @@ summary_item() {
 echo "check: summary"
 summary_item "Python lint (make lint)" "$litellm_py_files" "no litellm/ Python files in scope"
 summary_item "tests/e2e checks (basedpyright + raw HTTP client ban)" "$e2e_py_files" "no tests/e2e Python files in scope"
+summary_item "test-tree lint (ruff-tests.toml + test-quality budget)" "$tests_py_changed" "no tests/ Python files in scope"
 summary_item "dashboard lint (prettier + eslint + lint budgets)" "$ui_prettier_changed$ui_eslint_changed" "no dashboard files in scope"
 summary_item "dashboard API-type sync (npm run gen:api)" "$spec_files" "no litellm/proxy, litellm/types, or generator files in scope"
 
-if [ -z "$litellm_py_files$e2e_py_files$ui_prettier_changed$ui_eslint_changed$spec_files" ]; then
+if [ -z "$litellm_py_files$e2e_py_files$tests_py_changed$ui_prettier_changed$ui_eslint_changed$spec_files" ]; then
     echo "check: NOTE - no gating lint check matches the files in scope, so nothing ran:" >&2
     printf '%s\n' "$scope" | sed 's/^/    /' >&2
     echo "  A pass here is a no-op, not a lint verdict." >&2
diff --git a/tests/test_litellm/test_pre_commit_lint.py b/tests/test_litellm/test_pre_commit_lint.py
index aa2260e89ea..22ae0e1bfaa 100644
--- a/tests/test_litellm/test_pre_commit_lint.py
+++ b/tests/test_litellm/test_pre_commit_lint.py
@@ -40,6 +40,9 @@ case "$*" in
             sleep 60
         fi
         ;;
+    lint-test-quality)
+        [ "${STUB_FAIL:-}" = "test-quality" ] && exit 1
+        ;;
 esac
 exit 0
 """
@@ -69,6 +72,10 @@ case "$*" in
     *orjson*)
         [ -n "${STUB_BARRIER_DIR:-}" ] && barrier_sync genapi "python dashboard"
         ;;
+    "run --no-sync ruff check --config ruff-tests.toml"*)
+        [ -n "${STUB_ARGS_DIR:-}" ] && echo "$*" >> "$STUB_ARGS_DIR/ruff_tests.args"
+        [ "${STUB_FAIL:-}" = "tests-ruff" ] && exit 1
+        ;;
 esac
 exit 0
 """
@@ -153,6 +160,13 @@ def _set_base_ref(repo: Path) -> None:
     )
 
 
+def _stage_file(repo: Path, relative: str, body: str) -> None:
+    path = repo / relative
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(body)
+    subprocess.run(["git", "add", relative], cwd=repo, check=True)
+
+
 def test_nothing_staged_scopes_to_working_tree_diff_and_runs_checks(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _commit_all(repo, "base")
@@ -405,6 +419,7 @@ def test_run_ends_with_a_summary_of_ran_and_skipped_blocks(tmp_path: Path) -> No
     assert "ran:     dashboard lint (prettier + eslint + lint budgets)" in proc.stdout
     assert "ran:     dashboard API-type sync (npm run gen:api)" in proc.stdout
     assert "skipped: tests/e2e checks (basedpyright + raw HTTP client ban) (no tests/e2e Python files in scope)" in proc.stdout
+    assert "skipped: test-tree lint (ruff-tests.toml + test-quality budget) (no tests/ Python files in scope)" in proc.stdout
     assert "check: PASS" in proc.stdout
     assert "check: FAIL" not in proc.stdout
 
@@ -412,20 +427,93 @@ def test_run_ends_with_a_summary_of_ran_and_skipped_blocks(tmp_path: Path) -> No
 def test_staged_files_matching_no_check_print_an_explicit_noop_note_and_nonempty_log(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _commit_all(repo, "base")
-    tests_dir = repo / "tests" / "test_litellm"
-    tests_dir.mkdir(parents=True)
-    (tests_dir / "test_x.py").write_text("def test_x() -> None: ...\n")
-    subprocess.run(["git", "add", "tests"], cwd=repo, check=True)
+    _stage_file(repo, "scripts/tool.py", "def main() -> None: ...\n")
     proc = _run(repo, bin_dir, {})
     assert proc.returncode == 0, proc.stdout + proc.stderr
     assert "no gating lint check matches the files in scope, so nothing ran" in proc.stdout
-    assert "tests/test_litellm/test_x.py" in proc.stdout
+    assert "scripts/tool.py" in proc.stdout
     assert "a no-op, not a lint verdict" in proc.stdout
     assert "check: PASS" in proc.stdout
     assert "linting Python" not in proc.stdout
     log = (repo / ".git" / "pre_commit_lint.log").read_text()
     assert "check: summary" in log
     assert "skipped: Python lint (make lint) (no litellm/ Python files in scope)" in log
+    assert "skipped: test-tree lint (ruff-tests.toml + test-quality budget) (no tests/ Python files in scope)" in log
+
+
+def test_tests_only_change_runs_test_tree_ruff_on_staged_files_and_the_quality_gate(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    _stage_file(repo, "tests/test_b.py", "def test_b() -> None: ...\n")
+    _stage_file(repo, "tests/fixtures/data.json", "{}\n")
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    ruff_args = (args_dir / "ruff_tests.args").read_text().splitlines()
+    assert ruff_args == ["run --no-sync ruff check --config ruff-tests.toml tests/test_a.py tests/test_b.py"]
+    assert "ran:     test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
+    assert "no gating lint check matches" not in proc.stdout
+    assert "linting Python" not in proc.stdout
+    assert "check: PASS" in proc.stdout
+
+
+@pytest.mark.parametrize(
+    ("fail", "message"),
+    [
+        ("tests-ruff", "Test-tree ruff failed"),
+        ("test-quality", "Test-quality budget failed"),
+    ],
+)
+def test_a_failing_test_tree_check_fails_a_tests_only_run(tmp_path: Path, fail: str, message: str) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _commit_all(repo, "base")
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    proc = _run(repo, bin_dir, {"STUB_FAIL": fail})
+    assert proc.returncode == 1
+    assert message in proc.stdout + proc.stderr
+    assert "check: FAIL" in proc.stdout
+
+
+def test_tests_changed_alongside_litellm_files_defer_to_make_lint(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    _stage_file(repo, "litellm/foo.py", "x = 2\n")
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir), "STUB_FAIL": "test-quality"})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert "linting Python" in proc.stdout
+    assert not (args_dir / "ruff_tests.args").exists()
+    assert "ran:     test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
+
+
+def test_deleted_test_file_still_runs_the_quality_gate_without_feeding_ruff_the_missing_file(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    subprocess.run(["git", "rm", "-q", "tests/test_a.py"], cwd=repo, check=True)
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir), "STUB_FAIL": "test-quality"})
+    assert proc.returncode == 1
+    assert "Test-quality budget failed" in proc.stdout + proc.stderr
+    assert not (args_dir / "ruff_tests.args").exists()
+
+
+def test_partial_staging_warns_when_test_files_are_left_unstaged(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    _commit_all(repo, "base")
+    _stage_file(repo, "notes.md", "hi\n")
+    (repo / "tests" / "test_a.py").write_text("def test_a() -> None:\n    assert True\n")
+    proc = _run(repo, bin_dir, {})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert "SKIPPED test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
+    assert "tests/test_a.py" in proc.stdout
+    assert "make lint-test-quality" not in proc.stdout
 
 
 def test_run_queues_through_the_machine_wide_gate_slot_lock(tmp_path: Path) -> None:

From d22962248c1c659279e7389e6c2e1d1b640b400a Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 00:07:25 -0700
Subject: [PATCH 30/79] fix(check): run CI's whole-tree test ruff and widen the
 test-tree trigger

The scoped xargs list missed a ruff-tests.toml rule change and skipped ruff on
deletions, so the block now runs test-linting.yml's exact command over tests/.
ruff-tests.toml, test-quality-budget.json, and scripts/check_test_quality.py
trigger the block too, and it sits after the background launches so the
dashboard and gen:api jobs overlap it.
---
 scripts/pre_commit_lint.sh                 |  36 ++++----
 tests/test_litellm/test_pre_commit_lint.py | 102 ++++++++++++++++-----
 2 files changed, 96 insertions(+), 42 deletions(-)

diff --git a/scripts/pre_commit_lint.sh b/scripts/pre_commit_lint.sh
index ad1b4e793ab..ab84d2d518a 100755
--- a/scripts/pre_commit_lint.sh
+++ b/scripts/pre_commit_lint.sh
@@ -12,7 +12,8 @@
 #   - litellm/ Python  -> `make lint` (test-linting.yml's lint job)
 #   - tests/e2e Python -> `make lint-e2e-basedpyright` (test-linting.yml's e2e type-check step)
 #                         + raw HTTP client ban (test-code-quality.yml's check_e2e_no_raw_requests)
-#   - tests/ Python    -> ruff over ruff-tests.toml + `make lint-test-quality` (test-linting.yml's
+#   - tests/ Python, ruff-tests.toml, test-quality-budget.json, scripts/check_test_quality.py
+#                      -> ruff over ruff-tests.toml + `make lint-test-quality` (test-linting.yml's
 #                         test-tree ruff and test-quality budget steps)
 #   - dashboard        -> prettier + eslint + lint budgets (test-litellm-ui-build.yml's frontend-lint)
 #   - proxy/types      -> regenerate the lazy OpenAPI snapshot and dashboard API types, fail on drift (check-ui-api-types.yml)
@@ -90,18 +91,17 @@ existing_files() {
 
 litellm_py_pattern='^litellm/.*\.py$'
 e2e_py_pattern='^tests/e2e/.*\.py$'
-tests_py_pattern='^tests/.*\.py$'
+test_tree_pattern='^(tests/.*\.py|ruff-tests\.toml|test-quality-budget\.json|scripts/check_test_quality\.py)$'
 spec_pattern='^(litellm/(proxy|types)/.*|ui/litellm-dashboard/(scripts/gen-api-types\.mjs|package\.json|package-lock\.json|src/lib/http/schema\.d\.ts))$'
 ui_prettier_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs|json|css|scss|md|mdx|yml|yaml|html)$'
 ui_eslint_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs)$'
 
 # CI's lint job (test-linting.yml) is make lint: litellm/ checks plus two test-tree
 # steps (ruff over ruff-tests.toml and the test-quality budget). Trigger the slow
-# make lint on litellm/ files only; a tests-only commit runs just those two steps.
+# make lint on litellm/ files only; without them, the test-tree steps run on their own below.
 litellm_py_files=$(scope_match "$litellm_py_pattern")
 e2e_py_files=$(scope_match "$e2e_py_pattern")
-tests_py_changed=$(scope_match "$tests_py_pattern")
-tests_py_files=$(printf '%s\n' "$tests_py_changed" | existing_files)
+test_tree_files=$(scope_match "$test_tree_pattern")
 # ruff format (and CI's format step) skip enterprise; the rest of make lint covers it.
 fmt_files=$(printf '%s\n' "$litellm_py_files" | grep -v '^litellm/enterprise/' | existing_files)
 # check-ui-api-types.yml triggers on any file under litellm/proxy or litellm/types
@@ -141,7 +141,7 @@ if [ -n "$staged" ]; then
     }
     warn_skipped "Python lint (make lint)" "$litellm_py_pattern" "$litellm_py_files"
     warn_skipped "tests/e2e checks (basedpyright + raw HTTP client ban)" "$e2e_py_pattern" "$e2e_py_files"
-    warn_skipped "test-tree lint (ruff-tests.toml + test-quality budget)" "$tests_py_pattern" "$tests_py_changed"
+    warn_skipped "test-tree lint (ruff-tests.toml + test-quality budget)" "$test_tree_pattern" "$test_tree_files"
     warn_skipped "dashboard lint (prettier + eslint + lint budgets)" "$ui_prettier_pattern" "$ui_prettier_changed"
     warn_skipped "dashboard API-type sync (npm run gen:api)" "$spec_pattern" "$spec_files"
 fi
@@ -231,16 +231,6 @@ if [ -n "$e2e_py_files" ]; then
         || { echo "✗ Raw HTTP client import in tests/e2e. Route the call through tests/e2e/e2e_http.py, then re-run make check." >&2; status=1; }
 fi
 
-if [ -n "$tests_py_changed" ] && [ -z "$litellm_py_files" ]; then
-    if [ -n "$tests_py_files" ]; then
-        echo "check: linting the test tree (ruff check --config ruff-tests.toml, scoped tests files)"
-        printf '%s\n' "$tests_py_files" | xargs uv run --no-sync ruff check --config ruff-tests.toml \
-            || { echo "✗ Test-tree ruff failed. Fix the errors above, then re-run make check." >&2; status=1; }
-    fi
-    echo "check: checking the test-quality budget (make lint-test-quality)"
-    make lint-test-quality || { echo "✗ Test-quality budget failed. Fix the errors above, then re-run make check." >&2; status=1; }
-fi
-
 dashboard_checks() {
     echo "check: linting dashboard (prettier + eslint + lint budgets)"
     if [ ! -d ui/litellm-dashboard/node_modules ]; then
@@ -304,6 +294,15 @@ if [ -n "$spec_files" ]; then
     set +m
 fi
 
+if [ -n "$test_tree_files" ] && [ -z "$litellm_py_files" ]; then
+    echo "check: linting the test tree (ruff check --config ruff-tests.toml tests)"
+    uv run --no-sync ruff check --config ruff-tests.toml tests \
+        || { echo "✗ Test-tree ruff failed. Fix the errors above, then re-run make check." >&2; status=1; }
+    echo "check: checking the test-quality budget (make lint-test-quality)"
+    make lint-test-quality \
+        || { echo "✗ Test-quality budget failed. Fix the errors above, then re-run make check." >&2; status=1; }
+fi
+
 if [ -n "${python_pid:-}" ]; then
     wait "$python_pid" || status=1
     cat "$python_log"; rm -f "$python_log"
@@ -329,11 +328,12 @@ summary_item() {
 echo "check: summary"
 summary_item "Python lint (make lint)" "$litellm_py_files" "no litellm/ Python files in scope"
 summary_item "tests/e2e checks (basedpyright + raw HTTP client ban)" "$e2e_py_files" "no tests/e2e Python files in scope"
-summary_item "test-tree lint (ruff-tests.toml + test-quality budget)" "$tests_py_changed" "no tests/ Python files in scope"
+summary_item "test-tree lint (ruff-tests.toml + test-quality budget)" "$test_tree_files" \
+    "no tests/ Python files or test-tree lint inputs in scope"
 summary_item "dashboard lint (prettier + eslint + lint budgets)" "$ui_prettier_changed$ui_eslint_changed" "no dashboard files in scope"
 summary_item "dashboard API-type sync (npm run gen:api)" "$spec_files" "no litellm/proxy, litellm/types, or generator files in scope"
 
-if [ -z "$litellm_py_files$e2e_py_files$tests_py_changed$ui_prettier_changed$ui_eslint_changed$spec_files" ]; then
+if [ -z "$litellm_py_files$e2e_py_files$test_tree_files$ui_prettier_changed$ui_eslint_changed$spec_files" ]; then
     echo "check: NOTE - no gating lint check matches the files in scope, so nothing ran:" >&2
     printf '%s\n' "$scope" | sed 's/^/    /' >&2
     echo "  A pass here is a no-op, not a lint verdict." >&2
diff --git a/tests/test_litellm/test_pre_commit_lint.py b/tests/test_litellm/test_pre_commit_lint.py
index 22ae0e1bfaa..e17abf2fa9f 100644
--- a/tests/test_litellm/test_pre_commit_lint.py
+++ b/tests/test_litellm/test_pre_commit_lint.py
@@ -11,6 +11,12 @@ import pytest
 
 ROOT = Path(__file__).resolve().parents[2]
 SCRIPT = ROOT / "scripts" / "pre_commit_lint.sh"
+WHOLE_TREE_RUFF = "run --no-sync ruff check --config ruff-tests.toml tests"
+TEST_TREE_RAN = "ran:     test-tree lint (ruff-tests.toml + test-quality budget)"
+TEST_TREE_SKIPPED = (
+    "skipped: test-tree lint (ruff-tests.toml + test-quality budget) "
+    "(no tests/ Python files or test-tree lint inputs in scope)"
+)
 
 BARRIER_HELPER = """barrier_sync() {
     touch "$STUB_BARRIER_DIR/$1.started"
@@ -30,6 +36,7 @@ BARRIER_HELPER = """barrier_sync() {
 
 MAKE_STUB = """#!/bin/sh
 . "$STUB_BIN/barrier.sh"
+[ -n "${STUB_ARGS_DIR:-}" ] && echo "$*" >> "$STUB_ARGS_DIR/make.args"
 case "$*" in
     lint)
         [ "${STUB_FAIL:-}" = "make-lint" ] && exit 1
@@ -419,7 +426,7 @@ def test_run_ends_with_a_summary_of_ran_and_skipped_blocks(tmp_path: Path) -> No
     assert "ran:     dashboard lint (prettier + eslint + lint budgets)" in proc.stdout
     assert "ran:     dashboard API-type sync (npm run gen:api)" in proc.stdout
     assert "skipped: tests/e2e checks (basedpyright + raw HTTP client ban) (no tests/e2e Python files in scope)" in proc.stdout
-    assert "skipped: test-tree lint (ruff-tests.toml + test-quality budget) (no tests/ Python files in scope)" in proc.stdout
+    assert TEST_TREE_SKIPPED in proc.stdout
     assert "check: PASS" in proc.stdout
     assert "check: FAIL" not in proc.stdout
 
@@ -438,41 +445,83 @@ def test_staged_files_matching_no_check_print_an_explicit_noop_note_and_nonempty
     log = (repo / ".git" / "pre_commit_lint.log").read_text()
     assert "check: summary" in log
     assert "skipped: Python lint (make lint) (no litellm/ Python files in scope)" in log
-    assert "skipped: test-tree lint (ruff-tests.toml + test-quality budget) (no tests/ Python files in scope)" in log
+    assert TEST_TREE_SKIPPED in log
 
 
-def test_tests_only_change_runs_test_tree_ruff_on_staged_files_and_the_quality_gate(tmp_path: Path) -> None:
+def _recorded(args_dir: Path, name: str) -> list[str]:
+    path = args_dir / name
+    return path.read_text().splitlines() if path.exists() else []
+
+
+def test_tests_only_change_runs_the_whole_test_tree_ruff_and_the_quality_gate(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _commit_all(repo, "base")
     args_dir = tmp_path / "args"
     args_dir.mkdir()
     _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
-    _stage_file(repo, "tests/test_b.py", "def test_b() -> None: ...\n")
     _stage_file(repo, "tests/fixtures/data.json", "{}\n")
     proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
     assert proc.returncode == 0, proc.stdout + proc.stderr
-    ruff_args = (args_dir / "ruff_tests.args").read_text().splitlines()
-    assert ruff_args == ["run --no-sync ruff check --config ruff-tests.toml tests/test_a.py tests/test_b.py"]
-    assert "ran:     test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
+    assert _recorded(args_dir, "ruff_tests.args") == [WHOLE_TREE_RUFF]
+    assert _recorded(args_dir, "make.args") == ["lint-test-quality"]
+    assert TEST_TREE_RAN in proc.stdout
     assert "no gating lint check matches" not in proc.stdout
     assert "linting Python" not in proc.stdout
     assert "check: PASS" in proc.stdout
 
 
 @pytest.mark.parametrize(
-    ("fail", "message"),
-    [
-        ("tests-ruff", "Test-tree ruff failed"),
-        ("test-quality", "Test-quality budget failed"),
-    ],
+    "changed",
+    ["ruff-tests.toml", "test-quality-budget.json", "scripts/check_test_quality.py", "tests/e2e/test_x.py"],
 )
-def test_a_failing_test_tree_check_fails_a_tests_only_run(tmp_path: Path, fail: str, message: str) -> None:
+def test_test_tree_lint_inputs_trigger_the_test_tree_checks(tmp_path: Path, changed: str) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    _stage_file(repo, changed, "x = 1\n")
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert _recorded(args_dir, "ruff_tests.args") == [WHOLE_TREE_RUFF]
+    assert "lint-test-quality" in _recorded(args_dir, "make.args")
+    assert TEST_TREE_RAN in proc.stdout
+
+
+def test_nothing_staged_tests_only_working_tree_change_runs_the_test_tree_checks(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    _commit_all(repo, "base")
+    _set_base_ref(repo)
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    (repo / "tests" / "test_a.py").write_text("def test_a() -> None:\n    assert True\n")
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert "nothing staged; scoping to the working tree's diff" in proc.stdout
+    assert _recorded(args_dir, "ruff_tests.args") == [WHOLE_TREE_RUFF]
+    assert _recorded(args_dir, "make.args") == ["lint-test-quality"]
+
+
+def test_a_failing_test_tree_ruff_fails_the_run_and_still_runs_the_quality_gate(tmp_path: Path) -> None:
+    repo, bin_dir = _sandbox(tmp_path)
+    _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
+    _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir), "STUB_FAIL": "tests-ruff"})
+    assert proc.returncode == 1
+    assert "Test-tree ruff failed" in proc.stdout + proc.stderr
+    assert "check: FAIL" in proc.stdout
+    assert _recorded(args_dir, "make.args") == ["lint-test-quality"]
+
+
+def test_a_failing_quality_gate_fails_a_tests_only_run(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _commit_all(repo, "base")
     _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
-    proc = _run(repo, bin_dir, {"STUB_FAIL": fail})
+    proc = _run(repo, bin_dir, {"STUB_FAIL": "test-quality"})
     assert proc.returncode == 1
-    assert message in proc.stdout + proc.stderr
+    assert "Test-quality budget failed" in proc.stdout + proc.stderr
     assert "check: FAIL" in proc.stdout
 
 
@@ -486,34 +535,39 @@ def test_tests_changed_alongside_litellm_files_defer_to_make_lint(tmp_path: Path
     proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir), "STUB_FAIL": "test-quality"})
     assert proc.returncode == 0, proc.stdout + proc.stderr
     assert "linting Python" in proc.stdout
-    assert not (args_dir / "ruff_tests.args").exists()
-    assert "ran:     test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
+    assert _recorded(args_dir, "ruff_tests.args") == []
+    assert _recorded(args_dir, "make.args") == ["lint"]
+    assert TEST_TREE_RAN in proc.stdout
 
 
-def test_deleted_test_file_still_runs_the_quality_gate_without_feeding_ruff_the_missing_file(tmp_path: Path) -> None:
+def test_deleted_test_file_still_runs_the_test_tree_checks(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
     _commit_all(repo, "base")
     args_dir = tmp_path / "args"
     args_dir.mkdir()
     subprocess.run(["git", "rm", "-q", "tests/test_a.py"], cwd=repo, check=True)
-    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir), "STUB_FAIL": "test-quality"})
-    assert proc.returncode == 1
-    assert "Test-quality budget failed" in proc.stdout + proc.stderr
-    assert not (args_dir / "ruff_tests.args").exists()
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
+    assert proc.returncode == 0, proc.stdout + proc.stderr
+    assert _recorded(args_dir, "ruff_tests.args") == [WHOLE_TREE_RUFF]
+    assert _recorded(args_dir, "make.args") == ["lint-test-quality"]
+    assert TEST_TREE_RAN in proc.stdout
 
 
 def test_partial_staging_warns_when_test_files_are_left_unstaged(tmp_path: Path) -> None:
     repo, bin_dir = _sandbox(tmp_path)
     _stage_file(repo, "tests/test_a.py", "def test_a() -> None: ...\n")
     _commit_all(repo, "base")
+    args_dir = tmp_path / "args"
+    args_dir.mkdir()
     _stage_file(repo, "notes.md", "hi\n")
     (repo / "tests" / "test_a.py").write_text("def test_a() -> None:\n    assert True\n")
-    proc = _run(repo, bin_dir, {})
+    proc = _run(repo, bin_dir, {"STUB_ARGS_DIR": str(args_dir)})
     assert proc.returncode == 0, proc.stdout + proc.stderr
     assert "SKIPPED test-tree lint (ruff-tests.toml + test-quality budget)" in proc.stdout
     assert "tests/test_a.py" in proc.stdout
-    assert "make lint-test-quality" not in proc.stdout
+    assert _recorded(args_dir, "ruff_tests.args") == []
+    assert _recorded(args_dir, "make.args") == []
 
 
 def test_run_queues_through_the_machine_wide_gate_slot_lock(tmp_path: Path) -> None:

From 36b0d80d3a1459f116f07e45b205e953e21c8978 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 00:09:00 -0700
Subject: [PATCH 31/79] test(cost): pin the nested-reasoning helper's clamps

The strip in text_tokens_without_nested_reasoning is capped at the reasoning
share, the reported text, and the over-sum past completion_tokens. Dropping the
caps to a bare over-sum passed every existing test, so this pins each cap with
a parametrized helper test plus one billing test where text over-reports past
the reasoning share and only the nested share may be netted out
---
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 23 ++++++++++++++
 tests/test_litellm/types/test_types_utils.py  | 31 ++++++++++++++++++-
 2 files changed, 53 insertions(+), 1 deletion(-)

diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index e8eae73df3f..c9fe53651cd 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -4821,6 +4821,29 @@ def test_generic_cost_per_token_keeps_billing_reasoning_reported_beside_text_tok
     assert completion_cost == pytest.approx(44 * info["output_cost_per_token"])
 
 
+def test_generic_cost_per_token_strips_only_the_reasoning_share_when_text_over_reports(
+    _local_model_cost_map: None,
+) -> None:
+    """Text over-reported past the reasoning share keeps its extra tokens billed; only the nested reasoning is netted out."""
+
+    model = "gpt-realtime-2.1-mini"
+    usage = Usage(
+        prompt_tokens=120,
+        completion_tokens=100,
+        total_tokens=220,
+        completion_tokens_details=CompletionTokensDetailsWrapper(text_tokens=100, audio_tokens=70, reasoning_tokens=10),
+    )
+
+    _, completion_cost = generic_cost_per_token(model=model, usage=usage, custom_llm_provider="openai")
+    breakdown = get_token_type_cost_breakdown(model=model, custom_llm_provider="openai", usage=usage)
+
+    info = litellm.get_model_info(model=model, custom_llm_provider="openai")
+    assert breakdown.reasoning_cost == pytest.approx(10 * info["output_cost_per_token"])
+    assert completion_cost == pytest.approx(
+        100 * info["output_cost_per_token"] + 70 * info["output_cost_per_audio_token"]
+    )
+
+
 def test_generic_cost_per_token_bills_nested_reasoning_once_beside_audio_output(_local_model_cost_map: None) -> None:
     """Audio-output realtime usage nests reasoning inside text_tokens next to audio_tokens; text is billed net of it."""
 
diff --git a/tests/test_litellm/types/test_types_utils.py b/tests/test_litellm/types/test_types_utils.py
index 554604ab200..5f44ba1773e 100644
--- a/tests/test_litellm/types/test_types_utils.py
+++ b/tests/test_litellm/types/test_types_utils.py
@@ -3,7 +3,7 @@ from typing import Final
 import pytest
 
 
-from litellm.types.utils import HiddenParams, all_litellm_params
+from litellm.types.utils import HiddenParams, all_litellm_params, text_tokens_without_nested_reasoning
 
 
 def test_rust_is_a_known_litellm_param():
@@ -768,3 +768,32 @@ def test_image_response_keeps_background():
     response = ImageResponse(created=1, data=[{"b64_json": "aGk="}], background="transparent", output_format="png")
     assert response.background == "transparent"
     assert response.model_dump()["background"] == "transparent"
+
+
+@pytest.mark.parametrize(
+    ("completion_tokens", "text_tokens", "reasoning_tokens", "other_modality_tokens", "expected_text_tokens"),
+    (
+        pytest.param(50, 30, 20, 0, 30, id="details_sum_to_completion_is_a_no_op"),
+        pytest.param(34, 30, 24, 0, 10, id="strip_is_capped_at_the_over_sum"),
+        pytest.param(100, 100, 10, 70, 90, id="only_the_reasoning_share_is_stripped_when_text_over_reports_further"),
+        pytest.param(10, 5, 20, 0, 0, id="text_never_goes_negative_when_reasoning_exceeds_it"),
+    ),
+)
+def test_text_tokens_without_nested_reasoning_clamps(
+    completion_tokens: int,
+    text_tokens: int,
+    reasoning_tokens: int,
+    other_modality_tokens: int,
+    expected_text_tokens: int,
+) -> None:
+    """The strip never exceeds the reasoning share, the reported text, or the over-sum past completion_tokens."""
+
+    assert (
+        text_tokens_without_nested_reasoning(
+            completion_tokens=completion_tokens,
+            text_tokens=text_tokens,
+            reasoning_tokens=reasoning_tokens,
+            other_modality_tokens=other_modality_tokens,
+        )
+        == expected_text_tokens
+    )

From 827554954d305b7c8c24f524c961cf7a27e3a029 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 00:16:14 -0700
Subject: [PATCH 32/79] fix(image_handling): answer every SSRF rejection with
 one message so error text cannot probe internal hostnames

---
 .../prompt_templates/image_handling.py        | 10 ++---
 litellm/litellm_core_utils/url_utils.py       |  8 +---
 .../litellm_core_utils/test_image_handling.py | 41 ++++++++-----------
 .../litellm_core_utils/test_url_utils.py      | 19 +--------
 4 files changed, 24 insertions(+), 54 deletions(-)

diff --git a/litellm/litellm_core_utils/prompt_templates/image_handling.py b/litellm/litellm_core_utils/prompt_templates/image_handling.py
index 99efbd13320..24f3b8bca7f 100644
--- a/litellm/litellm_core_utils/prompt_templates/image_handling.py
+++ b/litellm/litellm_core_utils/prompt_templates/image_handling.py
@@ -15,7 +15,7 @@ import litellm
 from litellm import verbose_logger
 from litellm.caching.caching import InMemoryCache
 from litellm.constants import MAX_IMAGE_URL_DOWNLOAD_SIZE_MB
-from litellm.litellm_core_utils.url_utils import HostResolutionError, SSRFError, async_safe_get, safe_get
+from litellm.litellm_core_utils.url_utils import SSRFError, async_safe_get, safe_get
 from litellm.types.llms.openai import AllMessageValues
 
 MAX_IMGS_IN_MEMORY: Final = 10
@@ -80,13 +80,9 @@ def _process_image_response(response: Response, url: str) -> str:
 
 def _rejected_image_fetch(url: str, verdict: SSRFError) -> "litellm.ImageFetchError":
     verbose_logger.warning("Image fetch of %s rejected before any request went out: %s", url, verdict)
-    if isinstance(verdict, HostResolutionError):
-        return litellm.ImageFetchError(
-            f"Error: Unable to fetch image from URL. The image host could not be resolved. url={url}"
-        )
     return litellm.ImageFetchError(
-        "Error: Unable to fetch image from URL. The proxy's URL policy rejected this host; "
-        f"an admin can allow it with `user_url_allowed_hosts` in general_settings. url={url}"
+        "Error: Unable to fetch image from URL. The proxy could not resolve this host or its URL policy rejected it; "
+        f"an admin can check the proxy log and `user_url_allowed_hosts` in general_settings. url={url}"
     )
 
 
diff --git a/litellm/litellm_core_utils/url_utils.py b/litellm/litellm_core_utils/url_utils.py
index f4388d5fd12..fa070a648f5 100644
--- a/litellm/litellm_core_utils/url_utils.py
+++ b/litellm/litellm_core_utils/url_utils.py
@@ -93,10 +93,6 @@ class SSRFError(ValueError):
     """Raised when a URL targets a blocked network."""
 
 
-class HostResolutionError(SSRFError):
-    pass
-
-
 def encode_url_path_segment(value: object, *, field_name: str = "path parameter") -> str:
     """Percent-encode one user-controlled URL path segment.
 
@@ -328,10 +324,10 @@ def validate_url(url: str) -> tuple[str, str]:
     try:
         addrinfo: Final = socket.getaddrinfo(hostname, effective_port, proto=socket.IPPROTO_TCP)
     except socket.gaierror as e:
-        raise HostResolutionError(f"DNS resolution failed for '{hostname}': {e}")
+        raise SSRFError(f"DNS resolution failed for '{hostname}': {e}")
 
     if not addrinfo:
-        raise HostResolutionError(f"No addresses found for '{hostname}'")
+        raise SSRFError(f"No addresses found for '{hostname}'")
 
     if not is_allowlisted:
         for addrinfo_entry in addrinfo:
diff --git a/tests/test_litellm/litellm_core_utils/test_image_handling.py b/tests/test_litellm/litellm_core_utils/test_image_handling.py
index 3a909298aba..8fa4bd6c14d 100644
--- a/tests/test_litellm/litellm_core_utils/test_image_handling.py
+++ b/tests/test_litellm/litellm_core_utils/test_image_handling.py
@@ -17,7 +17,7 @@ from litellm.litellm_core_utils.prompt_templates.image_handling import (
     async_inline_remote_media,
     convert_url_to_base64,
 )
-from litellm.litellm_core_utils.url_utils import HostResolutionError, SSRFError
+from litellm.litellm_core_utils.url_utils import SSRFError
 
 
 @pytest.fixture(autouse=True)
@@ -425,32 +425,25 @@ async def test_async_inline_remote_media_cancels_the_other_fetches_when_one_fail
 
 
 _SSRF_VERDICTS = (
-    (
-        SSRFError(
-            "URL targets a blocked address (10.0.0.8). If this is a legitimate internal service, "
-            "add the host to `user_url_allowed_hosts` in general_settings."
-        ),
-        "The proxy's URL policy rejected this host; an admin can allow it with `user_url_allowed_hosts`",
+    SSRFError(
+        "URL targets a blocked address (10.0.0.8). If this is a legitimate internal service, "
+        "add the host to `user_url_allowed_hosts` in general_settings."
     ),
-    (
-        HostResolutionError(
-            "DNS resolution failed for 'internal.example': [Errno 8] nodename nor servname provided, or not known"
-        ),
-        "The image host could not be resolved",
-    ),
-    (HostResolutionError("No addresses found for 'internal.example'"), "The image host could not be resolved"),
+    SSRFError("DNS resolution failed for 'internal.example': [Errno 8] nodename nor servname provided, or not known"),
+    SSRFError("No addresses found for 'internal.example'"),
 )
 
 
 def _assert_verdict_free_messages(messages, url):
-    for message, (verdict, expected_guidance) in zip(messages, _SSRF_VERDICTS, strict=True):
-        assert expected_guidance in message
-        assert url in message
-        assert "10.0.0.8" not in message
-        assert "DNS" not in message
-        assert "No addresses" not in message
-        if isinstance(verdict, HostResolutionError):
-            assert "user_url_allowed_hosts" not in message
+    assert len(messages) == len(_SSRF_VERDICTS)
+    assert len(set(messages)) == 1, "a caller must not be able to tell a blocked host from one that does not resolve"
+    message = messages[0]
+    assert "The proxy could not resolve this host or its URL policy rejected it" in message
+    assert "user_url_allowed_hosts" in message
+    assert url in message
+    assert "10.0.0.8" not in message
+    assert "DNS" not in message
+    assert "No addresses" not in message
 
 
 async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeypatch):
@@ -458,7 +451,7 @@ async def test_async_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_r
     messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    for verdict, _ in _SSRF_VERDICTS:
+    for verdict in _SSRF_VERDICTS:
 
         async def block(client, fetched_url, verdict=verdict, **kwargs):
             attempts.append(fetched_url)
@@ -478,7 +471,7 @@ def test_convert_url_to_base64_hides_the_ssrf_verdict_and_does_not_retry(monkeyp
     messages = []
     url = f"http://internal.example/{uuid.uuid4()}.png"
 
-    for verdict, _ in _SSRF_VERDICTS:
+    for verdict in _SSRF_VERDICTS:
 
         def block(client, fetched_url, verdict=verdict, **kwargs):
             attempts.append(fetched_url)
diff --git a/tests/test_litellm/litellm_core_utils/test_url_utils.py b/tests/test_litellm/litellm_core_utils/test_url_utils.py
index 151c7ceb5bd..fccdc1a2a0a 100644
--- a/tests/test_litellm/litellm_core_utils/test_url_utils.py
+++ b/tests/test_litellm/litellm_core_utils/test_url_utils.py
@@ -9,7 +9,6 @@ import pytest
 import litellm
 from litellm.litellm_core_utils import url_utils
 from litellm.litellm_core_utils.url_utils import (
-    HostResolutionError,
     SSRFError,
     _is_blocked_ip,
     assert_same_origin,
@@ -162,24 +161,10 @@ class TestValidateUrl:
         assert "/path" in rewritten
         assert "key=value" in rewritten
 
-    def test_dns_failure_raises_a_host_resolution_error(self, mock_dns_failure):
-        with pytest.raises(HostResolutionError, match="DNS resolution failed"):
+    def test_dns_failure_raises(self, mock_dns_failure):
+        with pytest.raises(SSRFError, match="DNS resolution failed"):
             validate_url("http://this-domain-does-not-exist-xyz123.invalid/test")
 
-    def test_empty_resolution_raises_a_host_resolution_error(self, monkeypatch):
-        monkeypatch.setattr(url_utils.socket, "getaddrinfo", lambda *args, **kwargs: [])
-        with pytest.raises(HostResolutionError, match="No addresses found"):
-            validate_url("http://this-domain-resolves-to-nothing.invalid/test")
-
-    def test_blocked_address_is_not_a_host_resolution_error(self, monkeypatch):
-        def fake(host, port, *a, **kw):
-            return [(socket.AF_INET, socket.SOCK_STREAM, 6, "", ("10.0.0.8", port or 80))]
-
-        monkeypatch.setattr(url_utils.socket, "getaddrinfo", fake)
-        with pytest.raises(SSRFError, match="blocked address") as raised:
-            validate_url("http://internal.example/test")
-        assert not isinstance(raised.value, HostResolutionError)
-
     def test_blocks_localhost_hostname(self, monkeypatch):
         def fake(host, port, *a, **kw):
             return [

From e3366dddf44c7450a856ed472c2966798413ef36 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 00:26:16 -0700
Subject: [PATCH 33/79] fix(check): trigger the test-tree checks on the gate
 script and drop the scope comment

---
 scripts/pre_commit_lint.sh                 | 8 +++-----
 tests/test_litellm/test_pre_commit_lint.py | 8 +++++++-
 2 files changed, 10 insertions(+), 6 deletions(-)

diff --git a/scripts/pre_commit_lint.sh b/scripts/pre_commit_lint.sh
index ab84d2d518a..f245803408c 100755
--- a/scripts/pre_commit_lint.sh
+++ b/scripts/pre_commit_lint.sh
@@ -12,7 +12,8 @@
 #   - litellm/ Python  -> `make lint` (test-linting.yml's lint job)
 #   - tests/e2e Python -> `make lint-e2e-basedpyright` (test-linting.yml's e2e type-check step)
 #                         + raw HTTP client ban (test-code-quality.yml's check_e2e_no_raw_requests)
-#   - tests/ Python, ruff-tests.toml, test-quality-budget.json, scripts/check_test_quality.py
+#   - tests/ Python, ruff-tests.toml, test-quality-budget.json, scripts/check_test_quality.py,
+#     scripts/test_quality_gate.py
 #                      -> ruff over ruff-tests.toml + `make lint-test-quality` (test-linting.yml's
 #                         test-tree ruff and test-quality budget steps)
 #   - dashboard        -> prettier + eslint + lint budgets (test-litellm-ui-build.yml's frontend-lint)
@@ -91,14 +92,11 @@ existing_files() {
 
 litellm_py_pattern='^litellm/.*\.py$'
 e2e_py_pattern='^tests/e2e/.*\.py$'
-test_tree_pattern='^(tests/.*\.py|ruff-tests\.toml|test-quality-budget\.json|scripts/check_test_quality\.py)$'
+test_tree_pattern='^(tests/.*\.py|ruff-tests\.toml|test-quality-budget\.json|scripts/(check_test_quality|test_quality_gate)\.py)$'
 spec_pattern='^(litellm/(proxy|types)/.*|ui/litellm-dashboard/(scripts/gen-api-types\.mjs|package\.json|package-lock\.json|src/lib/http/schema\.d\.ts))$'
 ui_prettier_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs|json|css|scss|md|mdx|yml|yaml|html)$'
 ui_eslint_pattern='^ui/litellm-dashboard/.*\.(js|jsx|ts|tsx|mjs|cjs)$'
 
-# CI's lint job (test-linting.yml) is make lint: litellm/ checks plus two test-tree
-# steps (ruff over ruff-tests.toml and the test-quality budget). Trigger the slow
-# make lint on litellm/ files only; without them, the test-tree steps run on their own below.
 litellm_py_files=$(scope_match "$litellm_py_pattern")
 e2e_py_files=$(scope_match "$e2e_py_pattern")
 test_tree_files=$(scope_match "$test_tree_pattern")
diff --git a/tests/test_litellm/test_pre_commit_lint.py b/tests/test_litellm/test_pre_commit_lint.py
index e17abf2fa9f..b84cb8aa657 100644
--- a/tests/test_litellm/test_pre_commit_lint.py
+++ b/tests/test_litellm/test_pre_commit_lint.py
@@ -472,7 +472,13 @@ def test_tests_only_change_runs_the_whole_test_tree_ruff_and_the_quality_gate(tm
 
 @pytest.mark.parametrize(
     "changed",
-    ["ruff-tests.toml", "test-quality-budget.json", "scripts/check_test_quality.py", "tests/e2e/test_x.py"],
+    [
+        "ruff-tests.toml",
+        "test-quality-budget.json",
+        "scripts/check_test_quality.py",
+        "scripts/test_quality_gate.py",
+        "tests/e2e/test_x.py",
+    ],
 )
 def test_test_tree_lint_inputs_trigger_the_test_tree_checks(tmp_path: Path, changed: str) -> None:
     repo, bin_dir = _sandbox(tmp_path)

From 2d2b5dabf27405fd413dfbfd71aef5a304775a75 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 01:44:16 -0700
Subject: [PATCH 34/79] fix(test-quality-gate): tear the base worktree down on
 SIGTERM and SIGHUP

---
 scripts/test_quality_gate.py                 | 22 ++++--
 tests/test_litellm/test_test_quality_gate.py | 77 ++++++++++++++++++++
 2 files changed, 92 insertions(+), 7 deletions(-)

diff --git a/scripts/test_quality_gate.py b/scripts/test_quality_gate.py
index 7d34b194f1c..4f79c9488b1 100644
--- a/scripts/test_quality_gate.py
+++ b/scripts/test_quality_gate.py
@@ -34,13 +34,14 @@ import argparse
 import json
 import re
 import shutil
+import signal
 import subprocess
 import sys
 import tempfile
 from collections import Counter
 from collections.abc import Mapping, Sequence
 from pathlib import Path
-from types import MappingProxyType
+from types import FrameType, MappingProxyType
 from typing import Final, NamedTuple
 
 REPO_ROOT: Final = Path(__file__).resolve().parent.parent
@@ -48,6 +49,7 @@ CHECKER: Final = REPO_ROOT / "scripts" / "check_test_quality.py"
 BUDGET_PATH: Final = REPO_ROOT / "test-quality-budget.json"
 TARGET: Final = "tests"
 DEFAULT_BASE: Final = "origin/litellm_internal_staging"
+TERMINATION_SIGNALS: Final = (signal.SIGTERM, signal.SIGHUP)
 
 _HUNK: Final = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,(\d+))? @@", re.MULTILINE)
 _FILE_HEADER: Final = re.compile(r"^\+\+\+ b/(.+)$", re.MULTILINE)
@@ -116,22 +118,28 @@ def count_by_rule(violations: Sequence[Violation]) -> Mapping[str, int]:
     return MappingProxyType(dict(Counter(v.code for v in violations)))
 
 
-def base_counts(ref: str) -> Mapping[str, int]:
+def _exit_on_termination(signum: int, _frame: FrameType | None) -> None:
+    raise SystemExit(128 + signum)
+
+
+def base_counts(ref: str, repo_root: Path = REPO_ROOT, checker: Path = CHECKER) -> Mapping[str, int]:
     """Rule counts at `ref`, measured with the *current* rule logic rather than
     whatever the checker looked like at that commit."""
+    for termination in TERMINATION_SIGNALS:
+        signal.signal(termination, _exit_on_termination)
     parent: Final = Path(tempfile.mkdtemp(prefix="tq_base_"))
     worktree: Final = parent / "wt"
     try:
-        _run(["git", "worktree", "add", "--detach", str(worktree), ref])
+        _run(["git", "worktree", "add", "--detach", str(worktree), ref], cwd=repo_root)
         (worktree / "scripts").mkdir(parents=True, exist_ok=True)
-        checker: Final = worktree / "scripts" / "check_test_quality.py"
-        shutil.copy(CHECKER, checker)
-        return count_by_rule(_check(worktree, checker))
+        base_checker: Final = worktree / "scripts" / "check_test_quality.py"
+        shutil.copy(checker, base_checker)
+        return count_by_rule(_check(worktree, base_checker))
     finally:
         # Teardown must never raise, or it masks the real error when the body failed.
         subprocess.run(
             ["git", "worktree", "remove", "--force", str(worktree)],
-            cwd=REPO_ROOT, capture_output=True, text=True,
+            cwd=repo_root, capture_output=True, text=True,
         )
         shutil.rmtree(parent, ignore_errors=True)
 
diff --git a/tests/test_litellm/test_test_quality_gate.py b/tests/test_litellm/test_test_quality_gate.py
index 8cce6bc735a..0664153f671 100644
--- a/tests/test_litellm/test_test_quality_gate.py
+++ b/tests/test_litellm/test_test_quality_gate.py
@@ -9,7 +9,13 @@ file:line.
 """
 
 import importlib.util
+import os
+import signal
+import subprocess
 import sys
+import time
+from collections.abc import Callable
+from contextlib import suppress
 from pathlib import Path
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
@@ -23,6 +29,15 @@ _spec.loader.exec_module(gate)
 
 _BUDGET = {"TQ001": {"limit": 10}, "TQ003": {"limit": 5}}
 
+_SCAN_BASE = (
+    "import importlib.util, pathlib, sys\n"
+    "spec = importlib.util.spec_from_file_location('test_quality_gate', sys.argv[1])\n"
+    "gate = importlib.util.module_from_spec(spec)\n"
+    "sys.modules[spec.name] = gate\n"
+    "spec.loader.exec_module(gate)\n"
+    "gate.base_counts('HEAD', repo_root=pathlib.Path(sys.argv[2]), checker=pathlib.Path(sys.argv[3]))\n"
+)
+
 
 def test_a_rule_within_its_limit_is_not_a_breach():
     assert gate.evaluate({"TQ001": 10}, {"TQ001": 10}, _BUDGET) == ()
@@ -146,3 +161,65 @@ def test_the_shipped_budget_covers_every_rule_the_checker_can_emit():
     budget = json.loads((_REPO_ROOT / "test-quality-budget.json").read_text())
     assert set(budget) == {"TQ001", "TQ002", "TQ003", "TQ004", "TQ005", "TQ006", "TQ007", "TQ008"}
     assert all(spec["limit"] >= 0 for spec in budget.values())
+
+
+def _git(cwd: Path, *args: str) -> str:
+    proc = subprocess.run(["git", *args], cwd=cwd, capture_output=True, text=True)
+    assert proc.returncode == 0, proc.stderr
+    return proc.stdout.strip()
+
+
+def _committed_repo(tmp_path: Path) -> Path:
+    repo = tmp_path / "repo"
+    (repo / "tests").mkdir(parents=True)
+    (repo / "tests" / "test_seed.py").write_text("def test_seed():\n    assert True\n")
+    _git(repo, "init", "-q", "-b", "main")
+    _git(repo, "config", "user.email", "gate@example.com")
+    _git(repo, "config", "user.name", "gate")
+    _git(repo, "config", "commit.gpgsign", "false")
+    _git(repo, "add", "-A")
+    _git(repo, "commit", "-q", "-m", "seed")
+    return repo
+
+
+def _wait_until(predicate: Callable[[], bool], timeout_seconds: float) -> bool:
+    deadline = time.monotonic() + timeout_seconds
+    while time.monotonic() < deadline:
+        if predicate():
+            return True
+        time.sleep(0.05)
+    return predicate()
+
+
+def _reap(process: subprocess.Popen[bytes]) -> None:
+    with suppress(subprocess.TimeoutExpired):
+        process.wait(timeout=10)
+    if process.poll() is None:
+        process.kill()
+        process.wait(timeout=10)
+
+
+def _registered_worktrees(repo: Path) -> int:
+    listing = _git(repo, "worktree", "list", "--porcelain")
+    return sum(line.startswith("worktree ") for line in listing.splitlines())
+
+
+def test_a_terminated_base_scan_still_removes_its_worktree(tmp_path: Path) -> None:
+    repo = _committed_repo(tmp_path)
+    scanning = tmp_path / "scanning"
+    slow_checker = tmp_path / "slow_checker.py"
+    slow_checker.write_text(f"import pathlib, time\npathlib.Path({str(scanning)!r}).touch()\ntime.sleep(30)\n")
+    temp_dir = tmp_path / "tmp"
+    temp_dir.mkdir()
+    scan = subprocess.Popen(
+        [sys.executable, "-c", _SCAN_BASE, str(_MODULE_PATH), str(repo), str(slow_checker)],
+        env={**os.environ, "TMPDIR": str(temp_dir)},
+    )
+    try:
+        assert _wait_until(scanning.exists, 30), "the base scan never reached the checker"
+        scan.send_signal(signal.SIGTERM)
+        assert scan.wait(timeout=30) == 128 + signal.SIGTERM
+    finally:
+        _reap(scan)
+    assert _registered_worktrees(repo) == 1
+    assert list(temp_dir.iterdir()) == []

From 952f082e3ee80eb6bec0855a752ee3c101652744 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 02:07:14 -0700
Subject: [PATCH 35/79] fix(test-quality-gate): keep a termination signal the
 parent already ignores ignored

The SIGTERM/SIGHUP teardown handlers were installed unconditionally, so a base scan started
under nohup (SIGHUP inherited as SIG_IGN) would start dying on hangups it was told to ignore.
Install them only where the disposition is still the default, and cover the ignored case with a
regression test that hangs up a scan started with SIGHUP ignored and expects it to finish.
---
 scripts/test_quality_gate.py                 |  9 +++-
 tests/test_litellm/test_test_quality_gate.py | 56 ++++++++++++++++----
 2 files changed, 53 insertions(+), 12 deletions(-)

diff --git a/scripts/test_quality_gate.py b/scripts/test_quality_gate.py
index 4f79c9488b1..c3316915b5e 100644
--- a/scripts/test_quality_gate.py
+++ b/scripts/test_quality_gate.py
@@ -122,11 +122,16 @@ def _exit_on_termination(signum: int, _frame: FrameType | None) -> None:
     raise SystemExit(128 + signum)
 
 
+def _install_termination_handlers() -> None:
+    for termination in TERMINATION_SIGNALS:
+        if signal.getsignal(termination) == signal.SIG_DFL:
+            signal.signal(termination, _exit_on_termination)
+
+
 def base_counts(ref: str, repo_root: Path = REPO_ROOT, checker: Path = CHECKER) -> Mapping[str, int]:
     """Rule counts at `ref`, measured with the *current* rule logic rather than
     whatever the checker looked like at that commit."""
-    for termination in TERMINATION_SIGNALS:
-        signal.signal(termination, _exit_on_termination)
+    _install_termination_handlers()
     parent: Final = Path(tempfile.mkdtemp(prefix="tq_base_"))
     worktree: Final = parent / "wt"
     try:
diff --git a/tests/test_litellm/test_test_quality_gate.py b/tests/test_litellm/test_test_quality_gate.py
index 0664153f671..b3511eabbfb 100644
--- a/tests/test_litellm/test_test_quality_gate.py
+++ b/tests/test_litellm/test_test_quality_gate.py
@@ -17,6 +17,7 @@ import time
 from collections.abc import Callable
 from contextlib import suppress
 from pathlib import Path
+from typing import NamedTuple
 
 _REPO_ROOT = Path(__file__).resolve().parents[2]
 _MODULE_PATH = _REPO_ROOT / "scripts" / "test_quality_gate.py"
@@ -37,6 +38,7 @@ _SCAN_BASE = (
     "spec.loader.exec_module(gate)\n"
     "gate.base_counts('HEAD', repo_root=pathlib.Path(sys.argv[2]), checker=pathlib.Path(sys.argv[3]))\n"
 )
+_SCAN_BASE_WITH_SIGHUP_IGNORED = "import signal\nsignal.signal(signal.SIGHUP, signal.SIG_IGN)\n" + _SCAN_BASE
 
 
 def test_a_rule_within_its_limit_is_not_a_breach():
@@ -204,22 +206,56 @@ def _registered_worktrees(repo: Path) -> int:
     return sum(line.startswith("worktree ") for line in listing.splitlines())
 
 
-def test_a_terminated_base_scan_still_removes_its_worktree(tmp_path: Path) -> None:
+class _StalledScan(NamedTuple):
+    process: subprocess.Popen[bytes]
+    repo: Path
+    release: Path
+    temp_dir: Path
+
+
+def _base_scan_stalled_in_its_checker(tmp_path: Path, driver: str) -> _StalledScan:
     repo = _committed_repo(tmp_path)
     scanning = tmp_path / "scanning"
+    release = tmp_path / "release"
     slow_checker = tmp_path / "slow_checker.py"
-    slow_checker.write_text(f"import pathlib, time\npathlib.Path({str(scanning)!r}).touch()\ntime.sleep(30)\n")
+    slow_checker.write_text(
+        "import pathlib, time\n"
+        f"pathlib.Path({str(scanning)!r}).touch()\n"
+        f"while not pathlib.Path({str(release)!r}).exists():\n"
+        "    time.sleep(0.05)\n"
+    )
     temp_dir = tmp_path / "tmp"
     temp_dir.mkdir()
     scan = subprocess.Popen(
-        [sys.executable, "-c", _SCAN_BASE, str(_MODULE_PATH), str(repo), str(slow_checker)],
+        [sys.executable, "-c", driver, str(_MODULE_PATH), str(repo), str(slow_checker)],
         env={**os.environ, "TMPDIR": str(temp_dir)},
     )
-    try:
-        assert _wait_until(scanning.exists, 30), "the base scan never reached the checker"
-        scan.send_signal(signal.SIGTERM)
-        assert scan.wait(timeout=30) == 128 + signal.SIGTERM
-    finally:
+    if not _wait_until(scanning.exists, 30):
         _reap(scan)
-    assert _registered_worktrees(repo) == 1
-    assert list(temp_dir.iterdir()) == []
+        raise AssertionError("the base scan never reached the checker")
+    return _StalledScan(scan, repo, release, temp_dir)
+
+
+def test_a_terminated_base_scan_still_removes_its_worktree(tmp_path: Path) -> None:
+    stalled = _base_scan_stalled_in_its_checker(tmp_path, _SCAN_BASE)
+    try:
+        stalled.process.send_signal(signal.SIGTERM)
+        assert stalled.process.wait(timeout=30) == 128 + signal.SIGTERM
+    finally:
+        _reap(stalled.process)
+    assert _registered_worktrees(stalled.repo) == 1
+    assert list(stalled.temp_dir.iterdir()) == []
+
+
+def test_a_base_scan_keeps_ignoring_the_hangup_its_parent_ignored(tmp_path: Path) -> None:
+    stalled = _base_scan_stalled_in_its_checker(tmp_path, _SCAN_BASE_WITH_SIGHUP_IGNORED)
+    try:
+        stalled.process.send_signal(signal.SIGHUP)
+        time.sleep(1)
+        assert stalled.process.poll() is None, "a hangup the parent ignored killed the scan"
+        stalled.release.touch()
+        assert stalled.process.wait(timeout=30) == 0
+    finally:
+        _reap(stalled.process)
+    assert _registered_worktrees(stalled.repo) == 1
+    assert list(stalled.temp_dir.iterdir()) == []

From 292f926fcd5b4882e3cda1e28c6fa9211a7860d4 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 15:56:17 -0700
Subject: [PATCH 36/79] fix(oci): stream Cohere tool-calling answers once

OCI Cohere restates the whole assistant text on the chunk that carries
the tool calls and again on the terminal chunk that carries chatHistory.
Only the terminal restatement was dropped, so a tool-calling turn streamed
the text twice. Treat a toolCalls-bearing chunk as a restatement too and
drop its text once deltas were already emitted.

Resolves LIT-6819
---
 litellm/llms/oci/chat/cohere.py               | 52 ++++-------------
 .../oci/chat/test_oci_cohere_tool_calls.py    | 57 +++++++++++++++++++
 2 files changed, 69 insertions(+), 40 deletions(-)

diff --git a/litellm/llms/oci/chat/cohere.py b/litellm/llms/oci/chat/cohere.py
index 6e9bb83b0a0..1b494ebad47 100644
--- a/litellm/llms/oci/chat/cohere.py
+++ b/litellm/llms/oci/chat/cohere.py
@@ -290,19 +290,14 @@ def handle_cohere_stream_chunk(
 ) -> ModelResponseStream:
     """Parse a single Cohere SSE chunk into a LiteLLM ModelResponseStream.
 
-    ``prior_tool_calls_emitted`` lets the caller signal whether tool calls
-    were already emitted in earlier chunks of the same stream. When set, the
-    terminal consolidation chunk's tool calls are suppressed (they would
-    duplicate prior deltas); otherwise they are passed through so a stream
-    that delivers tool calls only on the terminal chunk doesn't silently
-    drop them.
-
-    ``prior_text_emitted`` plays the analogous role for the ``text`` field:
-    when set, the terminal consolidation chunk's ``text`` is suppressed
-    (it would re-emit the full assembled response on top of prior deltas);
-    when unset (e.g. a degenerate stream that delivers the entire response
-    in a single SSE event carrying both ``chatHistory`` and ``finishReason``),
-    the text is passed through so the response content isn't silently lost.
+    OCI Cohere streams the answer as single-token ``text`` deltas, then restates
+    the whole assembled ``text`` on every chunk that carries ``toolCalls`` or
+    ``chatHistory`` (the tool-calls event and the terminal event). Once the
+    caller reports that earlier chunks already emitted text
+    (``prior_text_emitted``), those restatements are dropped so the client does
+    not see the answer twice; a stream whose only text lives on such a chunk
+    keeps it. ``prior_tool_calls_emitted`` plays the same role for the tool
+    calls the terminal ``chatHistory`` chunk repeats.
     """
     try:
         typed_chunk: Final = CohereStreamChunk.model_validate(dict_chunk)
@@ -315,33 +310,10 @@ def handle_cohere_stream_chunk(
     if typed_chunk.index is None:
         typed_chunk.index = 0
 
-    # OCI Cohere's terminal SSE event re-sends the full assembled response in
-    # `text` alongside a populated `chatHistory` and a non-null `finishReason`.
-    # Emitting that text would concatenate the whole response onto the
-    # already-streamed deltas. We require both signals to be present so that a
-    # future API change which adds `chatHistory` to intermediate chunks (or a
-    # rare early-populated case) doesn't silently drop legitimate token deltas.
-    is_terminal_consolidation: Final = typed_chunk.chatHistory is not None and typed_chunk.finishReason is not None
-    # On non-terminal text-free chunks (e.g. tool-call-only or keep-alive
-    # chunks) emit ``content=None`` rather than ``content=""`` so downstream
-    # stream-mergers that distinguish "no text in this delta" from "an
-    # explicitly empty text delta" behave correctly.
-    #
-    # We only suppress the terminal chunk's ``text`` when the caller has
-    # confirmed that text deltas were already emitted earlier — otherwise
-    # (e.g. a degenerate stream that delivers the whole response in a
-    # single SSE event), passing it through is the only chance to surface it.
-    text: Final[str | None] = None if (is_terminal_consolidation and prior_text_emitted) else typed_chunk.text
-
-    # Tool calls on the terminal consolidation chunk (whether from
-    # `typed_chunk.toolCalls` or from `chatHistory`) typically restate what
-    # was already streamed in intermediate chunks. Re-emitting them would
-    # mint fresh `uuid4` IDs and cause downstream consumers to execute each
-    # tool call twice. We only suppress when the caller has confirmed that
-    # tool calls were already emitted earlier — otherwise (e.g. a short
-    # response that delivers tool calls exclusively on the terminal chunk),
-    # passing them through is the only chance to surface them.
-    cohere_tool_calls = None if (is_terminal_consolidation and prior_tool_calls_emitted) else typed_chunk.toolCalls
+    restates_text: Final = typed_chunk.chatHistory is not None or typed_chunk.toolCalls is not None
+    restates_tool_calls: Final = typed_chunk.chatHistory is not None
+    text: Final[str | None] = None if (restates_text and prior_text_emitted) else typed_chunk.text
+    cohere_tool_calls: Final = None if (restates_tool_calls and prior_tool_calls_emitted) else typed_chunk.toolCalls
 
     tool_calls: list[dict[str, object]] | None = None
     if cohere_tool_calls:
diff --git a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
index 5dd44d72d68..97b2fd88cb9 100644
--- a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
+++ b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
@@ -445,6 +445,63 @@ class TestOCICohereToolCalls:
         assert result.choices[0].index == 0
         assert result.choices[0].finish_reason == "stop"  # COMPLETE is mapped to stop
 
+    _TOOL_TURN_TEXT = "I will use the tool to find out the weather in Paris."
+    _TOOL_TURN_DELTAS = ["I", " will", " use", " the", " tool", " to", " find", " out", " the", " weather", " in", " Paris", "."]
+    _TOOL_TURN_CALLS = [{"name": "get_weather", "parameters": {"city": "Paris"}}]
+    _TOOL_TURN_HISTORY = [
+        {"role": "USER", "message": "Briefly say what you will do, then find out the weather in Paris using the tool."},
+        {"role": "CHATBOT", "message": _TOOL_TURN_TEXT, "toolCalls": _TOOL_TURN_CALLS},
+    ]
+    _TOOL_TURN_TERMINAL_TOGETHER = [
+        {
+            "apiFormat": "COHERE",
+            "text": _TOOL_TURN_TEXT,
+            "chatHistory": _TOOL_TURN_HISTORY,
+            "finishReason": "COMPLETE",
+            "toolCalls": _TOOL_TURN_CALLS,
+        },
+    ]
+    _TOOL_TURN_TERMINAL_SPLIT = [
+        {"apiFormat": "COHERE", "text": _TOOL_TURN_TEXT, "chatHistory": _TOOL_TURN_HISTORY, "toolCalls": _TOOL_TURN_CALLS},
+        {"apiFormat": "COHERE", "finishReason": "COMPLETE"},
+    ]
+
+    @staticmethod
+    def _drain_cohere_stream(events):
+        wrapper = OCIStreamWrapper(completion_stream=MagicMock(), model="cohere.command-a-03-2025", logging_obj=MagicMock())
+        chunks = [wrapper.chunk_creator(f"data: {json.dumps(event)}") for event in events]
+        content = "".join(chunk.choices[0].delta.content or "" for chunk in chunks)
+        tool_calls = [call for chunk in chunks for call in (chunk.choices[0].delta.tool_calls or [])]
+        finish_reasons = [chunk.choices[0].finish_reason for chunk in chunks if chunk.choices[0].finish_reason]
+        return content, tool_calls, finish_reasons
+
+    @pytest.mark.parametrize("terminal_events", [_TOOL_TURN_TERMINAL_TOGETHER, _TOOL_TURN_TERMINAL_SPLIT])
+    def test_cohere_tool_turn_streams_the_answer_once(self, terminal_events):
+        """OCI restates the whole answer on the tool-calls chunk and again on the terminal chunk;
+        the client must read it exactly once, with one tool call and one finish reason."""
+        deltas = [{"apiFormat": "COHERE", "text": token} for token in self._TOOL_TURN_DELTAS]
+        tool_calls_event = {"apiFormat": "COHERE", "text": self._TOOL_TURN_TEXT, "toolCalls": self._TOOL_TURN_CALLS}
+
+        content, tool_calls, finish_reasons = self._drain_cohere_stream([*deltas, tool_calls_event, *terminal_events])
+
+        assert content == self._TOOL_TURN_TEXT
+        assert [(call["function"]["name"], call["function"]["arguments"]) for call in tool_calls] == [
+            ("get_weather", '{"city": "Paris"}')
+        ]
+        assert finish_reasons == ["stop"]
+
+    def test_cohere_tool_turn_without_preamble_deltas_keeps_the_only_text(self):
+        """When the tool-calls chunk carries the only copy of the text, dropping it would lose the answer."""
+        tool_calls_event = {"apiFormat": "COHERE", "text": self._TOOL_TURN_TEXT, "toolCalls": self._TOOL_TURN_CALLS}
+
+        content, tool_calls, finish_reasons = self._drain_cohere_stream(
+            [tool_calls_event, *self._TOOL_TURN_TERMINAL_TOGETHER]
+        )
+
+        assert content == self._TOOL_TURN_TEXT
+        assert len(tool_calls) == 1
+        assert finish_reasons == ["stop"]
+
     def test_cohere_parameter_mapping_excludes_tool_choice(self):
         """Test that tool_choice is excluded from Cohere parameter mapping"""
         config = OCIChatConfig()

From 43bb55d84934762a775bef3c32f6ed5ad8bc5ce6 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 16:38:23 -0700
Subject: [PATCH 37/79] style(oci): wrap the Cohere tool-turn test fixtures to
 120 columns

---
 .../llms/oci/chat/test_oci_cohere_tool_calls.py   | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
index 97b2fd88cb9..729a2d25f41 100644
--- a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
+++ b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
@@ -446,7 +446,9 @@ class TestOCICohereToolCalls:
         assert result.choices[0].finish_reason == "stop"  # COMPLETE is mapped to stop
 
     _TOOL_TURN_TEXT = "I will use the tool to find out the weather in Paris."
-    _TOOL_TURN_DELTAS = ["I", " will", " use", " the", " tool", " to", " find", " out", " the", " weather", " in", " Paris", "."]
+    _TOOL_TURN_DELTAS = [
+        "I", " will", " use", " the", " tool", " to", " find", " out", " the", " weather", " in", " Paris", ".",
+    ]
     _TOOL_TURN_CALLS = [{"name": "get_weather", "parameters": {"city": "Paris"}}]
     _TOOL_TURN_HISTORY = [
         {"role": "USER", "message": "Briefly say what you will do, then find out the weather in Paris using the tool."},
@@ -462,13 +464,20 @@ class TestOCICohereToolCalls:
         },
     ]
     _TOOL_TURN_TERMINAL_SPLIT = [
-        {"apiFormat": "COHERE", "text": _TOOL_TURN_TEXT, "chatHistory": _TOOL_TURN_HISTORY, "toolCalls": _TOOL_TURN_CALLS},
+        {
+            "apiFormat": "COHERE",
+            "text": _TOOL_TURN_TEXT,
+            "chatHistory": _TOOL_TURN_HISTORY,
+            "toolCalls": _TOOL_TURN_CALLS,
+        },
         {"apiFormat": "COHERE", "finishReason": "COMPLETE"},
     ]
 
     @staticmethod
     def _drain_cohere_stream(events):
-        wrapper = OCIStreamWrapper(completion_stream=MagicMock(), model="cohere.command-a-03-2025", logging_obj=MagicMock())
+        wrapper = OCIStreamWrapper(
+            completion_stream=MagicMock(), model="cohere.command-a-03-2025", logging_obj=MagicMock()
+        )
         chunks = [wrapper.chunk_creator(f"data: {json.dumps(event)}") for event in events]
         content = "".join(chunk.choices[0].delta.content or "" for chunk in chunks)
         tool_calls = [call for chunk in chunks for call in (chunk.choices[0].delta.tool_calls or [])]

From 347ea8f7ca08b2c5ce8fde4430f867b47ace4489 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 16:54:50 -0700
Subject: [PATCH 38/79] fix: keep provider id and metadata on Responses API
 bridged chat completions

---
 .../transformation.py                         |  23 +++-
 ...responses_transformation_transformation.py | 106 ++++++++++++++++++
 2 files changed, 126 insertions(+), 3 deletions(-)

diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py
index d70f947469a..331cd081fcd 100644
--- a/litellm/completion_extras/litellm_responses_transformation/transformation.py
+++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py
@@ -33,7 +33,7 @@ from litellm.responses.sse_output_recovery import (
     record_output_item_chunk,
     record_output_text_chunk,
 )
-from litellm.responses.utils import normalize_responses_api_stream_options
+from litellm.responses.utils import ResponsesAPIRequestUtils, normalize_responses_api_stream_options
 from litellm.types.llms.openai import (
     REASONING_EFFORT,
     ChatCompletionAnnotation,
@@ -54,7 +54,7 @@ if TYPE_CHECKING:
     )
     from pydantic import BaseModel
 
-    from litellm import LiteLLMLoggingObj, ModelResponse
+    from litellm import LiteLLMLoggingObj
     from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
     from litellm.types.llms.openai import (
         ALL_RESPONSES_API_TOOL_PARAMS,
@@ -69,6 +69,15 @@ if TYPE_CHECKING:
     from litellm.types.utils import Choices
 
 
+_CHAT_COMPLETION_FIELDS: Final = frozenset((*ModelResponse.model_fields, "usage"))
+
+
+def _upstream_response_id(response_id: str | None) -> str | None:
+    if response_id is None:
+        return None
+    return ResponsesAPIRequestUtils.decode_previous_response_id_to_original_previous_response_id(response_id)
+
+
 class _ReasoningSummaryText(TypedDict):
     type: str
     text: str
@@ -904,6 +913,12 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
             ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(raw_response.usage),
         )
 
+        model_response.id = _upstream_response_id(raw_response.id) or raw_response.id
+        provider_extras: Final = raw_response.model_extra.items() if raw_response.model_extra else ()
+        for key, value in provider_extras:
+            if key not in _CHAT_COMPLETION_FIELDS and value is not None:
+                setattr(model_response, key, value)
+
         # Preserve hidden params from the ResponsesAPIResponse, especially the headers
         # which contain important provider information like x-request-id
         raw_response_hidden_params: Final = getattr(raw_response, "_hidden_params", {})
@@ -1359,14 +1374,16 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
         if event_type == "response.created":
             # Initial response creation event
             verbose_logger.debug("Chat provider: response.created -> %s", parsed_chunk)
+            created_response: Final = parsed_chunk.get("response")
             return ModelResponseStream(
+                id=_upstream_response_id(created_response.get("id")) if created_response else None,
                 choices=[
                     StreamingChoices(
                         index=0,
                         delta=Delta(content=""),
                         finish_reason=None,
                     )
-                ]
+                ],
             )
         elif event_type == "response.output_item.added":
             # New output item added
diff --git a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
index 6590718878d..97b18382a68 100644
--- a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
+++ b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
@@ -3952,3 +3952,109 @@ def test_convert_chat_completion_messages_to_responses_api_tool_result_with_tool
 
     function_call_output = next(item for item in response if item.get("type") == "function_call_output")
     assert function_call_output["output"] == [{"type": "input_text", "text": "1 tool found"}]
+
+
+def _litellm_encoded_response_id(upstream_id: str) -> str:
+    """The id a Responses API call hands back through LiteLLM: the provider's id wrapped with the
+    deployment it came from, the way ``/v1/responses`` clients see it."""
+    import base64
+
+    tagged = f"litellm:custom_llm_provider:azure;model_id:deployment-1;response_id:{upstream_id}"
+    return "resp_" + base64.b64encode(tagged.encode()).decode()
+
+
+def test_transform_response_keeps_upstream_id_and_provider_extras():
+    """A chat completion bridged through the Responses API must answer with the provider's own
+    response id, decoded out of the deployment-tagged id LiteLLM wraps around it, and every
+    top-level field the provider adds beyond the Responses schema (Azure's ``content_filters``,
+    ``service_tier``), the way the native chat path passes unknown top-level fields through,
+    instead of a locally minted ``chatcmpl-`` id and nothing else."""
+    from unittest.mock import Mock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        LiteLLMResponsesTransformationHandler,
+    )
+    from litellm.types.llms.openai import ResponsesAPIResponse
+    from litellm.types.utils import ModelResponse, Usage
+
+    content_filters = [
+        {"blocked": False, "source_type": "prompt", "content_filter_results": {"hate": {"filtered": False}}}
+    ]
+    raw_response = ResponsesAPIResponse.model_validate(
+        {
+            "id": _litellm_encoded_response_id("resp_azure_123"),
+            "created_at": 1734366691,
+            "object": "response",
+            "model": "gpt-5.6",
+            "status": "completed",
+            "output": [
+                {
+                    "type": "function_call",
+                    "id": "fc_1",
+                    "call_id": "call_1",
+                    "name": "lookup_weather",
+                    "arguments": '{"city": "Seattle"}',
+                    "status": "completed",
+                }
+            ],
+            "parallel_tool_calls": True,
+            "tool_choice": "auto",
+            "tools": [],
+            "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+            "service_tier": "default",
+            "content_filters": content_filters,
+            "max_tool_calls": None,
+        }
+    )
+    model_response = ModelResponse(
+        id="chatcmpl-local",
+        created=1734366691,
+        model=None,
+        object="chat.completion",
+        choices=[],
+        usage=Usage(completion_tokens=0, prompt_tokens=0, total_tokens=0),
+    )
+
+    result = LiteLLMResponsesTransformationHandler().transform_response(
+        model="gpt-5.6",
+        raw_response=raw_response,
+        model_response=model_response,
+        logging_obj=Mock(),
+        request_data={"model": "gpt-5.6"},
+        messages=[{"role": "user", "content": "What is the weather in Seattle?"}],
+        optional_params={},
+        litellm_params={},
+        encoding=Mock(),
+    )
+    dumped = result.model_dump()
+
+    assert dumped["id"] == "resp_azure_123"
+    assert dumped["object"] == "chat.completion"
+    assert dumped["service_tier"] == "default"
+    assert dumped["content_filters"] == content_filters
+    assert "max_tool_calls" not in dumped, "a null provider field must not appear as a null top-level key"
+    assert "output" not in dumped and "status" not in dumped, (
+        "Responses schema fields must not leak into the chat response"
+    )
+    assert dumped["choices"][0]["message"]["tool_calls"][0]["function"]["name"] == "lookup_weather"
+
+
+def test_streaming_chunks_carry_the_upstream_response_id():
+    """Every streamed chunk of a bridged chat completion must carry the provider's response id
+    from ``response.created`` rather than a locally minted ``chatcmpl-`` id, so a client can
+    correlate the stream with the provider's request the same way the non-streaming path does."""
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+
+    iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
+    encoded_id = _litellm_encoded_response_id("resp_azure_stream")
+    events = [
+        {"type": "response.created", "response": {"id": encoded_id, "output": []}},
+        {"type": "response.output_text.delta", "delta": "Hel"},
+        {"type": "response.completed", "response": {"id": encoded_id, "output": [{"type": "message"}]}},
+    ]
+
+    ids = [iterator.chunk_parser(event).id for event in events]
+
+    assert ids == ["resp_azure_stream"] * len(events), f"streamed chunks did not carry the upstream id: {ids}"

From 089b4b8ff458681bcbfa9741e85952e3c6aa0ece Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 16:59:24 -0700
Subject: [PATCH 39/79] test: drop docstrings from the bridge id regression
 tests

---
 ..._litellm_responses_transformation_transformation.py | 10 ----------
 1 file changed, 10 deletions(-)

diff --git a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
index 97b18382a68..1599449382b 100644
--- a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
+++ b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
@@ -3955,8 +3955,6 @@ def test_convert_chat_completion_messages_to_responses_api_tool_result_with_tool
 
 
 def _litellm_encoded_response_id(upstream_id: str) -> str:
-    """The id a Responses API call hands back through LiteLLM: the provider's id wrapped with the
-    deployment it came from, the way ``/v1/responses`` clients see it."""
     import base64
 
     tagged = f"litellm:custom_llm_provider:azure;model_id:deployment-1;response_id:{upstream_id}"
@@ -3964,11 +3962,6 @@ def _litellm_encoded_response_id(upstream_id: str) -> str:
 
 
 def test_transform_response_keeps_upstream_id_and_provider_extras():
-    """A chat completion bridged through the Responses API must answer with the provider's own
-    response id, decoded out of the deployment-tagged id LiteLLM wraps around it, and every
-    top-level field the provider adds beyond the Responses schema (Azure's ``content_filters``,
-    ``service_tier``), the way the native chat path passes unknown top-level fields through,
-    instead of a locally minted ``chatcmpl-`` id and nothing else."""
     from unittest.mock import Mock
 
     from litellm.completion_extras.litellm_responses_transformation.transformation import (
@@ -4040,9 +4033,6 @@ def test_transform_response_keeps_upstream_id_and_provider_extras():
 
 
 def test_streaming_chunks_carry_the_upstream_response_id():
-    """Every streamed chunk of a bridged chat completion must carry the provider's response id
-    from ``response.created`` rather than a locally minted ``chatcmpl-`` id, so a client can
-    correlate the stream with the provider's request the same way the non-streaming path does."""
     from litellm.completion_extras.litellm_responses_transformation.transformation import (
         OpenAiResponsesToChatCompletionStreamIterator,
     )

From 64cbe6d0aa332bb6295dd37d6eb2066629c3af07 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 19:21:40 -0700
Subject: [PATCH 40/79] fix(bridge): carry provider metadata on streamed chats
 and keep served ids in spend logs

The terminal chunk of a bridged streaming chat now carries the same provider
fields the non-streaming response does (service_tier, content_filters), so
streaming clients see them on the final chunk

The passthrough drops Responses API bookkeeping (background, top_logprobs,
store, ...) by subtracting the OpenAI SDK's Response schema from the fields it
copies instead of a hand-kept denylist

Spend log rows for /v1/messages calls served through the Responses adapter keep
the id the client was handed instead of the decoded upstream id
---
 .../transformation.py                         |  25 ++++-
 litellm/litellm_core_utils/litellm_logging.py |   9 +-
 ...responses_transformation_transformation.py | 105 +++++++++++++++++-
 .../test_litellm_logging.py                   |  33 ++++++
 4 files changed, 162 insertions(+), 10 deletions(-)

diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py
index 331cd081fcd..87350b5479c 100644
--- a/litellm/completion_extras/litellm_responses_transformation/transformation.py
+++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py
@@ -5,8 +5,11 @@ Handler for transforming /chat/completions api requests to litellm.responses req
 import json
 import os
 from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, TypedDict, Union, cast, get_args
 
+from openai.types.chat import ChatCompletion
+from openai.types.responses import Response
 from openai.types.responses.custom_tool_param import CustomToolParam
 from openai.types.responses.response_input_param import (
     FunctionCallOutput,
@@ -43,6 +46,7 @@ from litellm.types.llms.openai import (
     ChatCompletionToolParamFunctionChunk,
     Reasoning,
     ResponsesAPIOptionalRequestParams,
+    ResponsesAPIResponse,
     ResponsesAPIStreamEvents,
 )
 from litellm.types.utils import GenericStreamingChunk, ModelResponseStream
@@ -70,6 +74,19 @@ if TYPE_CHECKING:
 
 
 _CHAT_COMPLETION_FIELDS: Final = frozenset((*ModelResponse.model_fields, "usage"))
+_RESPONSES_API_ONLY_FIELDS: Final = frozenset((*Response.model_fields, *ResponsesAPIResponse.model_fields)) - frozenset(
+    ChatCompletion.model_fields
+)
+
+
+def _provider_metadata(response_fields: Mapping[str, object] | None) -> Mapping[str, object]:
+    return MappingProxyType(
+        {
+            key: value
+            for key, value in (response_fields.items() if response_fields else ())
+            if value is not None and key not in _CHAT_COMPLETION_FIELDS and key not in _RESPONSES_API_ONLY_FIELDS
+        }
+    )
 
 
 def _upstream_response_id(response_id: str | None) -> str | None:
@@ -914,10 +931,8 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
         )
 
         model_response.id = _upstream_response_id(raw_response.id) or raw_response.id
-        provider_extras: Final = raw_response.model_extra.items() if raw_response.model_extra else ()
-        for key, value in provider_extras:
-            if key not in _CHAT_COMPLETION_FIELDS and value is not None:
-                setattr(model_response, key, value)
+        for key, value in _provider_metadata(raw_response.model_extra).items():
+            setattr(model_response, key, value)
 
         # Preserve hidden params from the ResponsesAPIResponse, especially the headers
         # which contain important provider information like x-request-id
@@ -1551,6 +1566,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                 from litellm.responses.utils import ResponseAPILoggingUtils
 
                 usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(response_data.get("usage"))
+            provider_metadata: Final = _provider_metadata(response_data)
             return ModelResponseStream(
                 choices=[
                     StreamingChoices(
@@ -1563,6 +1579,7 @@ class OpenAiResponsesToChatCompletionStreamIterator(BaseModelResponseIterator):
                     )
                 ],
                 usage=usage,
+                provider_specific_fields=dict(provider_metadata) or None,  # mutable-ok: field is typed dict
             )
         else:
             pass
diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index c31c4323157..dbd58766488 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -3918,11 +3918,12 @@ class Logging(LiteLLMLoggingBaseClass):
             LiteLLMResponsesTransformationHandler,
         )
 
+        served_id: Final = _provider_response_id(result)
         try:
-            return LiteLLMResponsesTransformationHandler().transform_response(
+            translated: Final = LiteLLMResponsesTransformationHandler().transform_response(
                 model=self.model,
                 raw_response=result,
-                model_response=litellm.ModelResponse(id=_provider_response_id(result)),
+                model_response=litellm.ModelResponse(id=served_id),
                 logging_obj=self,
                 request_data={},
                 messages=[],
@@ -3930,6 +3931,8 @@ class Logging(LiteLLMLoggingBaseClass):
                 litellm_params={},
                 encoding=litellm.encoding,
             )
+            translated.id = served_id or translated.id
+            return translated
         except Exception as e:
             verbose_logger.debug(
                 "Responses API -> ModelResponse translation failed for "
@@ -3937,7 +3940,7 @@ class Logging(LiteLLMLoggingBaseClass):
                 "usage-only ModelResponse to keep the spend_logs row.",
                 str(e),
             )
-            model_response: Final = litellm.ModelResponse(id=_provider_response_id(result))
+            model_response: Final = litellm.ModelResponse(id=served_id)
             model_response.model = self.model
             usage: Final = getattr(result, "usage", None)
             if usage is not None and ResponseAPILoggingUtils._is_response_api_usage(usage):
diff --git a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
index 1599449382b..d4d47b145d1 100644
--- a/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
+++ b/tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
@@ -3955,10 +3955,11 @@ def test_convert_chat_completion_messages_to_responses_api_tool_result_with_tool
 
 
 def _litellm_encoded_response_id(upstream_id: str) -> str:
-    import base64
+    from litellm.responses.utils import ResponsesAPIRequestUtils
 
-    tagged = f"litellm:custom_llm_provider:azure;model_id:deployment-1;response_id:{upstream_id}"
-    return "resp_" + base64.b64encode(tagged.encode()).decode()
+    return ResponsesAPIRequestUtils._build_responses_api_response_id(
+        custom_llm_provider="azure", model_id="deployment-1", response_id=upstream_id
+    )
 
 
 def test_transform_response_keeps_upstream_id_and_provider_extras():
@@ -3997,6 +3998,9 @@ def test_transform_response_keeps_upstream_id_and_provider_extras():
             "service_tier": "default",
             "content_filters": content_filters,
             "max_tool_calls": None,
+            "background": False,
+            "top_logprobs": 0,
+            "store": True,
         }
     )
     model_response = ModelResponse(
@@ -4029,9 +4033,63 @@ def test_transform_response_keeps_upstream_id_and_provider_extras():
     assert "output" not in dumped and "status" not in dumped, (
         "Responses schema fields must not leak into the chat response"
     )
+    assert not {"background", "top_logprobs", "store"} & dumped.keys(), (
+        "Responses API bookkeeping must not ride along as chat metadata"
+    )
     assert dumped["choices"][0]["message"]["tool_calls"][0]["function"]["name"] == "lookup_weather"
 
 
+def test_bridged_response_is_priced_by_the_reported_service_tier():
+    from unittest.mock import Mock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        LiteLLMResponsesTransformationHandler,
+    )
+    from litellm.types.llms.openai import ResponsesAPIResponse
+    from litellm.types.utils import ModelResponse
+
+    raw_response = ResponsesAPIResponse.model_validate(
+        {
+            "id": "resp_flex",
+            "created_at": 1734366691,
+            "object": "response",
+            "model": "gpt-5.4",
+            "status": "completed",
+            "output": [
+                {
+                    "type": "message",
+                    "id": "msg_1",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [{"type": "output_text", "text": "hi", "annotations": []}],
+                }
+            ],
+            "usage": {"input_tokens": 1000, "output_tokens": 100, "total_tokens": 1100},
+            "service_tier": "flex",
+        }
+    )
+
+    result = LiteLLMResponsesTransformationHandler().transform_response(
+        model="gpt-5.4",
+        raw_response=raw_response,
+        model_response=ModelResponse(),
+        logging_obj=Mock(),
+        request_data={"model": "gpt-5.4"},
+        messages=[{"role": "user", "content": "hi"}],
+        optional_params={},
+        litellm_params={},
+        encoding=Mock(),
+    )
+    pricing = litellm.model_cost["gpt-5.4"]
+    flex_cost = 1000 * pricing["input_cost_per_token_flex"] + 100 * pricing["output_cost_per_token_flex"]
+    standard_cost = 1000 * pricing["input_cost_per_token"] + 100 * pricing["output_cost_per_token"]
+
+    cost = litellm.completion_cost(completion_response=result, custom_llm_provider="openai")
+
+    assert cost == pytest.approx(flex_cost)
+    assert cost < standard_cost
+
+
 def test_streaming_chunks_carry_the_upstream_response_id():
     from litellm.completion_extras.litellm_responses_transformation.transformation import (
         OpenAiResponsesToChatCompletionStreamIterator,
@@ -4048,3 +4106,44 @@ def test_streaming_chunks_carry_the_upstream_response_id():
     ids = [iterator.chunk_parser(event).id for event in events]
 
     assert ids == ["resp_azure_stream"] * len(events), f"streamed chunks did not carry the upstream id: {ids}"
+
+
+def test_streaming_final_chunk_carries_provider_metadata():
+    from unittest.mock import MagicMock
+
+    from litellm.completion_extras.litellm_responses_transformation.transformation import (
+        OpenAiResponsesToChatCompletionStreamIterator,
+    )
+    from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+
+    iterator = OpenAiResponsesToChatCompletionStreamIterator(streaming_response=None, sync_stream=True)
+    content_filters = [{"blocked": False, "source_type": "completion", "content_filter_results": {}}]
+    events = [
+        {"type": "response.created", "response": {"id": "resp_azure_stream", "output": []}},
+        {"type": "response.output_text.delta", "delta": "Hello"},
+        {
+            "type": "response.completed",
+            "response": {
+                "id": "resp_azure_stream",
+                "output": [{"type": "message"}],
+                "usage": {"input_tokens": 3, "output_tokens": 1, "total_tokens": 4},
+                "service_tier": "default",
+                "content_filters": content_filters,
+                "background": False,
+            },
+        },
+    ]
+    stream = CustomStreamWrapper(
+        completion_stream=iter([iterator.chunk_parser(event) for event in events]),
+        model="gpt-5.6",
+        custom_llm_provider="azure",
+        logging_obj=MagicMock(),
+    )
+
+    chunks = [chunk.model_dump() for chunk in stream]
+
+    assert chunks[-1]["choices"][0]["finish_reason"] == "stop"
+    assert chunks[-1]["service_tier"] == "default"
+    assert chunks[-1]["content_filters"] == content_filters
+    assert "background" not in chunks[-1]
+    assert all("service_tier" not in chunk for chunk in chunks[:-1])
diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
index 16a99713a06..b5d56f390a9 100644
--- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
+++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
@@ -4391,6 +4391,39 @@ def test_handle_anthropic_messages_response_logging_translates_bare_responses_ap
     assert result.usage.total_tokens == 18  # type: ignore[attr-defined]
 
 
+def test_handle_anthropic_messages_response_logging_keeps_the_served_response_id():
+    from openai.types.responses import ResponseOutputMessage, ResponseOutputText
+
+    from litellm.responses.utils import ResponsesAPIRequestUtils
+    from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse
+
+    served_id = ResponsesAPIRequestUtils._build_responses_api_response_id(
+        custom_llm_provider="openai", model_id="deployment-1", response_id="resp_upstream"
+    )
+    logging_obj = _anthropic_messages_logging_obj()
+    result = logging_obj._handle_anthropic_messages_response_logging(
+        result=ResponsesAPIResponse(
+            id=served_id,
+            created_at=1700000000,
+            output=[
+                ResponseOutputMessage(
+                    id="msg-1",
+                    type="message",
+                    role="assistant",
+                    status="completed",
+                    content=[ResponseOutputText(annotations=[], text="hi", type="output_text")],
+                )
+            ],
+            usage=ResponseAPIUsage(input_tokens=2, output_tokens=1, total_tokens=3),
+            service_tier="flex",
+        )
+    )
+
+    assert isinstance(result, ModelResponse)
+    assert result.id == served_id, "the spend log row must keep the id the caller was served"
+    assert result.service_tier == "flex"
+
+
 def test_handle_anthropic_messages_response_logging_passes_model_response_through():
     """Anthropic-native path already yields a ModelResponse; it must be returned unchanged."""
     logging_obj = _anthropic_messages_logging_obj()

From 2e2fce5e583f31889dd4a75bd364cf0c2ba3cbe3 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 22:25:13 -0700
Subject: [PATCH 41/79] fix(router): skip the refusing deployment when retrying
 a non-transient error

BadRequestErrorRetries and ContentPolicyViolationErrorRetries did let a retry
happen, but the retry re-picked the deployment that had just refused, since a
400 never puts a deployment in cooldown. On a weighted model group the caller
got the same 400 back after every configured retry, and the existing 401/403
"retry on another deployment" rule broke the same way

A retry after a non-transient status now carries the deployments that already
answered this request in the per-request exclusion list weighted failover
already honors, so the next attempt lands on a sibling. Single-deployment
groups still retry in place, and 408/429/5xx retries are untouched

Adds live e2e coverage for reliability.retry.context_window.succeeds_within_retries
and renames the two litellm.utils deployment filters that are now called from
outside the module
---
 basedpyright-code-budget.json                 |  4 +-
 litellm/router.py                             | 45 +++++++++-
 litellm/utils.py                              |  4 +-
 tests/e2e/models.py                           |  1 +
 tests/e2e/router/reliability_support.py       | 19 +++-
 .../router/test_reliability_retries_e2e.py    | 88 +++++++++++++------
 tests/test_litellm/test_router.py             | 82 +++++++++++++++++
 .../test_router_order_fallback.py             | 18 ++--
 .../test_router_weighted_failover.py          | 16 ++--
 9 files changed, 223 insertions(+), 54 deletions(-)

diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json
index 0b0a61192e6..57ca267e504 100644
--- a/basedpyright-code-budget.json
+++ b/basedpyright-code-budget.json
@@ -84,7 +84,7 @@
     "limit": 56
   },
   "reportPrivateUsage": {
-    "limit": 1808
+    "limit": 1804
   },
   "reportRedeclaration": {
     "limit": 8
@@ -135,7 +135,7 @@
     "limit": 21
   },
   "reportUnusedFunction": {
-    "limit": 138
+    "limit": 136
   },
   "reportUnusedImport": {
     "limit": 542
diff --git a/litellm/router.py b/litellm/router.py
index 6d6efad9f42..f72f41313a0 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -401,6 +401,7 @@ def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream])
 
 _NO_SESSION_KWARGS: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({})
 _SESSION_ADAPTER: Final = TypeAdapter(Mapping[str, object])
+_EXCLUDED_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
 
 
 def _with_router_resolved_session_model(session: object, model_name: str) -> Mapping[str, Mapping[str, object]]:
@@ -7458,6 +7459,28 @@ class Router:
                     Context_Policy_Fallbacks={content_policy_fallbacks}",
             )
 
+    @staticmethod
+    def _deployment_ids_to_skip_on_retry(
+        exception: Exception,
+        already_skipped: object,
+        healthy_deployments: list[dict],  # mutable-ok: matches the routing filters' list contract
+    ) -> tuple[str, ...]:
+        failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
+        status_code: Final = getattr(exception, "status_code", None)
+        if not failed_deployment_id or status_code is None:
+            return ()
+        if litellm._should_retry(status_code):  # pyright: ignore[reportPrivateUsage]  # as in should_retry_this_error
+            return ()
+        already_skipped_ids: Final = _EXCLUDED_DEPLOYMENT_IDS_ADAPTER.validate_python(already_skipped or ())
+        skipped: Final = frozenset((*already_skipped_ids, failed_deployment_id))
+        same_order_candidates: Final = litellm.utils.get_order_filtered_deployments(healthy_deployments)
+        if not litellm.utils.get_excluded_filtered_deployments(same_order_candidates, excluded_deployment_ids=skipped):
+            return ()
+        verbose_router_logger.debug(
+            "Retry skips deployments that already answered %s to this request: %s", status_code, sorted(skipped)
+        )
+        return tuple(sorted(skipped))
+
     @tracer.wrap()
     async def async_function_with_retries(self, *args, **kwargs):
         verbose_router_logger.debug("Inside async function with retries.")
@@ -7553,6 +7576,13 @@ class Router:
             ## LOGGING
             if num_retries > 0:
                 kwargs = self.log_retry(kwargs=kwargs, e=original_exception)
+                skipped_deployment_ids: Final = self._deployment_ids_to_skip_on_retry(
+                    exception=original_exception,
+                    already_skipped=kwargs.get("_excluded_deployment_ids"),
+                    healthy_deployments=_healthy_deployments,
+                )
+                if skipped_deployment_ids:
+                    kwargs["_excluded_deployment_ids"] = skipped_deployment_ids
             else:
                 raise
 
@@ -7622,6 +7652,13 @@ class Router:
                         except Exception:
                             raise e
 
+                    retry_skipped_deployment_ids = self._deployment_ids_to_skip_on_retry(
+                        exception=e,
+                        already_skipped=kwargs.get("_excluded_deployment_ids"),
+                        healthy_deployments=_healthy_deployments,
+                    )
+                    if retry_skipped_deployment_ids:
+                        kwargs["_excluded_deployment_ids"] = retry_skipped_deployment_ids
                     _timeout = self._time_to_sleep_before_retry(
                         e=e,
                         remaining_retries=remaining_retries,
@@ -12452,7 +12489,7 @@ class Router:
 
         ## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
         _target_order: Final = (request_kwargs or {}).pop("_target_order", None)
-        healthy_deployments = litellm.utils._get_order_filtered_deployments(
+        healthy_deployments = litellm.utils.get_order_filtered_deployments(
             cast(list[dict], healthy_deployments), target_order=_target_order
         )
 
@@ -12460,7 +12497,7 @@ class Router:
         ## this request via weighted-failover. Always honored, regardless of the
         ## router-level flag, so a stale exclusion key on kwargs cannot escape.
         _excluded_deployment_ids: Final = (request_kwargs or {}).pop("_excluded_deployment_ids", None)
-        healthy_deployments = litellm.utils._get_excluded_filtered_deployments(
+        healthy_deployments = litellm.utils.get_excluded_filtered_deployments(
             cast(list[dict], healthy_deployments),
             excluded_deployment_ids=_excluded_deployment_ids,
         )
@@ -13357,7 +13394,7 @@ class Router:
 
         ## ORDER FILTERING ## -> if user set 'order' in deployments, return deployments with lowest order (e.g. order=1 > order=2)
         _target_order: Final = (request_kwargs or {}).pop("_target_order", None)
-        healthy_deployments = litellm.utils._get_order_filtered_deployments(
+        healthy_deployments = litellm.utils.get_order_filtered_deployments(
             healthy_deployments, target_order=_target_order
         )
 
@@ -13365,7 +13402,7 @@ class Router:
         ## this request via weighted-failover. See async counterpart in
         ## async_get_healthy_deployments for details.
         _excluded_deployment_ids: Final = (request_kwargs or {}).pop("_excluded_deployment_ids", None)
-        healthy_deployments = litellm.utils._get_excluded_filtered_deployments(
+        healthy_deployments = litellm.utils.get_excluded_filtered_deployments(
             healthy_deployments,
             excluded_deployment_ids=_excluded_deployment_ids,
         )
diff --git a/litellm/utils.py b/litellm/utils.py
index 52c1859b525..238225eff99 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -4889,7 +4889,7 @@ def _get_deployment_order(deployment: dict | Any) -> int | None:
     return order
 
 
-def _get_order_filtered_deployments(healthy_deployments: list[dict], target_order: int | None = None) -> list:
+def get_order_filtered_deployments(healthy_deployments: list[dict], target_order: int | None = None) -> list:
     if target_order is not None:
         return [d for d in healthy_deployments if _get_deployment_order(d) == target_order]
 
@@ -4908,7 +4908,7 @@ def _get_order_filtered_deployments(healthy_deployments: list[dict], target_orde
     return healthy_deployments
 
 
-def _get_excluded_filtered_deployments(
+def get_excluded_filtered_deployments(
     healthy_deployments: list[dict],
     excluded_deployment_ids: Iterable[str] | None = None,
 ) -> list:
diff --git a/tests/e2e/models.py b/tests/e2e/models.py
index 2c6c0e9bbd4..c01687d0b31 100644
--- a/tests/e2e/models.py
+++ b/tests/e2e/models.py
@@ -291,6 +291,7 @@ class RouterSettingsOverride(BaseModel):
     context_window_fallbacks: list[dict[str, list[str]]] | None = None
     content_policy_fallbacks: list[dict[str, list[str]]] | None = None
     num_retries: int | None = None
+    model_group_retry_policy: dict[str, dict[str, int]] | None = None
     enable_tag_filtering: bool | None = None
 
 
diff --git a/tests/e2e/router/reliability_support.py b/tests/e2e/router/reliability_support.py
index 5822058003c..1efcb1a045b 100644
--- a/tests/e2e/router/reliability_support.py
+++ b/tests/e2e/router/reliability_support.py
@@ -73,10 +73,25 @@ def create_always_timing_out_deployment(proxy: ProxyClient, name: str) -> str:
     )
 
 
+def create_always_picked_small_context_deployment(proxy: ProxyClient, name: str) -> str:
+    """The always-picked half of a retry pair on the smallest-context model OpenAI
+    still serves: it holds all of the model group's shuffle weight, so an oversized
+    prompt opens on it and earns a real context-window refusal, which never benches
+    a deployment, so only the retry itself can steer the request off it."""
+    return proxy.register_model(
+        ModelNewBody(
+            model_name=name,
+            litellm_params=LiteLLMParamsBody(model=SMALL_CONTEXT_MODEL, api_key=REAL_KEY, weight=1),
+            model_info=ModelInfoBody(),
+        )
+    )
+
+
 def create_zero_weight_backup_deployment(proxy: ProxyClient, name: str) -> str:
     """The other half of a retry pair: healthy, but weight 0, so the weighted shuffle
-    never opens on it. It is reachable only once its sibling is benched and the
-    weighted pick falls through to a uniform one over what is left."""
+    never opens on it. It is reachable only once its sibling is out of the running,
+    benched by a cooldown or skipped by the retry, and the weighted pick falls through
+    to a uniform one over what is left."""
     return proxy.register_model(
         ModelNewBody(
             model_name=name,
diff --git a/tests/e2e/router/test_reliability_retries_e2e.py b/tests/e2e/router/test_reliability_retries_e2e.py
index 5441412935c..da45cb46a46 100644
--- a/tests/e2e/router/test_reliability_retries_e2e.py
+++ b/tests/e2e/router/test_reliability_retries_e2e.py
@@ -1,13 +1,17 @@
 """Live e2e: a request that fails on its first deployment is retried inside its own
 model group and still comes back a completion.
 
-The model group is a pair: an always-timing-out deployment that holds all of the
-group's shuffle weight, and a healthy backup at weight 0. The weighted pick always
-opens on the timing-out one, its first Timeout benches it (an
-`allowed_fails_policy` of `TimeoutErrorAllowedFails: 0`), and the retry falls
-through to the only deployment left. So the customer sees a completion and the
-proxy reports that it took a retry to get there, with no random first pick in the
-middle of it.
+Each model group is a pair: a deployment that always refuses and holds all of the
+group's shuffle weight, plus a healthy backup at weight 0. The weighted pick always
+opens on the refusing one, so the customer sees a completion only if the retry
+lands on the backup, and the proxy reports that it took a retry to get there, with
+no random first pick in the middle of it.
+
+The timeout pair relies on cooldown: the first Timeout benches the timing-out
+deployment (an `allowed_fails_policy` of `TimeoutErrorAllowedFails: 0`) and the
+retry falls through to the only deployment left. The context-window pair cannot:
+a 400 never benches a deployment, so the retry policy's `BadRequestErrorRetries`
+has to steer the retry off the deployment that just refused the prompt.
 """
 
 from __future__ import annotations
@@ -16,20 +20,48 @@ import pytest
 
 from complexity_router_client import ComplexityRouterClient
 from e2e_config import unique_marker
+from e2e_http import StreamingResponse
 from lifecycle import ResourceManager
 from models import RouterSettingsOverride
 from reliability_support import (
     chat_override,
     completion_tokens_of,
     content_of,
+    create_always_picked_small_context_deployment,
     create_always_timing_out_deployment,
     create_zero_weight_backup_deployment,
     finish_reason_of,
+    oversized_prompt,
 )
 
 pytestmark = pytest.mark.e2e
 
 
+def assert_retry_landed_on_backup(resp: StreamingResponse) -> None:
+    assert resp.status_code == 200, (
+        f"the retry should have landed on the healthy backup, got {resp.status_code}: {resp.body[:300]}"
+    )
+
+    attempted = resp.headers.get("x-litellm-attempted-retries")
+    assert attempted is not None, "response is missing the x-litellm-attempted-retries header"
+    assert int(attempted) >= 1, (
+        f"x-litellm-attempted-retries is {attempted!r}; a 200 with no retry means the request never "
+        "opened on the refusing deployment, so this proves nothing about retries"
+    )
+
+    content = content_of(resp)
+    finish_reason = finish_reason_of(resp)
+    completion_tokens = completion_tokens_of(resp) or 0
+    assert isinstance(content, str), (
+        f"the retry should have returned a completion body, got content {content!r} (body={resp.body[:300]})"
+    )
+    assert content or (finish_reason == "length" and completion_tokens > 0), (
+        f"the retry returned empty content with finish_reason={finish_reason!r}, "
+        f"completion_tokens={completion_tokens}; empty content is only acceptable when the budget "
+        f"was spent on non-visible reasoning (body={resp.body[:300]})"
+    )
+
+
 class TestReliabilityRetries:
     @pytest.mark.covers("reliability.retry.timeout.succeeds_within_retries")
     def test_timeout_on_first_deployment_succeeds_on_retry(
@@ -49,25 +81,27 @@ class TestReliabilityRetries:
             override=RouterSettingsOverride(num_retries=2),
         )
 
-        assert resp.status_code == 200, (
-            f"the retry should have landed on the healthy backup, got {resp.status_code}: {resp.body[:300]}"
+        assert_retry_landed_on_backup(resp)
+
+    @pytest.mark.covers("reliability.retry.context_window.succeeds_within_retries")
+    def test_context_window_refusal_on_first_deployment_succeeds_on_retry(
+        self, client: ComplexityRouterClient, resources: ResourceManager, scoped_key: str
+    ) -> None:
+        group = f"reliability-retry-{unique_marker()}"
+        small_context = create_always_picked_small_context_deployment(client.proxy, group)
+        resources.defer(lambda: client.proxy.delete_model(small_context))
+        backup = create_zero_weight_backup_deployment(client.proxy, group)
+        resources.defer(lambda: client.proxy.delete_model(backup))
+
+        resp = chat_override(
+            client.proxy,
+            scoped_key,
+            group,
+            oversized_prompt(unique_marker()),
+            override=RouterSettingsOverride(
+                num_retries=2,
+                model_group_retry_policy={group: {"BadRequestErrorRetries": 2}},
+            ),
         )
 
-        attempted = resp.headers.get("x-litellm-attempted-retries")
-        assert attempted is not None, "response is missing the x-litellm-attempted-retries header"
-        assert int(attempted) >= 1, (
-            f"x-litellm-attempted-retries is {attempted!r}; a 200 with no retry means the request never "
-            "opened on the timing-out deployment, so this proves nothing about retries"
-        )
-
-        content = content_of(resp)
-        finish_reason = finish_reason_of(resp)
-        completion_tokens = completion_tokens_of(resp) or 0
-        assert isinstance(content, str), (
-            f"the retry should have returned a completion body, got content {content!r} (body={resp.body[:300]})"
-        )
-        assert content or (finish_reason == "length" and completion_tokens > 0), (
-            f"the retry returned empty content with finish_reason={finish_reason!r}, "
-            f"completion_tokens={completion_tokens}; empty content is only acceptable when the budget "
-            f"was spent on non-visible reasoning (body={resp.body[:300]})"
-        )
+        assert_retry_landed_on_backup(resp)
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 7c044310e14..12c1516837a 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13116,6 +13116,7 @@ async def test_prompt_management_factory_marks_injection_for_every_deployment(mo
         ({"DefaultRetries": 0}, 502, litellm.BadGatewayError, 1),
         ({"DefaultRetries": 0, "ServiceUnavailableErrorRetries": 1}, 503, litellm.ServiceUnavailableError, 2),
         ({"ServiceUnavailableErrorRetries": 0}, 502, litellm.BadGatewayError, 3),
+        ({"BadRequestErrorRetries": 2}, 400, litellm.BadRequestError, 3),
     ],
 )
 async def test_router_retry_policy_controls_upstream_attempt_count(
@@ -13152,6 +13153,87 @@ async def test_router_retry_policy_controls_upstream_attempt_count(
     assert upstream.call_count == expected_upstream_calls
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "retry_policy,upstream_error",
+    [
+        (
+            {"BadRequestErrorRetries": 2},
+            {
+                "message": "This model's maximum context length is 16385 tokens",
+                "type": "invalid_request_error",
+                "code": "context_length_exceeded",
+            },
+        ),
+        (
+            {"ContentPolicyViolationErrorRetries": 2},
+            {
+                "message": "Your request was rejected as a result of our safety system",
+                "type": "invalid_request_error",
+                "code": "content_policy_violation",
+            },
+        ),
+    ],
+)
+async def test_router_retry_policy_400_retries_on_sibling_deployment(
+    monkeypatch: pytest.MonkeyPatch, retry_policy, upstream_error
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-5.6",
+                "litellm_params": {
+                    "model": "openai/gpt-5.6",
+                    "api_key": "sk-fake",
+                    "api_base": "https://rejecting.local/v1",
+                    "weight": 1,
+                },
+                "model_info": {"id": "rejecting"},
+            },
+            {
+                "model_name": "gpt-5.6",
+                "litellm_params": {
+                    "model": "openai/gpt-5.6",
+                    "api_key": "sk-fake",
+                    "api_base": "https://accepting.local/v1",
+                    "weight": 0,
+                },
+                "model_info": {"id": "accepting"},
+            },
+        ],
+        num_retries=2,
+        retry_policy=retry_policy,
+        disable_cooldowns=True,
+    )
+
+    with respx.mock(assert_all_called=True) as respx_mock:
+        rejecting = respx_mock.post("https://rejecting.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": upstream_error})
+        )
+        accepting = respx_mock.post("https://accepting.local/v1/chat/completions").mock(
+            return_value=httpx.Response(
+                200,
+                json={
+                    "id": "chatcmpl-lit-7036",
+                    "object": "chat.completion",
+                    "created": 1,
+                    "model": "gpt-5.6",
+                    "choices": [
+                        {"index": 0, "message": {"role": "assistant", "content": "hi back"}, "finish_reason": "stop"}
+                    ],
+                    "usage": {"prompt_tokens": 1, "completion_tokens": 2, "total_tokens": 3},
+                },
+            )
+        )
+        response = await router.acompletion(model="gpt-5.6", messages=[{"role": "user", "content": "hi"}])
+
+    assert rejecting.call_count == 1
+    assert accepting.call_count == 1
+    assert response.choices[0].message.content == "hi back"
+    assert response._hidden_params["additional_headers"]["x-litellm-attempted-retries"] == 1
+
+
 def _make_failure_logging_obj():
     return LiteLLMLogging(
         model="gpt-5.6",
diff --git a/tests/test_litellm/test_router_order_fallback.py b/tests/test_litellm/test_router_order_fallback.py
index fde870e5abe..93895bbde08 100644
--- a/tests/test_litellm/test_router_order_fallback.py
+++ b/tests/test_litellm/test_router_order_fallback.py
@@ -18,10 +18,10 @@ from litellm import Router
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.router_utils.prompt_caching_cache import PromptCachingCache
 from litellm.types.router import RouterRateLimitError
-from litellm.utils import _get_deployment_order, _get_order_filtered_deployments
+from litellm.utils import _get_deployment_order, get_order_filtered_deployments
 
 # ---------------------------------------------------------------------------
-# Unit tests for _get_order_filtered_deployments
+# Unit tests for get_order_filtered_deployments
 # ---------------------------------------------------------------------------
 
 
@@ -42,7 +42,7 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(2, "b"),
             self._make_deployment(1, "c"),
         ]
-        result = _get_order_filtered_deployments(deps)
+        result = get_order_filtered_deployments(deps)
         assert len(result) == 2
         assert all(d["model_info"]["id"] in ("a", "c") for d in result)
 
@@ -52,7 +52,7 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(2, "b"),
             self._make_deployment(3, "c"),
         ]
-        result = _get_order_filtered_deployments(deps, target_order=2)
+        result = get_order_filtered_deployments(deps, target_order=2)
         assert len(result) == 1
         assert result[0]["model_info"]["id"] == "b"
 
@@ -61,7 +61,7 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(1, "a"),
             self._make_deployment(2, "b"),
         ]
-        result = _get_order_filtered_deployments(deps, target_order=99)
+        result = get_order_filtered_deployments(deps, target_order=99)
         assert result == []
 
     def test_target_order_no_match_does_not_reselect_lower_order(self):
@@ -70,7 +70,7 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(2, "b"),
         ]
         remaining_after_pre_call = [deps[0]]
-        result = _get_order_filtered_deployments(remaining_after_pre_call, target_order=2)
+        result = get_order_filtered_deployments(remaining_after_pre_call, target_order=2)
         assert result == []
 
     def test_no_order_set_returns_all(self):
@@ -78,11 +78,11 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(None, "a"),
             self._make_deployment(None, "b"),
         ]
-        result = _get_order_filtered_deployments(deps)
+        result = get_order_filtered_deployments(deps)
         assert len(result) == 2
 
     def test_empty_list(self):
-        result = _get_order_filtered_deployments([])
+        result = get_order_filtered_deployments([])
         assert result == []
 
     def test_single_order_returns_all_with_that_order(self):
@@ -90,7 +90,7 @@ class TestGetOrderFilteredDeployments:
             self._make_deployment(1, "a"),
             self._make_deployment(1, "b"),
         ]
-        result = _get_order_filtered_deployments(deps)
+        result = get_order_filtered_deployments(deps)
         assert len(result) == 2
 
 
diff --git a/tests/test_litellm/test_router_weighted_failover.py b/tests/test_litellm/test_router_weighted_failover.py
index 162312a8c67..9f05654f23f 100644
--- a/tests/test_litellm/test_router_weighted_failover.py
+++ b/tests/test_litellm/test_router_weighted_failover.py
@@ -15,11 +15,11 @@ import pytest
 
 import litellm
 from litellm import Router
-from litellm.utils import _get_excluded_filtered_deployments
+from litellm.utils import get_excluded_filtered_deployments
 
 
 # ---------------------------------------------------------------------------
-# Unit tests for _get_excluded_filtered_deployments
+# Unit tests for get_excluded_filtered_deployments
 # ---------------------------------------------------------------------------
 
 
@@ -37,17 +37,17 @@ def _make_dep(dep_id: str, weight: Optional[int] = None) -> dict:
 class TestGetExcludedFilteredDeployments:
     def test_no_excluded_returns_all(self):
         deps = [_make_dep("a"), _make_dep("b")]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=None)
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=None)
         assert len(result) == 2
 
     def test_empty_excluded_returns_all(self):
         deps = [_make_dep("a"), _make_dep("b")]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=[])
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=[])
         assert len(result) == 2
 
     def test_drops_excluded(self):
         deps = [_make_dep("a"), _make_dep("b"), _make_dep("c")]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=["b"])
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=["b"])
         ids = sorted(d["model_info"]["id"] for d in result)
         assert ids == ["a", "c"]
 
@@ -57,12 +57,12 @@ class TestGetExcludedFilteredDeployments:
         # error. Returning the original list here would re-include the
         # just-failed deployment and let weighted failover re-pick it.
         deps = [_make_dep("a"), _make_dep("b")]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=["a", "b"])
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=["a", "b"])
         assert result == []
 
     def test_excluded_set_with_unknown_ids(self):
         deps = [_make_dep("a"), _make_dep("b")]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=["zzz"])
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=["zzz"])
         assert len(result) == 2
 
     def test_handles_missing_model_info(self):
@@ -70,7 +70,7 @@ class TestGetExcludedFilteredDeployments:
             {"model_name": "x", "litellm_params": {"model": "gpt-4o"}},  # no model_info
             _make_dep("b"),
         ]
-        result = _get_excluded_filtered_deployments(deps, excluded_deployment_ids=["b"])
+        result = get_excluded_filtered_deployments(deps, excluded_deployment_ids=["b"])
         assert len(result) == 1
 
 

From 8da43835a6be47f58a45ce2967c43f50867401c9 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 22:26:09 -0700
Subject: [PATCH 42/79] test(ci): guard against two workflow jobs publishing
 one check-run name

A ruleset's required status check names a check run and GitHub matches it by
that name alone, so two jobs publishing the same name leave the gate unable to
say which job proved it. The new code-quality check reads every workflow,
expands matrix values and local reusable-workflow calls the way Actions does,
and fails when one name has more than one job behind it.
---
 .github/workflows/test-code-quality.yml       |   6 +
 .../check_workflow_job_name_collisions.py     | 179 +++++++++++++++
 .../test_workflow_job_name_collisions.py      | 211 ++++++++++++++++++
 3 files changed, 396 insertions(+)
 create mode 100644 tests/code_coverage_tests/check_workflow_job_name_collisions.py
 create mode 100644 tests/code_coverage_tests/test_workflow_job_name_collisions.py

diff --git a/.github/workflows/test-code-quality.yml b/.github/workflows/test-code-quality.yml
index 36fd656c231..e6d2264fbf0 100644
--- a/.github/workflows/test-code-quality.yml
+++ b/.github/workflows/test-code-quality.yml
@@ -74,6 +74,12 @@ jobs:
       - name: check_workflow_startup_safety
         run: uv run --no-sync python ./tests/code_coverage_tests/check_workflow_startup_safety.py
 
+      - name: check_workflow_job_name_collisions
+        run: uv run --no-sync python ./tests/code_coverage_tests/check_workflow_job_name_collisions.py
+
+      - name: test_workflow_job_name_collisions
+        run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_workflow_job_name_collisions.py
+
       - name: test_e2e_changed_gate
         run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_e2e_changed_gate.py
 
diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
new file mode 100644
index 00000000000..42e11a3e81d
--- /dev/null
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -0,0 +1,179 @@
+"""Catch two workflow jobs that publish check runs under the same name.
+
+A ruleset's required status check names a check run and GitHub matches it by that
+name alone. When two jobs publish the same name the required context stops
+mapping to the job that proves it: the commit carries two check runs under one
+name and nothing says which one the ruleset required. Both being green hides the
+clash completely, so the context quietly stops meaning what the ruleset intended.
+
+`.github/workflows/auto-close-duplicates.yml` shipped a job id `test` while
+`.github/workflows/test-mcp.yml` already published the required `test` context,
+and commit ed5761daef4ae17152446d182c860630c38b7268 carried both check runs.
+This invariant has to be enforced here because CI cannot enforce it on itself.
+
+A job publishes its `name:` when it sets one and its job id otherwise. Two shapes
+expand that further. A `${{ matrix.key }}` reference publishes one check run per
+value the matrix supplies, so two shard lists that overlap collide even though
+their templates read differently. A job calling a local reusable workflow
+publishes one check run per job of the callee, named `<caller> / <callee>`, which
+is why a caller's name never collides with a plain job that happens to match it.
+A reference nothing resolves stays in the string, so two jobs carrying the same
+unresolved template still compare equal and their collision is still caught.
+"""
+
+import itertools
+import operator
+import re
+import sys
+from collections.abc import Iterator, Mapping, Sequence
+from pathlib import Path
+from typing import Final
+
+import yaml
+from pydantic import BaseModel, Field, ValidationError
+
+REPO_ROOT: Final = Path(__file__).resolve().parent.parent.parent
+WORKFLOWS_DIR: Final = REPO_ROOT / ".github" / "workflows"
+MATRIX_REF: Final = re.compile(r"\$\{\{\s*matrix\.(?P<key>[\w-]+)\s*\}\}")
+LOCAL_CALL_PREFIX: Final = "./"
+
+
+class CheckRunNameCollision(Exception):
+    pass
+
+
+class Job(BaseModel):
+    name: str | None = None
+    uses: str | None = None
+    strategy: Mapping[str, object] = Field(default_factory=dict)
+
+
+class Workflow(BaseModel):
+    jobs: Mapping[str, Job] = Field(default_factory=dict)
+
+
+def parse(text: str) -> tuple[Workflow, object] | None:
+    """The workflow plus its raw `on:` value, or None when the file is not a workflow."""
+    parsed: Final = yaml.safe_load(text)
+    if not isinstance(parsed, dict):
+        return None
+    try:
+        return Workflow.model_validate(parsed), parsed.get(True, parsed.get("on"))
+    except ValidationError:
+        return None
+
+
+def events(raw_on: object) -> frozenset[str]:
+    if isinstance(raw_on, Mapping):
+        return frozenset(str(key) for key in raw_on)
+    if isinstance(raw_on, str):
+        return frozenset({raw_on})
+    if isinstance(raw_on, Sequence):
+        return frozenset(str(event) for event in raw_on)
+    return frozenset()
+
+
+def publishes_check_runs(raw_on: object) -> bool:
+    """A `workflow_call`-only workflow posts its check runs through callers, never itself."""
+    return events(raw_on) != frozenset({"workflow_call"})
+
+
+def matrix_values(job: Job, key: str) -> tuple[str, ...]:
+    matrix: Final = job.strategy.get("matrix")
+    if not isinstance(matrix, Mapping):
+        return ()
+    listed: Final = matrix.get(key)
+    rows: Final = matrix.get("include")
+    from_list: Final = (
+        tuple(str(value) for value in listed if isinstance(value, (str, int, float)))
+        if isinstance(listed, Sequence) and not isinstance(listed, str)
+        else ()
+    )
+    from_rows: Final = (
+        tuple(str(row[key]) for row in rows if isinstance(row, Mapping) and isinstance(row.get(key), (str, int, float)))
+        if isinstance(rows, Sequence) and not isinstance(rows, str)
+        else ()
+    )
+    return tuple(dict.fromkeys(from_list + from_rows))
+
+
+def substituted(template: str, values: Mapping[str, str]) -> str:
+    return MATRIX_REF.sub(lambda ref: values.get(ref.group("key"), ref.group(0)), template)
+
+
+def expand(template: str, job: Job) -> tuple[str, ...]:
+    keys: Final = tuple(dict.fromkeys(ref.group("key") for ref in MATRIX_REF.finditer(template)))
+    candidates: Final = tuple((key, matrix_values(job, key)) for key in keys)
+    resolvable: Final = tuple((key, values) for key, values in candidates if values)
+    if not resolvable:
+        return (template,)
+    return tuple(
+        substituted(template, dict(zip((key for key, _ in resolvable), combination)))
+        for combination in itertools.product(*(values for _, values in resolvable))
+    )
+
+
+def callee_path(job: Job) -> str | None:
+    if job.uses is None or not job.uses.startswith(LOCAL_CALL_PREFIX):
+        return None
+    return job.uses[len(LOCAL_CALL_PREFIX) :].split("@")[0]
+
+
+def job_names(job_id: str, job: Job, workflows: Mapping[str, Workflow]) -> tuple[str, ...]:
+    prefixes: Final = expand(job.name or job_id, job)
+    callee: Final = workflows.get(callee_path(job) or "")
+    if callee is None:
+        return prefixes
+    suffixes: Final = tuple(
+        name
+        for callee_id, callee_job in callee.jobs.items()
+        for name in expand(callee_job.name or callee_id, callee_job)
+    )
+    return tuple(f"{prefix} / {suffix}" for prefix in prefixes for suffix in suffixes)
+
+
+def published(sources: Mapping[str, str]) -> Iterator[tuple[str, str]]:
+    parsed: Final = {rel: entry for rel, text in sources.items() if (entry := parse(text)) is not None}
+    workflows: Final = {rel: workflow for rel, (workflow, _) in parsed.items()}
+    for rel, (workflow, raw_on) in parsed.items():
+        if not publishes_check_runs(raw_on):
+            continue
+        for job_id, job in workflow.jobs.items():
+            for name in job_names(job_id, job, workflows):
+                yield name, f"{rel} job `{job_id}`"
+
+
+def owners_by_name(sources: Mapping[str, str]) -> Iterator[tuple[str, tuple[str, ...]]]:
+    for name, pairs in itertools.groupby(sorted(published(sources)), key=operator.itemgetter(0)):
+        yield name, tuple(owner for _, owner in pairs)
+
+
+def collisions(sources: Mapping[str, str]) -> tuple[str, ...]:
+    return tuple(
+        f"`{name}` is published by {len(owners)} jobs: {', '.join(owners)}. A required status check matching "
+        f"that name cannot say which job proves it; give one of them a distinct `name:` or job id."
+        for name, owners in owners_by_name(sources)
+        if len(owners) > 1
+    )
+
+
+def workflow_sources() -> Mapping[str, str]:
+    """Repo-relative posix paths to text, the keys `uses: ./...` resolves against."""
+    return {path.relative_to(REPO_ROOT).as_posix(): path.read_text() for path in sorted(WORKFLOWS_DIR.glob("*.y*ml"))}
+
+
+def main() -> None:
+    sources: Final = workflow_sources()
+    found: Final = collisions(sources)
+    if found:
+        raise CheckRunNameCollision("Check-run names are not unique:\n  - " + "\n  - ".join(found))
+
+    print(f"Check-run names are unique across {len(sources)} workflows")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except CheckRunNameCollision as exc:
+        print(f"ERROR: {exc}", file=sys.stderr)
+        sys.exit(1)
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
new file mode 100644
index 00000000000..35c1480ecd8
--- /dev/null
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -0,0 +1,211 @@
+import sys
+from pathlib import Path
+from typing import Final
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+
+from check_workflow_job_name_collisions import callee_path, collisions, parse, workflow_sources  # noqa: E402
+
+
+def test_every_workflow_in_the_repo_publishes_a_unique_check_run_name() -> None:
+    assert collisions(workflow_sources()) == ()
+
+
+def test_every_workflow_in_the_repo_parses_into_jobs() -> None:
+    unparsed: Final = tuple(
+        rel for rel, source in workflow_sources().items() if (entry := parse(source)) is None or not entry[0].jobs
+    )
+
+    assert unparsed == ()
+
+
+def test_every_local_reusable_call_in_the_repo_resolves_to_a_workflow() -> None:
+    sources: Final = workflow_sources()
+    parsed: Final = tuple(entry for text in sources.values() if (entry := parse(text)) is not None)
+    unresolved: Final = tuple(
+        job.uses
+        for workflow, _ in parsed
+        for job in workflow.jobs.values()
+        if (callee := callee_path(job)) is not None and callee not in sources
+    )
+
+    assert unresolved == ()
+
+
+def test_two_jobs_falling_back_to_the_same_job_id_collide() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+        "b.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`test` is published by 2 jobs" in found[0]
+    assert "a.yml job `test`" in found[0] and "b.yml job `test`" in found[0]
+
+
+def test_an_explicit_name_overrides_the_job_id_and_clears_the_collision() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  test:\n    name: Sweep tests\n    runs-on: ubuntu-latest\n",
+        "b.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+
+
+def test_an_explicit_name_matching_another_job_id_collides() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  sweep:\n    name: test\n    runs-on: ubuntu-latest\n",
+        "b.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`test` is published by 2 jobs" in found[0]
+
+
+def test_two_callers_of_one_reusable_workflow_collide_on_a_shared_matrix_value() -> None:
+    base: Final = "on:\n  workflow_call:\njobs:\n  run:\n    name: Run tests\n    runs-on: ubuntu-latest\n"
+    caller: Final = (
+        "on: pull_request\n"
+        "jobs:\n"
+        "  {job}:\n"
+        "    name: ${{{{ matrix.shard }}}}\n"
+        "    uses: ./.github/workflows/base.yml\n"
+        "    strategy:\n"
+        "      matrix:\n"
+        "        include:\n"
+        "          - shard: {shard}\n"
+    )
+    sources: Final = {
+        ".github/workflows/base.yml": base,
+        "unit.yml": caller.format(job="unit", shard="proxy-auth"),
+        "proxy-db.yml": caller.format(job="proxy-db", shard="proxy-auth"),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`proxy-auth / Run tests` is published by 2 jobs" in found[0]
+
+
+def test_distinct_matrix_values_through_one_reusable_workflow_do_not_collide() -> None:
+    base: Final = "on:\n  workflow_call:\njobs:\n  run:\n    name: Run tests\n    runs-on: ubuntu-latest\n"
+    caller: Final = (
+        "on: pull_request\n"
+        "jobs:\n"
+        "  {job}:\n"
+        "    name: ${{{{ matrix.shard }}}}\n"
+        "    uses: ./.github/workflows/base.yml\n"
+        "    strategy:\n"
+        "      matrix:\n"
+        "        include:\n"
+        "          - shard: {shard}\n"
+    )
+    sources: Final = {
+        ".github/workflows/base.yml": base,
+        "unit.yml": caller.format(job="unit", shard="proxy-auth"),
+        "proxy-db.yml": caller.format(job="proxy-db", shard="budgets"),
+    }
+
+    assert collisions(sources) == ()
+
+
+def test_a_reusable_caller_does_not_collide_with_a_plain_job_of_the_same_name() -> None:
+    sources: Final = {
+        ".github/workflows/base.yml": (
+            "on:\n  workflow_call:\njobs:\n  run:\n    name: Run tests\n    runs-on: ubuntu-latest\n"
+        ),
+        "unit.yml": (
+            "on: pull_request\n"
+            "jobs:\n"
+            "  unit:\n"
+            "    name: ${{ matrix.shard }}\n"
+            "    uses: ./.github/workflows/base.yml\n"
+            "    strategy:\n"
+            "      matrix:\n"
+            "        include:\n"
+            "          - shard: proxy-behavior\n"
+        ),
+        "postgres.yml": (
+            "on: pull_request\n"
+            "jobs:\n"
+            "  postgres:\n"
+            "    name: ${{ matrix.shard }}\n"
+            "    runs-on: ubuntu-latest\n"
+            "    strategy:\n"
+            "      matrix:\n"
+            "        include:\n"
+            "          - shard: proxy-behavior\n"
+        ),
+    }
+
+    assert collisions(sources) == ()
+
+
+def test_a_workflow_call_only_workflow_publishes_nothing_of_its_own() -> None:
+    sources: Final = {
+        "base.yml": "on:\n  workflow_call:\njobs:\n  run:\n    runs-on: ubuntu-latest\n",
+        "other.yml": "on:\n  workflow_call:\njobs:\n  run:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+
+
+def test_a_workflow_call_workflow_that_also_runs_on_pull_request_still_publishes() -> None:
+    sources: Final = {
+        "base.yml": "on:\n  workflow_call:\n  pull_request:\njobs:\n  run:\n    runs-on: ubuntu-latest\n",
+        "other.yml": "on: pull_request\njobs:\n  run:\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`run` is published by 2 jobs" in found[0]
+
+
+def test_a_matrix_list_supplies_values_the_same_way_include_rows_do() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\n"
+            "jobs:\n"
+            "  build:\n"
+            "    name: Analyze (${{ matrix.language }})\n"
+            "    runs-on: ubuntu-latest\n"
+            "    strategy:\n"
+            "      matrix:\n"
+            "        language: [python, go]\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  go:\n    name: Analyze (go)\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`Analyze (go)` is published by 2 jobs" in found[0]
+
+
+def test_two_jobs_sharing_one_unresolvable_template_still_collide() -> None:
+    template: Final = (
+        "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ matrix.shard }}}}\n    runs-on: ubuntu-latest\n"
+    )
+    sources: Final = {
+        "a.yml": template.format(job="one"),
+        "b.yml": template.format(job="two"),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "is published by 2 jobs" in found[0]
+
+
+def test_a_file_that_is_not_a_workflow_is_ignored() -> None:
+    sources: Final = {
+        "notes.yml": "just a string\n",
+        "a.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()

From 44b7a76439f1896736e61679b4cd0b5e66dd297a Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 22:26:38 -0700
Subject: [PATCH 43/79] fix(ci): stop the auto-close duplicates job colliding
 with the required test check

Its job id was `test`, the same check-run name test-mcp.yml publishes and
guard-internal-staging requires, so a commit touching the sweep carried two
check runs called `test`. Renaming it to `sweep-tests` leaves the required
context with exactly one job behind it.
---
 .github/workflows/auto-close-duplicates.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/auto-close-duplicates.yml b/.github/workflows/auto-close-duplicates.yml
index d8256917805..9a362b4c85c 100644
--- a/.github/workflows/auto-close-duplicates.yml
+++ b/.github/workflows/auto-close-duplicates.yml
@@ -22,7 +22,7 @@ on:
 permissions: {}
 
 jobs:
-  test:
+  sweep-tests:
     if: github.event_name == 'pull_request'
     runs-on: ubuntu-latest
     timeout-minutes: 5

From 7c7810df42a6f90f213b9998e9897da81b8cfb25 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 22:43:15 -0700
Subject: [PATCH 44/79] fix(router): ignore non-integer status codes when
 picking retry skips

CI's router_code_coverage gate wants every function in router.py called by
name from a test file with "router" in its name, and the new helper had no
direct caller, so the check-quality job failed on the first tip.

Covering it directly also turned up a hole. litellm._should_retry compares
the status code to 500, so a provider exception carrying a string status code
raises TypeError instead of answering. should_retry_this_error has the same
call, but the retry policy path skips it, which is exactly the path this
change enables, so the helper was the first to touch that value. Narrowing to
int leaves those exceptions on the old retry-in-place behavior.
---
 litellm/router.py                 |  2 +-
 tests/test_litellm/test_router.py | 27 +++++++++++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/litellm/router.py b/litellm/router.py
index f72f41313a0..6f1f5696847 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -7467,7 +7467,7 @@ class Router:
     ) -> tuple[str, ...]:
         failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
         status_code: Final = getattr(exception, "status_code", None)
-        if not failed_deployment_id or status_code is None:
+        if not failed_deployment_id or not isinstance(status_code, int):
             return ()
         if litellm._should_retry(status_code):  # pyright: ignore[reportPrivateUsage]  # as in should_retry_this_error
             return ()
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 12c1516837a..4bf75ad408c 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13234,6 +13234,33 @@ async def test_router_retry_policy_400_retries_on_sibling_deployment(
     assert response._hidden_params["additional_headers"]["x-litellm-attempted-retries"] == 1
 
 
+@pytest.mark.parametrize(
+    "status_code,failed_deployment_id,already_skipped,healthy_deployment_ids,expected",
+    [
+        (400, "rejecting", None, ["rejecting", "accepting"], ("rejecting",)),
+        (403, "rejecting", None, ["rejecting", "accepting"], ("rejecting",)),
+        (400, "second", ("first",), ["first", "second", "third"], ("first", "second")),
+        (429, "rejecting", None, ["rejecting", "accepting"], ()),
+        (503, "rejecting", None, ["rejecting", "accepting"], ()),
+        (400, "rejecting", None, ["rejecting"], ()),
+        (400, None, None, ["rejecting", "accepting"], ()),
+        (None, "rejecting", None, ["rejecting", "accepting"], ()),
+        ("400", "rejecting", None, ["rejecting", "accepting"], ()),
+    ],
+)
+def test_router_deployment_ids_to_skip_on_retry(
+    status_code, failed_deployment_id, already_skipped, healthy_deployment_ids, expected
+):
+    exception = Exception("upstream refused this request")
+    exception.status_code = status_code
+    exception.failed_deployment_id = failed_deployment_id
+    healthy_deployments = [{"model_info": {"id": deployment_id}} for deployment_id in healthy_deployment_ids]
+
+    assert (
+        litellm.Router._deployment_ids_to_skip_on_retry(exception, already_skipped, healthy_deployments) == expected
+    )
+
+
 def _make_failure_logging_obj():
     return LiteLLMLogging(
         model="gpt-5.6",

From baee7d8175a1e46085a477f45257dc8b9006b4c1 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 22:47:23 -0700
Subject: [PATCH 45/79] fix: evaluate job name expressions in the check-run
 collision guard

The guard only substituted a bare `${{ matrix.key }}`, so any name built from a
larger expression stayed in the string as its own template. `_test-unit-base.yml`
names its job with a ternary over `format()`, which meant every shard published
an opaque name and 23 of the 33 required contexts, all of them `<shard> / Run
tests`, were invisible to the very check meant to protect them.

Job names are now evaluated per matrix combination over the pieces a name can
hold: string literals, `matrix.<key>`, `format()`, `==` and `!=`, and the
`<cond> && <a> || <b>` idiom. All 33 required contexts now resolve, and nothing
in the repo leaves an expression unresolved. An expression the evaluator does not
understand still falls back to its verbatim template, so two jobs sharing one
stays a collision.

The test also drops its `sys.path.insert`, which the test-quality budget counts
under TQ003; pytest already puts the file's own directory on the path.
---
 .../check_workflow_job_name_collisions.py     | 109 +++++++++++++++---
 .../test_workflow_job_name_collisions.py      |  63 +++++++++-
 2 files changed, 152 insertions(+), 20 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 42e11a3e81d..849a906c3b8 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -12,13 +12,17 @@ and commit ed5761daef4ae17152446d182c860630c38b7268 carried both check runs.
 This invariant has to be enforced here because CI cannot enforce it on itself.
 
 A job publishes its `name:` when it sets one and its job id otherwise. Two shapes
-expand that further. A `${{ matrix.key }}` reference publishes one check run per
-value the matrix supplies, so two shard lists that overlap collide even though
-their templates read differently. A job calling a local reusable workflow
-publishes one check run per job of the callee, named `<caller> / <callee>`, which
-is why a caller's name never collides with a plain job that happens to match it.
-A reference nothing resolves stays in the string, so two jobs carrying the same
-unresolved template still compare equal and their collision is still caught.
+expand that further. A name carrying `${{ ... }}` publishes one check run per
+matrix combination it reads, so two shard lists that overlap collide even though
+their templates read differently. Each expression is evaluated per combination
+over the pieces a job name can hold: string literals, `matrix.<key>`, `format()`,
+`==` and `!=`, and the `<cond> && <a> || <b>` idiom, which is how the shards
+reach their real `<shard> / Run tests` names rather than staying opaque. A job
+calling a local reusable workflow publishes one check run per job of the callee,
+named `<caller> / <callee>`, which is why a caller's name never collides with a
+plain job that happens to match it. An expression nothing resolves stays in the
+string, so two jobs carrying the same unresolved template still compare equal and
+their collision is still caught.
 """
 
 import itertools
@@ -27,6 +31,7 @@ import re
 import sys
 from collections.abc import Iterator, Mapping, Sequence
 from pathlib import Path
+from types import MappingProxyType
 from typing import Final
 
 import yaml
@@ -34,7 +39,13 @@ from pydantic import BaseModel, Field, ValidationError
 
 REPO_ROOT: Final = Path(__file__).resolve().parent.parent.parent
 WORKFLOWS_DIR: Final = REPO_ROOT / ".github" / "workflows"
-MATRIX_REF: Final = re.compile(r"\$\{\{\s*matrix\.(?P<key>[\w-]+)\s*\}\}")
+EXPRESSION: Final = re.compile(r"\$\{\{(?P<body>.*?)\}\}", re.DOTALL)
+MATRIX_KEY: Final = re.compile(r"matrix\.(?P<key>[\w-]+)")
+MATRIX_REF: Final = re.compile(r"^matrix\.(?P<key>[\w-]+)$")
+LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
+FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
+COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
+NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
 LOCAL_CALL_PREFIX: Final = "./"
 
 
@@ -97,19 +108,85 @@ def matrix_values(job: Job, key: str) -> tuple[str, ...]:
     return tuple(dict.fromkeys(from_list + from_rows))
 
 
-def substituted(template: str, values: Mapping[str, str]) -> str:
-    return MATRIX_REF.sub(lambda ref: values.get(ref.group("key"), ref.group(0)), template)
+def scanned(state: tuple[int, bool], char: str) -> tuple[int, bool]:
+    depth, quoted = state
+    if char == "'":
+        return depth, not quoted
+    if quoted:
+        return depth, quoted
+    return depth + int(char == "(") - int(char == ")"), quoted
+
+
+def split_outside(text: str, token: str) -> tuple[str, ...]:
+    """`text` cut on every `token` that sits outside quotes and parentheses."""
+    states: Final = tuple(itertools.accumulate(text, scanned, initial=(0, False)))
+    cuts: Final = tuple(
+        index
+        for index in range(len(text) - len(token) + 1)
+        if text.startswith(token, index) and states[index] == (0, False)
+    )
+    starts: Final = (0, *(cut + len(token) for cut in cuts))
+    return tuple(text[start:end] for start, end in zip(starts, (*cuts, len(text))))
+
+
+def value_of(text: str, values: Mapping[str, str]) -> str | None:
+    expression: Final = text.strip()
+    literal: Final = LITERAL.match(expression)
+    if literal is not None:
+        return literal.group("text")
+    reference: Final = MATRIX_REF.match(expression)
+    if reference is not None:
+        return values.get(reference.group("key"))
+    call: Final = FORMAT_CALL.match(expression)
+    if call is None:
+        return None
+    arguments: Final = tuple(value_of(part, values) for part in split_outside(call.group("args"), ","))
+    resolved: Final = tuple(argument for argument in arguments if argument is not None)
+    if not resolved or len(resolved) != len(arguments):
+        return None
+    return resolved[0].format(*resolved[1:])
+
+
+def holds(condition: str, values: Mapping[str, str]) -> bool | None:
+    comparison: Final = COMPARISON.match(condition.strip())
+    if comparison is None:
+        return None
+    left: Final = value_of(comparison.group("left"), values)
+    right: Final = value_of(comparison.group("right"), values)
+    if left is None or right is None:
+        return None
+    return (left == right) == (comparison.group("operator") == "==")
+
+
+def evaluate(body: str, values: Mapping[str, str]) -> str | None:
+    """The single string this expression yields, or None when its shape is not understood."""
+    branches: Final = tuple(split_outside(alternative, "&&") for alternative in split_outside(body, "||"))
+    outcomes: Final = tuple(tuple(holds(part, values) for part in branch[:-1]) for branch in branches)
+    if any(outcome is None for branch in outcomes for outcome in branch):
+        return None
+    taken: Final = next((branch[-1] for branch, outcome in zip(branches, outcomes) if all(outcome)), None)
+    return None if taken is None else value_of(taken, values)
+
+
+def resolved_span(span: re.Match[str], values: Mapping[str, str]) -> str:
+    substitution: Final = evaluate(span.group("body"), values)
+    return span.group(0) if substitution is None else substitution
+
+
+def rendered(template: str, values: Mapping[str, str]) -> str:
+    return EXPRESSION.sub(lambda span: resolved_span(span, values), template)
 
 
 def expand(template: str, job: Job) -> tuple[str, ...]:
-    keys: Final = tuple(dict.fromkeys(ref.group("key") for ref in MATRIX_REF.finditer(template)))
-    candidates: Final = tuple((key, matrix_values(job, key)) for key in keys)
-    resolvable: Final = tuple((key, values) for key, values in candidates if values)
+    keys: Final = tuple(dict.fromkeys(ref.group("key") for ref in MATRIX_KEY.finditer(template)))
+    resolvable: Final = tuple((key, values) for key in keys if (values := matrix_values(job, key)))
     if not resolvable:
-        return (template,)
+        return (rendered(template, NO_MATRIX),)
     return tuple(
-        substituted(template, dict(zip((key for key, _ in resolvable), combination)))
-        for combination in itertools.product(*(values for _, values in resolvable))
+        dict.fromkeys(
+            rendered(template, dict(zip((key for key, _ in resolvable), combination)))
+            for combination in itertools.product(*(values for _, values in resolvable))
+        )
     )
 
 
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 35c1480ecd8..96820d50040 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -1,10 +1,30 @@
-import sys
-from pathlib import Path
 from typing import Final
 
-sys.path.insert(0, str(Path(__file__).resolve().parent))
+from check_workflow_job_name_collisions import callee_path, collisions, parse, published, workflow_sources
 
-from check_workflow_job_name_collisions import callee_path, collisions, parse, workflow_sources  # noqa: E402
+REUSABLE_BASE: Final = """on:
+  workflow_call:
+jobs:
+  run:
+    name: >-
+      ${{ matrix.python-version == '3.12' && 'Run tests'
+      || format('Run tests (Python {0})', matrix.python-version) }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.12", "3.13"]
+"""
+
+SHARD_CALLER: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ matrix.shard }}
+    uses: ./.github/workflows/base.yml
+    strategy:
+      matrix:
+        include:
+          - shard: core-utils
+"""
 
 
 def test_every_workflow_in_the_repo_publishes_a_unique_check_run_name() -> None:
@@ -209,3 +229,38 @@ def test_a_file_that_is_not_a_workflow_is_ignored() -> None:
     }
 
     assert collisions(sources) == ()
+
+
+def test_a_conditional_name_expands_to_the_branch_each_matrix_value_takes() -> None:
+    names: Final = frozenset(
+        name for name, _ in published({".github/workflows/base.yml": REUSABLE_BASE, "unit.yml": SHARD_CALLER})
+    )
+
+    assert names == frozenset({"core-utils / Run tests", "core-utils / Run tests (Python 3.13)"})
+
+
+def test_a_conditional_name_never_publishes_the_branch_its_condition_rules_out() -> None:
+    names: Final = frozenset(
+        name for name, _ in published({".github/workflows/base.yml": REUSABLE_BASE, "unit.yml": SHARD_CALLER})
+    )
+
+    assert "core-utils / Run tests (Python 3.12)" not in names
+
+
+def test_a_conditional_reusable_name_collides_with_a_plain_job_publishing_the_same_name() -> None:
+    sources: Final = {
+        ".github/workflows/base.yml": REUSABLE_BASE,
+        "unit.yml": SHARD_CALLER,
+        "postgres.yml": ("on: pull_request\njobs:\n  legacy:\n    name: core-utils / Run tests\n"),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`core-utils / Run tests` is published by 2 jobs" in found[0]
+
+
+def test_every_workflow_in_the_repo_resolves_every_expression_in_its_job_names() -> None:
+    unresolved: Final = tuple(f"{owner}: {name}" for name, owner in published(workflow_sources()) if "${{" in name)
+
+    assert unresolved == ()

From d6a727fe0f52e8d280235d1cc98c43462643d150 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 23:02:32 -0700
Subject: [PATCH 46/79] fix: keep matrix include rows whole when expanding job
 names

The guard read each matrix key's values independently and crossed them, so
a job name reading two keys off one include row published pairs no job ever
runs, which could fail a valid workflow on a required check

It now builds the combinations GitHub builds: the listed keys crossed, each
include row folded into the combinations it overwrites nothing in, and a row
that fits nowhere standing on its own
---
 .../check_workflow_job_name_collisions.py     | 73 +++++++++++++------
 .../test_workflow_job_name_collisions.py      | 46 ++++++++++++
 2 files changed, 95 insertions(+), 24 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 849a906c3b8..c683de7f08b 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -13,8 +13,9 @@ This invariant has to be enforced here because CI cannot enforce it on itself.
 
 A job publishes its `name:` when it sets one and its job id otherwise. Two shapes
 expand that further. A name carrying `${{ ... }}` publishes one check run per
-matrix combination it reads, so two shard lists that overlap collide even though
-their templates read differently. Each expression is evaluated per combination
+combination the matrix produces, with each `include` row's values held together
+rather than crossed with the other rows', so two shard lists that overlap collide
+even though their templates read differently. Each expression is evaluated per combination
 over the pieces a job name can hold: string literals, `matrix.<key>`, `format()`,
 `==` and `!=`, and the `<cond> && <a> || <b>` idiom, which is how the shards
 reach their real `<shard> / Run tests` names rather than staying opaque. A job
@@ -40,12 +41,13 @@ from pydantic import BaseModel, Field, ValidationError
 REPO_ROOT: Final = Path(__file__).resolve().parent.parent.parent
 WORKFLOWS_DIR: Final = REPO_ROOT / ".github" / "workflows"
 EXPRESSION: Final = re.compile(r"\$\{\{(?P<body>.*?)\}\}", re.DOTALL)
-MATRIX_KEY: Final = re.compile(r"matrix\.(?P<key>[\w-]+)")
 MATRIX_REF: Final = re.compile(r"^matrix\.(?P<key>[\w-]+)$")
 LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
 FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
 COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
 NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
+SCALAR: Final = (str, int, float)
+MATRIX_DIRECTIVES: Final = frozenset({"include", "exclude"})
 LOCAL_CALL_PREFIX: Final = "./"
 
 
@@ -89,23 +91,52 @@ def publishes_check_runs(raw_on: object) -> bool:
     return events(raw_on) != frozenset({"workflow_call"})
 
 
-def matrix_values(job: Job, key: str) -> tuple[str, ...]:
+def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, ...]], ...]:
+    return tuple(
+        (str(key), tuple(str(value) for value in values if isinstance(value, SCALAR)))
+        for key, values in matrix.items()
+        if str(key) not in MATRIX_DIRECTIVES and isinstance(values, Sequence) and not isinstance(values, str)
+    )
+
+
+def include_rows(matrix: Mapping[str, object]) -> tuple[Mapping[str, str], ...]:
+    rows: Final = matrix.get("include")
+    if not isinstance(rows, Sequence) or isinstance(rows, str):
+        return ()
+    return tuple(
+        MappingProxyType({str(key): str(value) for key, value in row.items() if isinstance(value, SCALAR)})
+        for row in rows
+        if isinstance(row, Mapping)
+    )
+
+
+def extends(row: Mapping[str, str], combination: Mapping[str, str]) -> bool:
+    """GitHub folds an `include` row into a combination only where it overwrites no listed value."""
+    return all(combination[key] == value for key, value in row.items() if key in combination)
+
+
+def extended(combination: Mapping[str, str], rows: Sequence[Mapping[str, str]]) -> Mapping[str, str]:
+    additions: Final = {key: value for row in rows if extends(row, combination) for key, value in row.items()}
+    return MappingProxyType({**combination, **additions})
+
+
+def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...]:
+    """One mapping per job the matrix produces, each `include` row's values staying together."""
     matrix: Final = job.strategy.get("matrix")
     if not isinstance(matrix, Mapping):
         return ()
-    listed: Final = matrix.get(key)
-    rows: Final = matrix.get("include")
-    from_list: Final = (
-        tuple(str(value) for value in listed if isinstance(value, (str, int, float)))
-        if isinstance(listed, Sequence) and not isinstance(listed, str)
+    listed: Final = listed_values(matrix)
+    rows: Final = include_rows(matrix)
+    crossed: Final = (
+        tuple(
+            MappingProxyType(dict(zip((key for key, _ in listed), values)))
+            for values in itertools.product(*(values for _, values in listed))
+        )
+        if listed
         else ()
     )
-    from_rows: Final = (
-        tuple(str(row[key]) for row in rows if isinstance(row, Mapping) and isinstance(row.get(key), (str, int, float)))
-        if isinstance(rows, Sequence) and not isinstance(rows, str)
-        else ()
-    )
-    return tuple(dict.fromkeys(from_list + from_rows))
+    standalone: Final = tuple(row for row in rows if not any(extends(row, combination) for combination in crossed))
+    return (*(extended(combination, rows) for combination in crossed), *standalone)
 
 
 def scanned(state: tuple[int, bool], char: str) -> tuple[int, bool]:
@@ -178,16 +209,10 @@ def rendered(template: str, values: Mapping[str, str]) -> str:
 
 
 def expand(template: str, job: Job) -> tuple[str, ...]:
-    keys: Final = tuple(dict.fromkeys(ref.group("key") for ref in MATRIX_KEY.finditer(template)))
-    resolvable: Final = tuple((key, values) for key in keys if (values := matrix_values(job, key)))
-    if not resolvable:
+    combinations: Final = matrix_combinations(job)
+    if not combinations:
         return (rendered(template, NO_MATRIX),)
-    return tuple(
-        dict.fromkeys(
-            rendered(template, dict(zip((key for key, _ in resolvable), combination)))
-            for combination in itertools.product(*(values for _, values in resolvable))
-        )
-    )
+    return tuple(dict.fromkeys(rendered(template, values) for values in combinations))
 
 
 def callee_path(job: Job) -> str | None:
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 96820d50040..359072326bd 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -27,6 +27,33 @@ jobs:
 """
 
 
+CORRELATED_ROWS: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ matrix.shard }} on ${{ matrix.test-path }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        include:
+          - shard: core-utils
+            test-path: tests/core
+          - shard: proxy
+            test-path: tests/proxy
+"""
+
+LISTED_PLUS_ROW: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ matrix.python-version }} ${{ matrix.label }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.12", "3.13"]
+        include:
+          - label: fast
+"""
+
+
 def test_every_workflow_in_the_repo_publishes_a_unique_check_run_name() -> None:
     assert collisions(workflow_sources()) == ()
 
@@ -260,6 +287,25 @@ def test_a_conditional_reusable_name_collides_with_a_plain_job_publishing_the_sa
     assert "`core-utils / Run tests` is published by 2 jobs" in found[0]
 
 
+def test_a_name_reading_two_matrix_keys_publishes_only_the_pairs_each_include_row_holds() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": CORRELATED_ROWS}))
+
+    assert names == frozenset({"core-utils on tests/core", "proxy on tests/proxy"})
+
+
+def test_a_name_reading_two_matrix_keys_never_publishes_a_pair_no_include_row_holds() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": CORRELATED_ROWS}))
+
+    assert "core-utils on tests/proxy" not in names
+    assert "proxy on tests/core" not in names
+
+
+def test_an_include_row_carrying_no_listed_key_extends_every_listed_combination() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": LISTED_PLUS_ROW}))
+
+    assert names == frozenset({"3.12 fast", "3.13 fast"})
+
+
 def test_every_workflow_in_the_repo_resolves_every_expression_in_its_job_names() -> None:
     unresolved: Final = tuple(f"{owner}: {name}" for name, owner in published(workflow_sources()) if "${{" in name)
 

From cb1ec76e46245196092a425e8d759f85e47df547 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 23:02:44 -0700
Subject: [PATCH 47/79] fix(router): keep retry skips on the active
 order-fallback target

The retry-skip guard checks that some other deployment could still answer
before it excludes the one that just refused, so a single-deployment group
keeps the old retry-in-place behavior. It asked that question at the group's
minimum order, but the router picks the retry's deployment at the order the
request has already escalated to.

So a group with a primary at order 1 and a backup at order 2 answered "yes,
order 1 still has a candidate" while the retry was pinned to order 2, and the
exclusion left order 2 with nothing. The caller got a no-deployments error in
place of the provider's own 400.

The helper now takes the active target order and filters by it, which is the
same value async_get_healthy_deployments reads off the request.
---
 litellm/router.py                 |  8 +++++++-
 tests/test_litellm/test_router.py | 30 ++++++++++++++++++++++++++++++
 2 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/litellm/router.py b/litellm/router.py
index 6f1f5696847..4901199902c 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -402,6 +402,7 @@ def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream])
 _NO_SESSION_KWARGS: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({})
 _SESSION_ADAPTER: Final = TypeAdapter(Mapping[str, object])
 _EXCLUDED_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
+_TARGET_ORDER_ADAPTER: Final = TypeAdapter(int | None)
 
 
 def _with_router_resolved_session_model(session: object, model_name: str) -> Mapping[str, Mapping[str, object]]:
@@ -7464,6 +7465,7 @@ class Router:
         exception: Exception,
         already_skipped: object,
         healthy_deployments: list[dict],  # mutable-ok: matches the routing filters' list contract
+        target_order: object = None,
     ) -> tuple[str, ...]:
         failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
         status_code: Final = getattr(exception, "status_code", None)
@@ -7473,7 +7475,9 @@ class Router:
             return ()
         already_skipped_ids: Final = _EXCLUDED_DEPLOYMENT_IDS_ADAPTER.validate_python(already_skipped or ())
         skipped: Final = frozenset((*already_skipped_ids, failed_deployment_id))
-        same_order_candidates: Final = litellm.utils.get_order_filtered_deployments(healthy_deployments)
+        same_order_candidates: Final = litellm.utils.get_order_filtered_deployments(
+            healthy_deployments, target_order=_TARGET_ORDER_ADAPTER.validate_python(target_order)
+        )
         if not litellm.utils.get_excluded_filtered_deployments(same_order_candidates, excluded_deployment_ids=skipped):
             return ()
         verbose_router_logger.debug(
@@ -7580,6 +7584,7 @@ class Router:
                     exception=original_exception,
                     already_skipped=kwargs.get("_excluded_deployment_ids"),
                     healthy_deployments=_healthy_deployments,
+                    target_order=kwargs.get("_target_order"),
                 )
                 if skipped_deployment_ids:
                     kwargs["_excluded_deployment_ids"] = skipped_deployment_ids
@@ -7656,6 +7661,7 @@ class Router:
                         exception=e,
                         already_skipped=kwargs.get("_excluded_deployment_ids"),
                         healthy_deployments=_healthy_deployments,
+                        target_order=kwargs.get("_target_order"),
                     )
                     if retry_skipped_deployment_ids:
                         kwargs["_excluded_deployment_ids"] = retry_skipped_deployment_ids
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 4bf75ad408c..0d7f2e7d3f6 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13261,6 +13261,36 @@ def test_router_deployment_ids_to_skip_on_retry(
     )
 
 
+@pytest.mark.parametrize(
+    "target_order,deployment_orders,expected",
+    [
+        (2, {"rejecting": 2, "sibling": 1}, ()),
+        (2, {"rejecting": 2, "sibling": 2}, ("rejecting",)),
+        (1, {"rejecting": 1, "sibling": 2}, ()),
+        (None, {"rejecting": 1, "sibling": 2}, ()),
+        (None, {"rejecting": 1, "sibling": 1}, ("rejecting",)),
+        (3, {"rejecting": 2, "sibling": 1}, ()),
+    ],
+)
+def test_router_deployment_ids_to_skip_on_retry_honors_order_fallback_target(
+    target_order, deployment_orders, expected
+):
+    exception = Exception("upstream refused this request")
+    exception.status_code = 400
+    exception.failed_deployment_id = "rejecting"
+    healthy_deployments = [
+        {"model_info": {"id": deployment_id}, "litellm_params": {"order": order}}
+        for deployment_id, order in deployment_orders.items()
+    ]
+
+    assert (
+        litellm.Router._deployment_ids_to_skip_on_retry(
+            exception, None, healthy_deployments, target_order=target_order
+        )
+        == expected
+    )
+
+
 def _make_failure_logging_obj():
     return LiteLLMLogging(
         model="gpt-5.6",

From 9f379b36b9c484ff045fadfcec213f47857670db Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 23:18:18 -0700
Subject: [PATCH 48/79] refactor: return the collision check's failure instead
 of raising it

The checker raised a custom exception and caught it two lines down in the
same module, which is the throw-then-catch the repo's coding guide rules
out. `main` now prints the same message and returns the exit code, so the
collision list stays a value the whole way out
---
 .../check_workflow_job_name_collisions.py        | 16 +++++-----------
 1 file changed, 5 insertions(+), 11 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index c683de7f08b..9dc13c2ea0c 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -51,10 +51,6 @@ MATRIX_DIRECTIVES: Final = frozenset({"include", "exclude"})
 LOCAL_CALL_PREFIX: Final = "./"
 
 
-class CheckRunNameCollision(Exception):
-    pass
-
-
 class Job(BaseModel):
     name: str | None = None
     uses: str | None = None
@@ -264,18 +260,16 @@ def workflow_sources() -> Mapping[str, str]:
     return {path.relative_to(REPO_ROOT).as_posix(): path.read_text() for path in sorted(WORKFLOWS_DIR.glob("*.y*ml"))}
 
 
-def main() -> None:
+def main() -> int:
     sources: Final = workflow_sources()
     found: Final = collisions(sources)
     if found:
-        raise CheckRunNameCollision("Check-run names are not unique:\n  - " + "\n  - ".join(found))
+        print("ERROR: Check-run names are not unique:\n  - " + "\n  - ".join(found), file=sys.stderr)
+        return 1
 
     print(f"Check-run names are unique across {len(sources)} workflows")
+    return 0
 
 
 if __name__ == "__main__":
-    try:
-        main()
-    except CheckRunNameCollision as exc:
-        print(f"ERROR: {exc}", file=sys.stderr)
-        sys.exit(1)
+    sys.exit(main())

From ecf7e4e766cff606c9bbdef62d1d3a67c95113d0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 23:36:28 -0700
Subject: [PATCH 49/79] fix(router): keep the provider's error when the retry
 skip empties the group

Before excluding the deployment that just refused, the retry-skip guard asked
whether another one could still answer. It asked by re-running a single routing
filter, the order filter, while deployment selection also applies cooldowns, the
context-window pre-call check, tag routing, and routing plugins.

Any filter the guard did not replicate made it answer yes while the real pick was
left with nothing. A group narrowed to one deployment by tag routing turned the
provider's own 400 into a no-deployments 429.

The skip now runs where every filter has already been applied, and it keeps the
deployments untouched when skipping would leave none. The caller gets the
provider's error either way, and a group with one eligible deployment retries in
place as it did before.
---
 litellm/router.py                             |  59 ++++---
 .../complexity_router/complexity_router.py    |   5 +-
 tests/test_litellm/test_router.py             | 153 +++++++++++++-----
 3 files changed, 150 insertions(+), 67 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index 4901199902c..93178dc8e26 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -401,8 +401,7 @@ def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream])
 
 _NO_SESSION_KWARGS: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({})
 _SESSION_ADAPTER: Final = TypeAdapter(Mapping[str, object])
-_EXCLUDED_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
-_TARGET_ORDER_ADAPTER: Final = TypeAdapter(int | None)
+_SKIPPED_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
 
 
 def _with_router_resolved_session_model(session: object, model_name: str) -> Mapping[str, Mapping[str, object]]:
@@ -7461,29 +7460,19 @@ class Router:
             )
 
     @staticmethod
-    def _deployment_ids_to_skip_on_retry(
-        exception: Exception,
-        already_skipped: object,
-        healthy_deployments: list[dict],  # mutable-ok: matches the routing filters' list contract
-        target_order: object = None,
-    ) -> tuple[str, ...]:
+    def _deployment_ids_to_skip_on_retry(exception: Exception, already_skipped: object) -> tuple[str, ...]:
         failed_deployment_id: Final[str | None] = getattr(exception, "failed_deployment_id", None)
         status_code: Final = getattr(exception, "status_code", None)
         if not failed_deployment_id or not isinstance(status_code, int):
             return ()
         if litellm._should_retry(status_code):  # pyright: ignore[reportPrivateUsage]  # as in should_retry_this_error
             return ()
-        already_skipped_ids: Final = _EXCLUDED_DEPLOYMENT_IDS_ADAPTER.validate_python(already_skipped or ())
-        skipped: Final = frozenset((*already_skipped_ids, failed_deployment_id))
-        same_order_candidates: Final = litellm.utils.get_order_filtered_deployments(
-            healthy_deployments, target_order=_TARGET_ORDER_ADAPTER.validate_python(target_order)
-        )
-        if not litellm.utils.get_excluded_filtered_deployments(same_order_candidates, excluded_deployment_ids=skipped):
-            return ()
+        already_skipped_ids: Final = _SKIPPED_DEPLOYMENT_IDS_ADAPTER.validate_python(already_skipped or ())
+        skipped: Final = tuple(sorted(frozenset((*already_skipped_ids, failed_deployment_id))))
         verbose_router_logger.debug(
-            "Retry skips deployments that already answered %s to this request: %s", status_code, sorted(skipped)
+            "Retry skips deployments that already answered %s to this request: %s", status_code, skipped
         )
-        return tuple(sorted(skipped))
+        return skipped
 
     @tracer.wrap()
     async def async_function_with_retries(self, *args, **kwargs):
@@ -7582,12 +7571,10 @@ class Router:
                 kwargs = self.log_retry(kwargs=kwargs, e=original_exception)
                 skipped_deployment_ids: Final = self._deployment_ids_to_skip_on_retry(
                     exception=original_exception,
-                    already_skipped=kwargs.get("_excluded_deployment_ids"),
-                    healthy_deployments=_healthy_deployments,
-                    target_order=kwargs.get("_target_order"),
+                    already_skipped=kwargs.get("_retry_skipped_deployment_ids"),
                 )
                 if skipped_deployment_ids:
-                    kwargs["_excluded_deployment_ids"] = skipped_deployment_ids
+                    kwargs["_retry_skipped_deployment_ids"] = skipped_deployment_ids
             else:
                 raise
 
@@ -7659,12 +7646,10 @@ class Router:
 
                     retry_skipped_deployment_ids = self._deployment_ids_to_skip_on_retry(
                         exception=e,
-                        already_skipped=kwargs.get("_excluded_deployment_ids"),
-                        healthy_deployments=_healthy_deployments,
-                        target_order=kwargs.get("_target_order"),
+                        already_skipped=kwargs.get("_retry_skipped_deployment_ids"),
                     )
                     if retry_skipped_deployment_ids:
-                        kwargs["_excluded_deployment_ids"] = retry_skipped_deployment_ids
+                        kwargs["_retry_skipped_deployment_ids"] = retry_skipped_deployment_ids
                     _timeout = self._time_to_sleep_before_retry(
                         e=e,
                         remaining_retries=remaining_retries,
@@ -12508,6 +12493,19 @@ class Router:
             excluded_deployment_ids=_excluded_deployment_ids,
         )
 
+        ## RETRY SKIP ## -> drop deployments that already refused this request with a
+        ## non-retryable status, unless that leaves nothing, so the caller still gets
+        ## the provider's own error instead of a no-deployments error.
+        _retry_skipped_deployment_ids: Final = (
+            request_kwargs.pop("_retry_skipped_deployment_ids", None) if request_kwargs else None
+        )
+        healthy_deployments = (
+            litellm.utils.get_excluded_filtered_deployments(
+                healthy_deployments, excluded_deployment_ids=_retry_skipped_deployment_ids
+            )
+            or healthy_deployments
+        )
+
         if len(healthy_deployments) == 0:
             exception: Final = await async_raise_no_deployment_exception(
                 litellm_router_instance=self,
@@ -13413,6 +13411,17 @@ class Router:
             excluded_deployment_ids=_excluded_deployment_ids,
         )
 
+        ## RETRY SKIP ## -> see async counterpart in async_get_healthy_deployments.
+        _retry_skipped_deployment_ids: Final = (
+            request_kwargs.pop("_retry_skipped_deployment_ids", None) if request_kwargs else None
+        )
+        healthy_deployments = (
+            litellm.utils.get_excluded_filtered_deployments(
+                healthy_deployments, excluded_deployment_ids=_retry_skipped_deployment_ids
+            )
+            or healthy_deployments
+        )
+
         if len(healthy_deployments) == 0:
             model_ids = self.get_model_ids(model_name=model)
             _cooldown_time = self.cooldown_cache.get_min_cooldown(
diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py
index 98a1eb7ac9e..53a56866e92 100644
--- a/litellm/router_strategy/complexity_router/complexity_router.py
+++ b/litellm/router_strategy/complexity_router/complexity_router.py
@@ -2832,8 +2832,9 @@ class ComplexityRouter(CustomLogger):
         where the prompt never arrives as messages.
 
         Probed on a COPY of request_kwargs because the owner pops routing bookkeeping off the
-        dict it is handed (`_target_order`, `_excluded_deployment_ids`), and this is a
-        speculative question about a model that may never be picked.
+        dict it is handed (`_target_order`, `_excluded_deployment_ids`,
+        `_retry_skipped_deployment_ids`), and this is a speculative question about a model
+        that may never be picked.
 
         Every way the owner says "nothing here can serve this" is a negative verdict: no healthy
         deployment for the group at all (BadRequestError, which ContextWindowExceededError
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 0d7f2e7d3f6..34c435af706 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13234,61 +13234,134 @@ async def test_router_retry_policy_400_retries_on_sibling_deployment(
     assert response._hidden_params["additional_headers"]["x-litellm-attempted-retries"] == 1
 
 
+_UPSTREAM_400 = {"message": "upstream refused this request", "type": "invalid_request_error", "code": "bad_request"}
+
+
+def _retry_skip_deployment(deployment_id, host, litellm_params=None, model_info=None):
+    return {
+        "model_name": "gpt-5.6",
+        "litellm_params": {
+            "model": "openai/gpt-5.6",
+            "api_key": "sk-fake",
+            "api_base": f"https://{host}.local/v1",
+            **(litellm_params or {}),
+        },
+        "model_info": {"id": deployment_id, **(model_info or {})},
+    }
+
+
 @pytest.mark.parametrize(
-    "status_code,failed_deployment_id,already_skipped,healthy_deployment_ids,expected",
+    "status_code,failed_deployment_id,already_skipped,expected",
     [
-        (400, "rejecting", None, ["rejecting", "accepting"], ("rejecting",)),
-        (403, "rejecting", None, ["rejecting", "accepting"], ("rejecting",)),
-        (400, "second", ("first",), ["first", "second", "third"], ("first", "second")),
-        (429, "rejecting", None, ["rejecting", "accepting"], ()),
-        (503, "rejecting", None, ["rejecting", "accepting"], ()),
-        (400, "rejecting", None, ["rejecting"], ()),
-        (400, None, None, ["rejecting", "accepting"], ()),
-        (None, "rejecting", None, ["rejecting", "accepting"], ()),
-        ("400", "rejecting", None, ["rejecting", "accepting"], ()),
+        (400, "rejecting", None, ("rejecting",)),
+        (403, "rejecting", None, ("rejecting",)),
+        (400, "second", ("first",), ("first", "second")),
+        (400, "first", ("first",), ("first",)),
+        (429, "rejecting", None, ()),
+        (503, "rejecting", None, ()),
+        (408, "rejecting", None, ()),
+        (400, None, None, ()),
+        (None, "rejecting", None, ()),
+        ("400", "rejecting", None, ()),
     ],
 )
-def test_router_deployment_ids_to_skip_on_retry(
-    status_code, failed_deployment_id, already_skipped, healthy_deployment_ids, expected
-):
+def test_router_deployment_ids_to_skip_on_retry(status_code, failed_deployment_id, already_skipped, expected):
     exception = Exception("upstream refused this request")
     exception.status_code = status_code
     exception.failed_deployment_id = failed_deployment_id
-    healthy_deployments = [{"model_info": {"id": deployment_id}} for deployment_id in healthy_deployment_ids]
 
-    assert (
-        litellm.Router._deployment_ids_to_skip_on_retry(exception, already_skipped, healthy_deployments) == expected
-    )
+    assert litellm.Router._deployment_ids_to_skip_on_retry(exception, already_skipped) == expected
 
 
 @pytest.mark.parametrize(
-    "target_order,deployment_orders,expected",
+    "deployment_ids,skipped,expected",
     [
-        (2, {"rejecting": 2, "sibling": 1}, ()),
-        (2, {"rejecting": 2, "sibling": 2}, ("rejecting",)),
-        (1, {"rejecting": 1, "sibling": 2}, ()),
-        (None, {"rejecting": 1, "sibling": 2}, ()),
-        (None, {"rejecting": 1, "sibling": 1}, ("rejecting",)),
-        (3, {"rejecting": 2, "sibling": 1}, ()),
+        (["rejecting", "sibling"], ("rejecting",), ["sibling"]),
+        (["rejecting"], ("rejecting",), ["rejecting"]),
+        (["rejecting", "sibling"], ("rejecting", "sibling"), ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], (), ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], None, ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], ("absent",), ["rejecting", "sibling"]),
     ],
 )
-def test_router_deployment_ids_to_skip_on_retry_honors_order_fallback_target(
-    target_order, deployment_orders, expected
-):
-    exception = Exception("upstream refused this request")
-    exception.status_code = 400
-    exception.failed_deployment_id = "rejecting"
-    healthy_deployments = [
-        {"model_info": {"id": deployment_id}, "litellm_params": {"order": order}}
-        for deployment_id, order in deployment_orders.items()
-    ]
-
-    assert (
-        litellm.Router._deployment_ids_to_skip_on_retry(
-            exception, None, healthy_deployments, target_order=target_order
-        )
-        == expected
+@pytest.mark.asyncio
+async def test_router_healthy_deployments_keep_the_last_candidate_a_retry_skipped(deployment_ids, skipped, expected):
+    router = litellm.Router(
+        model_list=[_retry_skip_deployment(deployment_id, deployment_id) for deployment_id in deployment_ids],
+        disable_cooldowns=True,
     )
+    request_kwargs = {"_retry_skipped_deployment_ids": skipped}
+
+    healthy_deployments = await router.async_get_healthy_deployments(model="gpt-5.6", request_kwargs=request_kwargs)
+
+    assert sorted(deployment["model_info"]["id"] for deployment in healthy_deployments) == sorted(expected)
+    assert "_retry_skipped_deployment_ids" not in request_kwargs
+
+
+@pytest.mark.asyncio
+async def test_router_retry_policy_400_keeps_upstream_error_on_order_fallback_hop(monkeypatch: pytest.MonkeyPatch):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    router = litellm.Router(
+        model_list=[
+            _retry_skip_deployment("order1", "order1", litellm_params={"order": 1}),
+            _retry_skip_deployment("order2", "order2", litellm_params={"order": 2}),
+        ],
+        num_retries=2,
+        retry_policy={"BadRequestErrorRetries": 2},
+        disable_cooldowns=True,
+    )
+
+    with respx.mock as respx_mock:
+        order1 = respx_mock.post("https://order1.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        order2 = respx_mock.post("https://order2.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        with pytest.raises(litellm.BadRequestError) as raised:
+            await router.acompletion(model="gpt-5.6", messages=[{"role": "user", "content": "hi"}])
+
+    assert "upstream refused this request" in str(raised.value)
+    assert "No deployments available" not in str(raised.value)
+    assert order1.call_count >= 1
+    assert order2.call_count >= 1
+
+
+@pytest.mark.asyncio
+async def test_router_retry_policy_400_keeps_upstream_error_when_tags_narrow_the_group(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    router = litellm.Router(
+        model_list=[
+            _retry_skip_deployment(
+                "tagged", "tagged", litellm_params={"tags": ["free"]}, model_info={"enable_tag_filtering": True}
+            ),
+            _retry_skip_deployment("untagged", "untagged", model_info={"enable_tag_filtering": True}),
+        ],
+        num_retries=2,
+        retry_policy={"BadRequestErrorRetries": 2},
+        disable_cooldowns=True,
+    )
+
+    with respx.mock as respx_mock:
+        tagged = respx_mock.post("https://tagged.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        untagged = respx_mock.post("https://untagged.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        with pytest.raises(litellm.BadRequestError) as raised:
+            await router.acompletion(
+                model="gpt-5.6",
+                messages=[{"role": "user", "content": "hi"}],
+                metadata={"tags": ["free"]},
+            )
+
+    assert "upstream refused this request" in str(raised.value)
+    assert "No deployments available" not in str(raised.value)
+    assert tagged.call_count == 3
+    assert untagged.call_count == 0
 
 
 def _make_failure_logging_obj():

From 6866eac96feed14e47e051896c45263fee53086a Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 5 Sep 2026 23:53:36 -0700
Subject: [PATCH 50/79] fix(router): ignore a retry skip list the caller sent
 itself

The retry skip travels as a request kwarg, and the router forwards keys it
does not recognize, so a client can put _retry_skipped_deployment_ids in its
own request body. The value went straight into a pydantic TypeAdapter and
then into a set(), so an int or an object raised TypeError and a string, a
list, or a dict raised a ValidationError, each of them replacing the 400 the
provider had actually returned.

Every read now goes through one narrowing function that keeps a tuple of
strings and skips nothing otherwise, so a forged value costs the caller
nothing beyond the retry landing on the same deployment again.
---
 litellm/router.py                 | 11 +++++----
 tests/test_litellm/test_router.py | 39 +++++++++++++++++++++++++++++++
 2 files changed, 46 insertions(+), 4 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index 93178dc8e26..e42bc2af398 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -401,7 +401,10 @@ def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream])
 
 _NO_SESSION_KWARGS: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({})
 _SESSION_ADAPTER: Final = TypeAdapter(Mapping[str, object])
-_SKIPPED_DEPLOYMENT_IDS_ADAPTER: Final = TypeAdapter(tuple[str, ...])
+
+
+def _as_retry_skipped_deployment_ids(value: object) -> tuple[str, ...]:
+    return tuple(item for item in value if isinstance(item, str)) if isinstance(value, tuple) else ()
 
 
 def _with_router_resolved_session_model(session: object, model_name: str) -> Mapping[str, Mapping[str, object]]:
@@ -7467,7 +7470,7 @@ class Router:
             return ()
         if litellm._should_retry(status_code):  # pyright: ignore[reportPrivateUsage]  # as in should_retry_this_error
             return ()
-        already_skipped_ids: Final = _SKIPPED_DEPLOYMENT_IDS_ADAPTER.validate_python(already_skipped or ())
+        already_skipped_ids: Final = _as_retry_skipped_deployment_ids(already_skipped)
         skipped: Final = tuple(sorted(frozenset((*already_skipped_ids, failed_deployment_id))))
         verbose_router_logger.debug(
             "Retry skips deployments that already answered %s to this request: %s", status_code, skipped
@@ -12496,7 +12499,7 @@ class Router:
         ## RETRY SKIP ## -> drop deployments that already refused this request with a
         ## non-retryable status, unless that leaves nothing, so the caller still gets
         ## the provider's own error instead of a no-deployments error.
-        _retry_skipped_deployment_ids: Final = (
+        _retry_skipped_deployment_ids: Final = _as_retry_skipped_deployment_ids(
             request_kwargs.pop("_retry_skipped_deployment_ids", None) if request_kwargs else None
         )
         healthy_deployments = (
@@ -13412,7 +13415,7 @@ class Router:
         )
 
         ## RETRY SKIP ## -> see async counterpart in async_get_healthy_deployments.
-        _retry_skipped_deployment_ids: Final = (
+        _retry_skipped_deployment_ids: Final = _as_retry_skipped_deployment_ids(
             request_kwargs.pop("_retry_skipped_deployment_ids", None) if request_kwargs else None
         )
         healthy_deployments = (
diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 34c435af706..696abcd65a5 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13263,6 +13263,10 @@ def _retry_skip_deployment(deployment_id, host, litellm_params=None, model_info=
         (400, None, None, ()),
         (None, "rejecting", None, ()),
         ("400", "rejecting", None, ()),
+        (400, "second", 7, ("second",)),
+        (400, "second", "first", ("second",)),
+        (400, "second", ["first"], ("second",)),
+        (400, "second", ("first", 7), ("first", "second")),
     ],
 )
 def test_router_deployment_ids_to_skip_on_retry(status_code, failed_deployment_id, already_skipped, expected):
@@ -13282,6 +13286,11 @@ def test_router_deployment_ids_to_skip_on_retry(status_code, failed_deployment_i
         (["rejecting", "sibling"], (), ["rejecting", "sibling"]),
         (["rejecting", "sibling"], None, ["rejecting", "sibling"]),
         (["rejecting", "sibling"], ("absent",), ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], 7, ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], "rejecting", ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], ["rejecting"], ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], {"rejecting": True}, ["rejecting", "sibling"]),
+        (["rejecting", "sibling"], ("rejecting", 7), ["sibling"]),
     ],
 )
 @pytest.mark.asyncio
@@ -13298,6 +13307,36 @@ async def test_router_healthy_deployments_keep_the_last_candidate_a_retry_skippe
     assert "_retry_skipped_deployment_ids" not in request_kwargs
 
 
+@pytest.mark.parametrize("client_supplied", [7, "rejecting", ["rejecting"], {"rejecting": True}, object()])
+@pytest.mark.asyncio
+async def test_router_retry_policy_400_keeps_upstream_error_when_a_client_forges_the_skip_list(
+    monkeypatch: pytest.MonkeyPatch, client_supplied
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    router = litellm.Router(
+        model_list=[_retry_skip_deployment("rejecting", "rejecting"), _retry_skip_deployment("sibling", "sibling")],
+        num_retries=2,
+        retry_policy={"BadRequestErrorRetries": 2},
+        disable_cooldowns=True,
+    )
+
+    with respx.mock as respx_mock:
+        respx_mock.post("https://rejecting.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        respx_mock.post("https://sibling.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        with pytest.raises(litellm.BadRequestError) as raised:
+            await router.acompletion(
+                model="gpt-5.6",
+                messages=[{"role": "user", "content": "hi"}],
+                _retry_skipped_deployment_ids=client_supplied,
+            )
+
+    assert "upstream refused this request" in str(raised.value)
+
+
 @pytest.mark.asyncio
 async def test_router_retry_policy_400_keeps_upstream_error_on_order_fallback_hop(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)

From 5bbc83e3defa48db50546aa81ba60a9cdb45ed58 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 00:04:31 -0700
Subject: [PATCH 51/79] fix: model the check-run names GitHub really publishes

The collision sweep read a job's name as its `name:` or bare job id, which is
wrong for a matrix job that sets no name: GitHub publishes `build (3.12)`, one
per combination. That missed real duplicates and invented ones that don't exist.
It also crossed every matrix value while ignoring `exclude`, so it checked
combinations no job ever runs.

Four smaller gaps went with it. Boolean matrix values reached a name as `True`
rather than `true`. A `format()` whose arguments cannot fill its placeholders
raised straight out of the script instead of leaving the name unresolved. A job
calling a reusable workflow only ever chained one level, and a call outside the
repo fell back to the caller's own name, which GitHub never posts. A job whose
`name:` was not a string failed validation and silently dropped every job in
that file, so the sweep now renders any scalar and reports a file it cannot read
instead of skipping it.
---
 .../check_workflow_job_name_collisions.py     | 174 +++++++++++-----
 .../test_workflow_job_name_collisions.py      | 196 +++++++++++++++++-
 2 files changed, 319 insertions(+), 51 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 9dc13c2ea0c..4eef3df1c11 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -11,19 +11,27 @@ clash completely, so the context quietly stops meaning what the ruleset intended
 and commit ed5761daef4ae17152446d182c860630c38b7268 carried both check runs.
 This invariant has to be enforced here because CI cannot enforce it on itself.
 
-A job publishes its `name:` when it sets one and its job id otherwise. Two shapes
-expand that further. A name carrying `${{ ... }}` publishes one check run per
-combination the matrix produces, with each `include` row's values held together
-rather than crossed with the other rows', so two shard lists that overlap collide
-even though their templates read differently. Each expression is evaluated per combination
-over the pieces a job name can hold: string literals, `matrix.<key>`, `format()`,
-`==` and `!=`, and the `<cond> && <a> || <b>` idiom, which is how the shards
-reach their real `<shard> / Run tests` names rather than staying opaque. A job
-calling a local reusable workflow publishes one check run per job of the callee,
-named `<caller> / <callee>`, which is why a caller's name never collides with a
-plain job that happens to match it. An expression nothing resolves stays in the
-string, so two jobs carrying the same unresolved template still compare equal and
-their collision is still caught.
+A job publishes its `name:` when it sets one, and otherwise its job id plus the
+values of the combination it runs, the way GitHub writes `build (3.12)`. A name
+carrying `${{ ... }}` publishes one check run per combination the matrix
+produces: `exclude` rows drop combinations before `include` rows fold into the
+survivors, and each `include` row's values stay together rather than crossing
+with the other rows', so two shard lists that overlap collide even though their
+templates read differently. Each expression is evaluated per combination over the
+pieces a job name can hold: string literals, `matrix.<key>`, `format()`, `==` and
+`!=`, and the `<cond> && <a> || <b>` idiom, which is how the shards reach their
+real `<shard> / Run tests` names rather than staying opaque. An expression
+nothing resolves stays in the string, so two jobs carrying the same unresolved
+template still compare equal and their collision is still caught.
+
+A job calling a local reusable workflow publishes one check run per job of the
+callee, named `<caller> / <callee>` and chained through however many levels of
+local calls it takes, which is why a caller's name never collides with a plain
+job that happens to match it. A call this sweep cannot follow, to a workflow
+outside the repo or to one that is not there, publishes nothing rather than a
+name GitHub never posts. A file under `.github/workflows/` that does not read as
+a workflow at all is reported rather than skipped, since skipping it silently
+would hide every job it holds.
 """
 
 import itertools
@@ -31,6 +39,7 @@ import operator
 import re
 import sys
 from collections.abc import Iterator, Mapping, Sequence
+from dataclasses import dataclass
 from pathlib import Path
 from types import MappingProxyType
 from typing import Final
@@ -46,13 +55,19 @@ LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
 FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
 COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
 NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
+NO_CALLERS: Final[frozenset[str]] = frozenset()
 SCALAR: Final = (str, int, float)
 MATRIX_DIRECTIVES: Final = frozenset({"include", "exclude"})
 LOCAL_CALL_PREFIX: Final = "./"
 
 
+@dataclass(frozen=True, slots=True)
+class Unreadable:
+    reason: str
+
+
 class Job(BaseModel):
-    name: str | None = None
+    name: object = None
     uses: str | None = None
     strategy: Mapping[str, object] = Field(default_factory=dict)
 
@@ -61,15 +76,20 @@ class Workflow(BaseModel):
     jobs: Mapping[str, Job] = Field(default_factory=dict)
 
 
-def parse(text: str) -> tuple[Workflow, object] | None:
-    """The workflow plus its raw `on:` value, or None when the file is not a workflow."""
-    parsed: Final = yaml.safe_load(text)
+def scalar_text(value: object) -> str:
+    """A YAML scalar the way GitHub renders it, so `true` never reaches a name as `True`."""
+    return str(value).lower() if isinstance(value, bool) else str(value)
+
+
+def parse(source: str) -> tuple[Workflow, object] | Unreadable:
+    """The workflow plus its raw `on:` value, or why the file does not read as one."""
+    parsed: Final = yaml.safe_load(source)
     if not isinstance(parsed, dict):
-        return None
+        return Unreadable("its top level is not a mapping of workflow keys")
     try:
         return Workflow.model_validate(parsed), parsed.get(True, parsed.get("on"))
-    except ValidationError:
-        return None
+    except ValidationError as error:
+        return Unreadable(f"{error.error_count()} of its job definitions have a shape GitHub would reject")
 
 
 def events(raw_on: object) -> frozenset[str]:
@@ -89,23 +109,28 @@ def publishes_check_runs(raw_on: object) -> bool:
 
 def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, ...]], ...]:
     return tuple(
-        (str(key), tuple(str(value) for value in values if isinstance(value, SCALAR)))
+        (str(key), tuple(scalar_text(value) for value in values if isinstance(value, SCALAR)))
         for key, values in matrix.items()
         if str(key) not in MATRIX_DIRECTIVES and isinstance(values, Sequence) and not isinstance(values, str)
     )
 
 
-def include_rows(matrix: Mapping[str, object]) -> tuple[Mapping[str, str], ...]:
-    rows: Final = matrix.get("include")
+def directive_rows(matrix: Mapping[str, object], directive: str) -> tuple[Mapping[str, str], ...]:
+    rows: Final = matrix.get(directive)
     if not isinstance(rows, Sequence) or isinstance(rows, str):
         return ()
     return tuple(
-        MappingProxyType({str(key): str(value) for key, value in row.items() if isinstance(value, SCALAR)})
+        MappingProxyType({str(key): scalar_text(value) for key, value in row.items() if isinstance(value, SCALAR)})
         for row in rows
         if isinstance(row, Mapping)
     )
 
 
+def drops(row: Mapping[str, str], combination: Mapping[str, str]) -> bool:
+    """GitHub removes a combination that carries every value one `exclude` row names."""
+    return all(combination.get(key) == value for key, value in row.items())
+
+
 def extends(row: Mapping[str, str], combination: Mapping[str, str]) -> bool:
     """GitHub folds an `include` row into a combination only where it overwrites no listed value."""
     return all(combination[key] == value for key, value in row.items() if key in combination)
@@ -116,23 +141,29 @@ def extended(combination: Mapping[str, str], rows: Sequence[Mapping[str, str]])
     return MappingProxyType({**combination, **additions})
 
 
+def crossed_values(listed: Sequence[tuple[str, tuple[str, ...]]]) -> tuple[Mapping[str, str], ...]:
+    if not listed:
+        return ()
+    return tuple(
+        MappingProxyType(dict(zip((key for key, _ in listed), values)))
+        for values in itertools.product(*(values for _, values in listed))
+    )
+
+
 def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...]:
-    """One mapping per job the matrix produces, each `include` row's values staying together."""
+    """One mapping per job the matrix produces, `exclude` applied before `include` as GitHub does."""
     matrix: Final = job.strategy.get("matrix")
     if not isinstance(matrix, Mapping):
         return ()
-    listed: Final = listed_values(matrix)
-    rows: Final = include_rows(matrix)
-    crossed: Final = (
-        tuple(
-            MappingProxyType(dict(zip((key for key, _ in listed), values)))
-            for values in itertools.product(*(values for _, values in listed))
-        )
-        if listed
-        else ()
+    rows: Final = directive_rows(matrix, "include")
+    dropped: Final = directive_rows(matrix, "exclude")
+    kept: Final = tuple(
+        combination
+        for combination in crossed_values(listed_values(matrix))
+        if not any(drops(row, combination) for row in dropped)
     )
-    standalone: Final = tuple(row for row in rows if not any(extends(row, combination) for combination in crossed))
-    return (*(extended(combination, rows) for combination in crossed), *standalone)
+    standalone: Final = tuple(row for row in rows if not any(extends(row, combination) for combination in kept))
+    return (*(extended(combination, rows) for combination in kept), *standalone)
 
 
 def scanned(state: tuple[int, bool], char: str) -> tuple[int, bool]:
@@ -156,6 +187,14 @@ def split_outside(text: str, token: str) -> tuple[str, ...]:
     return tuple(text[start:end] for start, end in zip(starts, (*cuts, len(text))))
 
 
+def formatted(template: str, arguments: Sequence[str]) -> str | None:
+    """A `format()` whose placeholders the arguments cannot fill resolves to nothing, never a crash."""
+    try:
+        return template.format(*arguments)
+    except (IndexError, KeyError, ValueError):
+        return None
+
+
 def value_of(text: str, values: Mapping[str, str]) -> str | None:
     expression: Final = text.strip()
     literal: Final = LITERAL.match(expression)
@@ -171,7 +210,7 @@ def value_of(text: str, values: Mapping[str, str]) -> str | None:
     resolved: Final = tuple(argument for argument in arguments if argument is not None)
     if not resolved or len(resolved) != len(arguments):
         return None
-    return resolved[0].format(*resolved[1:])
+    return formatted(resolved[0], resolved[1:])
 
 
 def holds(condition: str, values: Mapping[str, str]) -> bool | None:
@@ -211,27 +250,58 @@ def expand(template: str, job: Job) -> tuple[str, ...]:
     return tuple(dict.fromkeys(rendered(template, values) for values in combinations))
 
 
+def suffixed(job_id: str, combination: Mapping[str, str]) -> str:
+    """The name GitHub gives a job with no `name:`, its id plus the combination it runs."""
+    return f"{job_id} ({', '.join(combination.values())})" if combination else job_id
+
+
+def published_names(job_id: str, job: Job) -> tuple[str, ...]:
+    if job.name is not None:
+        return expand(scalar_text(job.name), job)
+    suffixes: Final = tuple(dict.fromkeys(suffixed(job_id, values) for values in matrix_combinations(job)))
+    return suffixes or (job_id,)
+
+
 def callee_path(job: Job) -> str | None:
     if job.uses is None or not job.uses.startswith(LOCAL_CALL_PREFIX):
         return None
     return job.uses[len(LOCAL_CALL_PREFIX) :].split("@")[0]
 
 
-def job_names(job_id: str, job: Job, workflows: Mapping[str, Workflow]) -> tuple[str, ...]:
-    prefixes: Final = expand(job.name or job_id, job)
-    callee: Final = workflows.get(callee_path(job) or "")
-    if callee is None:
+def job_names(
+    job_id: str, job: Job, workflows: Mapping[str, Workflow], callers: frozenset[str] = NO_CALLERS
+) -> tuple[str, ...]:
+    prefixes: Final = published_names(job_id, job)
+    if job.uses is None:
         return prefixes
+    path: Final = callee_path(job)
+    callee: Final = workflows.get(path or "")
+    if path is None or callee is None or path in callers:
+        return ()
     suffixes: Final = tuple(
         name
         for callee_id, callee_job in callee.jobs.items()
-        for name in expand(callee_job.name or callee_id, callee_job)
+        for name in job_names(callee_id, callee_job, workflows, callers | {path})
     )
     return tuple(f"{prefix} / {suffix}" for prefix in prefixes for suffix in suffixes)
 
 
+def readable(sources: Mapping[str, str]) -> Mapping[str, tuple[Workflow, object]]:
+    parsed: Final = {rel: parse(source) for rel, source in sources.items()}
+    return MappingProxyType({rel: entry for rel, entry in parsed.items() if not isinstance(entry, Unreadable)})
+
+
+def unreadable(sources: Mapping[str, str]) -> tuple[str, ...]:
+    parsed: Final = {rel: parse(source) for rel, source in sources.items()}
+    return tuple(
+        f"{rel} sits in the workflows directory but {entry.reason}, so none of its jobs were checked."
+        for rel, entry in sorted(parsed.items())
+        if isinstance(entry, Unreadable)
+    )
+
+
 def published(sources: Mapping[str, str]) -> Iterator[tuple[str, str]]:
-    parsed: Final = {rel: entry for rel, text in sources.items() if (entry := parse(text)) is not None}
+    parsed: Final = readable(sources)
     workflows: Final = {rel: workflow for rel, (workflow, _) in parsed.items()}
     for rel, (workflow, raw_on) in parsed.items():
         if not publishes_check_runs(raw_on):
@@ -260,16 +330,26 @@ def workflow_sources() -> Mapping[str, str]:
     return {path.relative_to(REPO_ROOT).as_posix(): path.read_text() for path in sorted(WORKFLOWS_DIR.glob("*.y*ml"))}
 
 
-def main() -> int:
-    sources: Final = workflow_sources()
+def report(header: str, problems: Sequence[str]) -> None:
+    if problems:
+        print(f"ERROR: {header}:\n  - " + "\n  - ".join(problems), file=sys.stderr)
+
+
+def exit_code(sources: Mapping[str, str]) -> int:
+    unread: Final = unreadable(sources)
     found: Final = collisions(sources)
-    if found:
-        print("ERROR: Check-run names are not unique:\n  - " + "\n  - ".join(found), file=sys.stderr)
+    report("Some workflows could not be read", unread)
+    report("Check-run names are not unique", found)
+    if unread or found:
         return 1
 
     print(f"Check-run names are unique across {len(sources)} workflows")
     return 0
 
 
+def main() -> int:
+    return exit_code(workflow_sources())
+
+
 if __name__ == "__main__":
     sys.exit(main())
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 359072326bd..3dd8a3d4088 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -1,6 +1,15 @@
 from typing import Final
 
-from check_workflow_job_name_collisions import callee_path, collisions, parse, published, workflow_sources
+from check_workflow_job_name_collisions import (
+    Unreadable,
+    callee_path,
+    collisions,
+    exit_code,
+    parse,
+    published,
+    unreadable,
+    workflow_sources,
+)
 
 REUSABLE_BASE: Final = """on:
   workflow_call:
@@ -53,6 +62,62 @@ jobs:
           - label: fast
 """
 
+NAMELESS_MATRIX: Final = """on: pull_request
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.12", "3.13"]
+"""
+
+EXCLUDED_PAIR: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ matrix.os }}-${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        os: [ubuntu, macos]
+        python-version: ["3.12", "3.13"]
+        exclude:
+          - os: macos
+            python-version: "3.13"
+"""
+
+EXCLUDED_KEY: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ matrix.os }}-${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        os: [ubuntu, macos]
+        python-version: ["3.12", "3.13"]
+        exclude:
+          - os: macos
+"""
+
+BOOLEAN_MATRIX: Final = """on: pull_request
+jobs:
+  unit:
+    name: cache ${{ matrix.cached }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        cached: [true, false]
+"""
+
+UNFILLABLE_FORMAT: Final = """on: pull_request
+jobs:
+  unit:
+    name: ${{ format('{0} {1}', matrix.shard) }}
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        shard: [core]
+"""
+
 
 def test_every_workflow_in_the_repo_publishes_a_unique_check_run_name() -> None:
     assert collisions(workflow_sources()) == ()
@@ -60,7 +125,9 @@ def test_every_workflow_in_the_repo_publishes_a_unique_check_run_name() -> None:
 
 def test_every_workflow_in_the_repo_parses_into_jobs() -> None:
     unparsed: Final = tuple(
-        rel for rel, source in workflow_sources().items() if (entry := parse(source)) is None or not entry[0].jobs
+        rel
+        for rel, source in workflow_sources().items()
+        if isinstance(entry := parse(source), Unreadable) or not entry[0].jobs
     )
 
     assert unparsed == ()
@@ -68,7 +135,7 @@ def test_every_workflow_in_the_repo_parses_into_jobs() -> None:
 
 def test_every_local_reusable_call_in_the_repo_resolves_to_a_workflow() -> None:
     sources: Final = workflow_sources()
-    parsed: Final = tuple(entry for text in sources.values() if (entry := parse(text)) is not None)
+    parsed: Final = tuple(entry for text in sources.values() if not isinstance(entry := parse(text), Unreadable))
     unresolved: Final = tuple(
         job.uses
         for workflow, _ in parsed
@@ -249,15 +316,28 @@ def test_two_jobs_sharing_one_unresolvable_template_still_collide() -> None:
     assert "is published by 2 jobs" in found[0]
 
 
-def test_a_file_that_is_not_a_workflow_is_ignored() -> None:
+def test_a_file_that_is_not_a_workflow_is_reported_rather_than_skipped() -> None:
     sources: Final = {
         "notes.yml": "just a string\n",
         "a.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
     }
 
+    found: Final = unreadable(sources)
+
+    assert len(found) == 1
+    assert "notes.yml" in found[0]
     assert collisions(sources) == ()
 
 
+def test_a_workflow_holding_a_job_shape_github_would_reject_is_reported() -> None:
+    sources: Final = {"a.yml": "on: pull_request\njobs:\n  test:\n    uses: [not, a, string]\n"}
+
+    found: Final = unreadable(sources)
+
+    assert len(found) == 1
+    assert "a.yml" in found[0]
+
+
 def test_a_conditional_name_expands_to_the_branch_each_matrix_value_takes() -> None:
     names: Final = frozenset(
         name for name, _ in published({".github/workflows/base.yml": REUSABLE_BASE, "unit.yml": SHARD_CALLER})
@@ -310,3 +390,111 @@ def test_every_workflow_in_the_repo_resolves_every_expression_in_its_job_names()
     unresolved: Final = tuple(f"{owner}: {name}" for name, owner in published(workflow_sources()) if "${{" in name)
 
     assert unresolved == ()
+
+
+def test_a_matrix_job_with_no_name_publishes_the_id_and_values_github_appends() -> None:
+    names: Final = frozenset(name for name, _ in published({"a.yml": NAMELESS_MATRIX}))
+
+    assert names == frozenset({"build (3.12)", "build (3.13)"})
+
+
+def test_a_matrix_job_with_no_name_does_not_collide_with_a_plain_job_carrying_its_id() -> None:
+    sources: Final = {
+        "a.yml": NAMELESS_MATRIX,
+        "b.yml": "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+
+
+def test_a_matrix_job_with_no_name_collides_with_the_suffixed_name_github_writes() -> None:
+    sources: Final = {
+        "a.yml": NAMELESS_MATRIX,
+        "b.yml": "on: pull_request\njobs:\n  legacy:\n    name: build (3.13)\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`build (3.13)` is published by 2 jobs" in found[0]
+
+
+def test_an_excluded_combination_publishes_no_check_run() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": EXCLUDED_PAIR}))
+
+    assert names == frozenset({"ubuntu-3.12", "ubuntu-3.13", "macos-3.12"})
+
+
+def test_an_exclude_row_naming_one_key_drops_every_combination_carrying_it() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": EXCLUDED_KEY}))
+
+    assert names == frozenset({"ubuntu-3.12", "ubuntu-3.13"})
+
+
+def test_a_boolean_matrix_value_renders_the_way_github_writes_it() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": BOOLEAN_MATRIX}))
+
+    assert names == frozenset({"cache true", "cache false"})
+
+
+def test_a_format_call_its_arguments_cannot_fill_leaves_the_name_unresolved() -> None:
+    names: Final = frozenset(name for name, _ in published({"unit.yml": UNFILLABLE_FORMAT}))
+
+    assert names == frozenset({"${{ format('{0} {1}', matrix.shard) }}"})
+
+
+def test_a_call_to_a_workflow_outside_the_repo_publishes_no_name() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  unit:\n    uses: BerriAI/other/.github/workflows/base.yml@main\n",
+        "b.yml": "on: pull_request\njobs:\n  unit:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"unit"})
+    assert collisions(sources) == ()
+
+
+def test_a_chain_of_local_reusable_calls_publishes_every_level_of_the_chain() -> None:
+    sources: Final = {
+        ".github/workflows/leaf.yml": (
+            "on:\n  workflow_call:\njobs:\n  run:\n    name: Leaf\n    runs-on: ubuntu-latest\n"
+        ),
+        ".github/workflows/mid.yml": (
+            "on:\n  workflow_call:\njobs:\n  call:\n    name: Mid\n    uses: ./.github/workflows/leaf.yml\n"
+        ),
+        "top.yml": "on: pull_request\njobs:\n  top:\n    name: Top\n    uses: ./.github/workflows/mid.yml\n",
+    }
+
+    names: Final = frozenset(name for name, _ in published(sources))
+
+    assert names == frozenset({"Top / Mid / Leaf"})
+
+
+def test_a_job_name_that_is_not_a_string_still_publishes_the_value_github_renders() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  sweep:\n    name: 2024\n    runs-on: ubuntu-latest\n",
+        "b.yml": 'on: pull_request\njobs:\n  other:\n    name: "2024"\n    runs-on: ubuntu-latest\n',
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`2024` is published by 2 jobs" in found[0]
+
+
+def test_the_check_fails_when_a_file_in_the_workflows_directory_cannot_be_read() -> None:
+    assert exit_code({"notes.yml": "just a string\n"}) == 1
+
+
+def test_the_check_fails_when_two_jobs_publish_one_check_run_name() -> None:
+    plain: Final = "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n"
+
+    assert exit_code({"a.yml": plain, "b.yml": plain}) == 1
+
+
+def test_the_check_passes_when_every_file_reads_and_every_name_is_unique() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  test:\n    runs-on: ubuntu-latest\n",
+        "b.yml": "on: pull_request\njobs:\n  sweep:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert exit_code(sources) == 0

From 0fcf0fe06c80683da3d4a7b7b63a0c0aa922382b Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 00:06:35 -0700
Subject: [PATCH 52/79] test(router): cover the retry skip-list narrowing
 helper

The router code coverage gate reads every function defined in router.py
and fails when no test file names it. _as_retry_skipped_deployment_ids
was only reached indirectly through the retry path, so the gate went red
on this PR's tip.

Test it directly instead: a tuple of strings survives, non-string items
inside the tuple are dropped, and every other shape a caller could send
narrows to an empty skip list.
---
 tests/test_litellm/test_router.py | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 696abcd65a5..05ac672691e 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -13277,6 +13277,26 @@ def test_router_deployment_ids_to_skip_on_retry(status_code, failed_deployment_i
     assert litellm.Router._deployment_ids_to_skip_on_retry(exception, already_skipped) == expected
 
 
+@pytest.mark.parametrize(
+    "value,expected",
+    [
+        (("first", "second"), ("first", "second")),
+        ((), ()),
+        (("first", 7, None, "second"), ("first", "second")),
+        (None, ()),
+        (7, ()),
+        ("first", ()),
+        (["first"], ()),
+        ({"first": True}, ()),
+        (object(), ()),
+    ],
+)
+def test_router_as_retry_skipped_deployment_ids_keeps_only_a_tuple_of_strings(value, expected):
+    from litellm.router import _as_retry_skipped_deployment_ids
+
+    assert _as_retry_skipped_deployment_ids(value) == expected
+
+
 @pytest.mark.parametrize(
     "deployment_ids,skipped,expected",
     [

From a08fb489cdb2375b25b19fdca0a0f17a90942574 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Sun, 6 Sep 2026 00:38:07 -0700
Subject: [PATCH 53/79] test(ui): accept trailing slashes in migration sidebar
 links

---
 tests/e2e/ui/tests/migration/migratedPages.spec.ts | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tests/e2e/ui/tests/migration/migratedPages.spec.ts b/tests/e2e/ui/tests/migration/migratedPages.spec.ts
index 547330190bd..bc96d9a6c66 100644
--- a/tests/e2e/ui/tests/migration/migratedPages.spec.ts
+++ b/tests/e2e/ui/tests/migration/migratedPages.spec.ts
@@ -29,12 +29,12 @@ async function expectRendered(page: Page) {
 }
 
 /**
- * Click a migrated page's sidebar link. Migrated items render as <a href=".../ui/<segment>">;
+ * Click a migrated page's sidebar link. Migrated items render as <a href=".../ui/<segment>/">;
  * nested ones live under collapsible groups whose children only render while the
  * group is open, so expand collapsed groups until the link is clickable.
  */
 async function clickSidebar(page: Page, segment: string) {
-  const link = sidebar(page).locator(`a[href$="/ui/${segment}"]`).first();
+  const link = sidebar(page).locator(`a[href$="/ui/${segment}"], a[href$="/ui/${segment}/"]`).first();
   const collapsedGroups = sidebar(page).getByRole("button", { expanded: false });
   for (let i = 0; i < 8 && !(await link.isVisible().catch(() => false)); i++) {
     const stillCollapsed = await collapsedGroups.count();

From d664ca139ef6cf2a2d79a8e98bdfb4985bc3ed32 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 00:53:24 -0700
Subject: [PATCH 54/79] chore(router): suppress the retry-skip kwargs writes
 and correct the filter docstring

The two writes that hand the skip list to the next attempt now carry a
`# rebind-ok` reason, which is the sanctioned escape hatch for an unavoidable
parameter mutation and matches how `log_retry` already writes into the same
kwargs dict a few lines above

`get_excluded_filtered_deployments`'s docstring said returning the unfiltered
list would re-include the deployment that just failed. The retry skip does
exactly that on purpose, so the docstring now says each caller decides what an
empty result means

The reliability registry cell the new e2e test claims is marked
`fail_before_fix: proven`: the same config returns 400 at the merge base and
200 off a sibling deployment at the tip
---
 litellm/router.py                            | 12 ++++++------
 litellm/utils.py                             | 10 ++++++----
 tests/e2e/coverage_registry/reliability.yaml |  2 +-
 3 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/litellm/router.py b/litellm/router.py
index e42bc2af398..85586ccab17 100644
--- a/litellm/router.py
+++ b/litellm/router.py
@@ -7572,12 +7572,12 @@ class Router:
             ## LOGGING
             if num_retries > 0:
                 kwargs = self.log_retry(kwargs=kwargs, e=original_exception)
-                skipped_deployment_ids: Final = self._deployment_ids_to_skip_on_retry(
+                first_skipped_ids: Final = self._deployment_ids_to_skip_on_retry(
                     exception=original_exception,
                     already_skipped=kwargs.get("_retry_skipped_deployment_ids"),
                 )
-                if skipped_deployment_ids:
-                    kwargs["_retry_skipped_deployment_ids"] = skipped_deployment_ids
+                if first_skipped_ids:
+                    kwargs["_retry_skipped_deployment_ids"] = first_skipped_ids  # rebind-ok: the next attempt reads it
             else:
                 raise
 
@@ -7647,12 +7647,12 @@ class Router:
                         except Exception:
                             raise e
 
-                    retry_skipped_deployment_ids = self._deployment_ids_to_skip_on_retry(
+                    skipped_ids = self._deployment_ids_to_skip_on_retry(
                         exception=e,
                         already_skipped=kwargs.get("_retry_skipped_deployment_ids"),
                     )
-                    if retry_skipped_deployment_ids:
-                        kwargs["_retry_skipped_deployment_ids"] = retry_skipped_deployment_ids
+                    if skipped_ids:
+                        kwargs["_retry_skipped_deployment_ids"] = skipped_ids  # rebind-ok: the next attempt reads it
                     _timeout = self._time_to_sleep_before_retry(
                         e=e,
                         remaining_retries=remaining_retries,
diff --git a/litellm/utils.py b/litellm/utils.py
index 238225eff99..f99d7e6a4b7 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -4919,10 +4919,12 @@ def get_excluded_filtered_deployments(
     across the remaining deployments in the same model group after one of them
     has failed.
 
-    If the filter would leave no deployments, an empty list is returned so the
-    caller raises its usual no-deployments error and the weighted-failover
-    helper falls through to the cross-group fallback path. Returning the
-    original unfiltered list here would re-include the just-failed deployment.
+    If the filter would leave no deployments, an empty list is returned and the
+    caller decides what that means. Weighted failover lets it raise the usual
+    no-deployments error and fall through to the cross-group fallback path; the
+    retry skip in `async_get_healthy_deployments` deliberately falls back to the
+    unfiltered list, so a request every deployment refused still comes back with
+    the provider's own error rather than a no-deployments one.
     """
     if not excluded_deployment_ids:
         return healthy_deployments
diff --git a/tests/e2e/coverage_registry/reliability.yaml b/tests/e2e/coverage_registry/reliability.yaml
index b50551ec105..6b69677d490 100644
--- a/tests/e2e/coverage_registry/reliability.yaml
+++ b/tests/e2e/coverage_registry/reliability.yaml
@@ -7,7 +7,7 @@
 - {id: reliability.retry.timeout.succeeds_within_retries, module: reliability, tier: P0, behavior: retry, variant: timeout, assertions: [succeeds_within_retries], exercised_on: [chat_completions, messages], source: "get_retry_from_policy.py:44", rationale: "Timeout retried per policy"}
 - {id: reliability.retry.429.succeeds_within_retries, module: reliability, tier: P0, behavior: retry, variant: "429", assertions: [succeeds_within_retries], exercised_on: [chat_completions, messages], source: "get_retry_from_policy.py:46", rationale: "429 retried per RateLimitErrorRetries policy"}
 - {id: reliability.retry.auth.succeeds_within_retries, module: reliability, tier: P1, behavior: retry, variant: auth, assertions: [succeeds_within_retries], exercised_on: [chat_completions, messages], source: "get_retry_from_policy.py:42", rationale: "Transient auth glitch retry"}
-- {id: reliability.retry.context_window.succeeds_within_retries, module: reliability, tier: P1, behavior: retry, variant: context_window, assertions: [succeeds_within_retries], exercised_on: [chat_completions, messages], source: "get_retry_from_policy.py:51", rationale: "Multi-attempt on context error"}
+- {id: reliability.retry.context_window.succeeds_within_retries, module: reliability, tier: P1, behavior: retry, variant: context_window, assertions: [succeeds_within_retries], exercised_on: [chat_completions, messages], source: "get_retry_from_policy.py:51", fail_before_fix: proven, rationale: "A context-window 400 under BadRequestErrorRetries retries onto a sibling deployment in the same model group, instead of coming straight back as the 400 the deployment that just refused it returned"}
 - {id: reliability.cooldown.5xx.trips_then_recovers, module: reliability, tier: P0, behavior: cooldown, variant: "5xx", assertions: [trips_then_recovers], exercised_on: [chat_completions, messages], source: "cooldown_handlers.py:40", rationale: "Deployment cools after repeated 5xx, recovers after cooldown_time"}
 - {id: reliability.cooldown.429.trips_then_recovers, module: reliability, tier: P0, behavior: cooldown, variant: "429", assertions: [trips_then_recovers], exercised_on: [chat_completions, messages], source: "cooldown_handlers.py:69", rationale: "Cools on 429, avoids hammering exhausted provider"}
 - {id: reliability.cooldown.auth.trips_then_recovers, module: reliability, tier: P1, behavior: cooldown, variant: auth, assertions: [trips_then_recovers], exercised_on: [chat_completions, messages], source: "cooldown_handlers.py:74", rationale: "Cools on 401 auth error"}

From b1cc1e8dae77a6ff1d0517bcedde40cddd55a9d4 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Sun, 6 Sep 2026 00:59:18 -0700
Subject: [PATCH 55/79] test(ui): assert navigation through accessible links
 and page content

---
 tests/e2e/ui/fixtures/migratedPages.ts        | 190 +++++++++++++-----
 tests/e2e/ui/helpers/navigation.ts            |  24 +++
 tests/e2e/ui/tests/migration/README.md        |  31 ++-
 .../ui/tests/migration/migratedPages.spec.ts  | 126 +++++-------
 tests/e2e/ui/tests/navigation/sidebar.spec.ts |  59 +++++-
 5 files changed, 282 insertions(+), 148 deletions(-)

diff --git a/tests/e2e/ui/fixtures/migratedPages.ts b/tests/e2e/ui/fixtures/migratedPages.ts
index 58939ca2b9a..bce09b49e10 100644
--- a/tests/e2e/ui/fixtures/migratedPages.ts
+++ b/tests/e2e/ui/fixtures/migratedPages.ts
@@ -1,50 +1,142 @@
-/**
- * Source of truth for the App Router migration E2E suites.
- *
- * Add an entry (legacy sidebar page id -> route segment) once a page's migration
- * has MERGED to the branch under test. Consumers pick it up automatically:
- *   - migration smoke (tests/migration/migratedPages.spec.ts), via MIGRATED_E2E_SEGMENTS:
- *       default mount:           npm run e2e:migration
- *       server-root-path mount:  SERVER_ROOT_PATH=/<root> npm run e2e:migration:root
- *   - navigation specs that assert per-page URLs (tests/navigation/sidebar.spec.ts)
- *
- * Keep this in lockstep with MIGRATED_PAGES in src/utils/migratedPages.ts.
- */
-export const MIGRATED_E2E_PAGES: Record<string, string> = {
-  "api-keys": "api-keys",
-  models: "models-and-endpoints",
-  api_ref: "api-reference",
-  "llm-playground": "playground",
-  projects: "projects",
-  "access-groups": "access-groups",
-  budgets: "budgets",
-  workflows: "workflows",
-  "guardrails-monitor": "guardrails-monitor",
-  "mcp-servers": "mcp-servers",
-  "search-tools": "search-tools",
-  "tag-management": "tag-management",
-  "vector-stores": "vector-stores",
-  memory: "memory",
-  policies: "policies",
-  guardrails: "guardrails",
-  prompts: "prompts",
-  "tool-policies": "tool-policies",
-  skills: "skills",
-  caching: "caching",
-  "cost-tracking": "cost-tracking",
-  "transform-request": "transform-request",
-  "ui-theme": "ui-theme",
-  logs: "logs",
-  "admin-panel": "admin-panel",
-  "logging-and-alerts": "logging-and-alerts",
-  "model-hub-table": "model-hub-table",
-  new_usage: "usage",
-  usage: "old-usage",
-  agents: "agents",
-  "router-settings": "router-settings",
-  users: "users",
-  teams: "teams",
-  organizations: "organizations",
-};
+export type MigratedPage = Readonly<{
+  segment: string;
+  linkName: string | RegExp;
+  group?: string;
+  content: Readonly<{ role: "heading" | "tab" | "button"; name: string }> | Readonly<{ text: string }>;
+  unlicensedText?: string;
+}>;
 
-export const MIGRATED_E2E_SEGMENTS: string[] = [...new Set(Object.values(MIGRATED_E2E_PAGES))];
+export const MIGRATED_E2E_PAGES: Readonly<Record<string, MigratedPage>> = {
+  "api-keys": { segment: "api-keys", linkName: "Virtual Keys", content: { role: "heading", name: "Virtual Keys" } },
+  models: {
+    segment: "models-and-endpoints",
+    linkName: "Models + Endpoints",
+    content: { role: "heading", name: "Model Management" },
+  },
+  api_ref: {
+    segment: "api-reference",
+    linkName: "API Reference",
+    content: { role: "heading", name: "OpenAI Compatible Proxy: API Reference" },
+  },
+  "llm-playground": { segment: "playground", linkName: "Playground", content: { role: "tab", name: "Chat" } },
+  projects: {
+    segment: "projects",
+    linkName: /^Projects(?: Beta)?$/,
+    content: { role: "heading", name: "Projects" },
+  },
+  "access-groups": {
+    segment: "access-groups",
+    linkName: "Access Groups",
+    content: { role: "heading", name: "Access Groups" },
+  },
+  budgets: { segment: "budgets", linkName: "Budgets", content: { role: "heading", name: "Budgets" } },
+  workflows: {
+    segment: "workflows",
+    linkName: "Workflow Runs",
+    group: "Agentic",
+    content: { text: "Workflow Runs" },
+  },
+  "guardrails-monitor": {
+    segment: "guardrails-monitor",
+    linkName: "Guardrails Monitor",
+    content: { role: "heading", name: "Guardrails Monitor" },
+  },
+  "mcp-servers": {
+    segment: "mcp-servers",
+    linkName: "MCP Servers",
+    content: { role: "heading", name: "MCP Servers" },
+  },
+  "search-tools": {
+    segment: "search-tools",
+    linkName: "Search Tools",
+    group: "Tools",
+    content: { role: "heading", name: "Search Tools" },
+  },
+  "tag-management": {
+    segment: "tag-management",
+    linkName: "Tag Management",
+    group: "Experimental",
+    content: { role: "heading", name: "Tag Management" },
+  },
+  "vector-stores": {
+    segment: "vector-stores",
+    linkName: "Vector Stores",
+    group: "Tools",
+    content: { role: "heading", name: "Vector Store Management" },
+  },
+  memory: { segment: "memory", linkName: "Memory", group: "Agentic", content: { role: "heading", name: "Memory" } },
+  policies: { segment: "policies", linkName: "Policies", content: { role: "tab", name: "Policy Simulator" } },
+  guardrails: { segment: "guardrails", linkName: "Guardrails", content: { role: "tab", name: "Guardrails" } },
+  prompts: {
+    segment: "prompts",
+    linkName: "Prompts",
+    group: "Experimental",
+    content: { role: "button", name: "Add New Prompt" },
+  },
+  "tool-policies": {
+    segment: "tool-policies",
+    linkName: "Tool Policies",
+    group: "Tools",
+    content: { role: "heading", name: "Tool Policies" },
+  },
+  skills: { segment: "skills", linkName: "Skills", content: { role: "heading", name: "Skills" } },
+  caching: { segment: "caching", linkName: "Response Cache", content: { role: "tab", name: "Cache Settings" } },
+  "cost-tracking": {
+    segment: "cost-tracking",
+    linkName: "Cost Tracking",
+    group: "Settings",
+    content: { text: "Cost Tracking Settings" },
+  },
+  "transform-request": {
+    segment: "transform-request",
+    linkName: "API Playground",
+    group: "Experimental",
+    content: { role: "heading", name: "Playground" },
+  },
+  "ui-theme": {
+    segment: "ui-theme",
+    linkName: "UI Theme",
+    group: "Settings",
+    content: { role: "heading", name: "UI Theme Customization" },
+  },
+  logs: { segment: "logs", linkName: "Logs", content: { role: "heading", name: "Request Logs" } },
+  "admin-panel": {
+    segment: "admin-panel",
+    linkName: "Admin Settings",
+    group: "Settings",
+    content: { role: "heading", name: "Admin Access" },
+  },
+  "logging-and-alerts": {
+    segment: "logging-and-alerts",
+    linkName: "Logging & Alerts",
+    group: "Settings",
+    content: { role: "tab", name: "Logging Callbacks" },
+  },
+  "model-hub-table": {
+    segment: "model-hub-table",
+    linkName: "AI Hub",
+    content: { role: "heading", name: "AI Hub" },
+  },
+  new_usage: { segment: "usage", linkName: "Usage", content: { role: "heading", name: "Usage View" } },
+  usage: {
+    segment: "old-usage",
+    linkName: "Old Usage",
+    group: "Experimental",
+    content: { role: "tab", name: "All Up" },
+  },
+  agents: { segment: "agents", linkName: "Agents", group: "Agentic", content: { role: "heading", name: "Agents" } },
+  "router-settings": {
+    segment: "router-settings",
+    linkName: "Router Settings",
+    group: "Settings",
+    content: { role: "heading", name: "Routing Settings" },
+  },
+  users: { segment: "users", linkName: "Internal Users", content: { role: "tab", name: "Users" } },
+  teams: { segment: "teams", linkName: "Teams", content: { role: "heading", name: "Teams" } },
+  organizations: {
+    segment: "organizations",
+    linkName: "Organizations",
+    content: { text: "Click on an organization ID to view its details." },
+    unlicensedText: "This is a LiteLLM Enterprise feature, and requires a valid key to use. Get a trial key here.",
+  },
+};
diff --git a/tests/e2e/ui/helpers/navigation.ts b/tests/e2e/ui/helpers/navigation.ts
index a58ece16f9c..4a7c4e7baa9 100644
--- a/tests/e2e/ui/helpers/navigation.ts
+++ b/tests/e2e/ui/helpers/navigation.ts
@@ -1,5 +1,29 @@
 import { Page } from "../fixtures/pages";
 import { Page as PlaywrightPage, expect } from "@playwright/test";
+import { UI_BASE_URL } from "../constants";
+
+export const sidebarLink = (page: PlaywrightPage, name: string | RegExp) =>
+  page.getByRole("complementary").getByRole("link", { name, exact: true });
+
+export async function clickSidebarLink(page: PlaywrightPage, name: string | RegExp, groupName?: string): Promise<void> {
+  const link = sidebarLink(page, name);
+  if (groupName && !(await link.isVisible())) {
+    const group = page.getByRole("complementary").getByRole("button", { name: groupName, exact: true });
+    await expect(group).toBeVisible();
+    if ((await group.getAttribute("aria-expanded")) === "false") {
+      await group.click();
+    }
+  }
+  await link.click();
+}
+
+export async function expectUiRoute(page: PlaywrightPage, segment: string): Promise<void> {
+  const root = (process.env.SERVER_ROOT_PATH ?? "").replace(/\/+$/, "");
+  const expected = new URL(`${root}/ui/${segment}`, UI_BASE_URL);
+  await expect(page, `navigate to ${expected.pathname}`).toHaveURL(
+    (url) => url.origin === expected.origin && url.pathname.replace(/\/+$/, "") === expected.pathname,
+  );
+}
 
 /**
  * Navigates to a specific page using the page query parameter.
diff --git a/tests/e2e/ui/tests/migration/README.md b/tests/e2e/ui/tests/migration/README.md
index d6b33598ec4..59933502463 100644
--- a/tests/e2e/ui/tests/migration/README.md
+++ b/tests/e2e/ui/tests/migration/README.md
@@ -1,17 +1,25 @@
 # App Router migration smoke
 
 A growing E2E smoke for pages migrated from the legacy `?page=` switch to App
-Router path routes. For each migrated page it clicks the page's sidebar link, checks
-the URL is the path route and the page renders, reloads it, then clicks off to a
-legacy page and back to confirm navigation still works. It runs in two situations:
-the default mount and a non-root `SERVER_ROOT_PATH` mount.
+Router path routes. For each page it clicks the sidebar link by its accessible
+name, verifies the destination's content, reloads it, then visits Virtual Keys
+and returns. It runs at the default mount and a non-root `SERVER_ROOT_PATH` mount
+
+Link selection does not depend on `href` formatting. URL assertions compare the
+origin and pathname, allowing a trailing slash, query string, and fragment while
+rejecting another route or mount. Reloads must return a successful document,
+and each journey must finish without uncaught browser errors
 
 ## Adding a page
 
-When a page's migration merges, add its route segment to
-`tests/e2e/ui/fixtures/migratedPages.ts` (keep it in lockstep with `MIGRATED_PAGES`
-in `ui/litellm-dashboard/src/utils/migratedPages.ts`). Both suites pick it up
-automatically.
+Add an entry to `tests/e2e/ui/fixtures/migratedPages.ts`, keyed by the legacy page
+ID. Specify its route segment, accessible link name, sidebar group if collapsed,
+and distinctive visible content such as a heading or tab. Keep expectations
+independent of the application's route table so an incorrect destination fails
+the test. Both navigation suites use this fixture
+
+For a licensed-only page, `unlicensedText` describes the expected upgrade notice.
+The authenticated session's license claim determines which content must render
 
 ## Running
 
@@ -31,4 +39,9 @@ SERVER_ROOT_PATH=/litellm npm run e2e:migration:root
 ```
 
 `globalSetup` logs in once per role; the admin storage state is reused for these
-tests. Under a non-root mount it logs in at `${SERVER_ROOT_PATH}/ui/login`.
+tests. Under a non-root mount it logs in at `${SERVER_ROOT_PATH}/ui/login`
+
+`tests/navigation/sidebar.spec.ts` also checks the navigation helpers against
+equivalent link formats on the live dashboard and a deep link containing a query
+string and fragment. The link-format cases change only the rendered `href`
+attribute to exercise the locator contract; destination pages and APIs remain live
diff --git a/tests/e2e/ui/tests/migration/migratedPages.spec.ts b/tests/e2e/ui/tests/migration/migratedPages.spec.ts
index bc96d9a6c66..f2dbd2dbc77 100644
--- a/tests/e2e/ui/tests/migration/migratedPages.spec.ts
+++ b/tests/e2e/ui/tests/migration/migratedPages.spec.ts
@@ -1,105 +1,73 @@
 import { test, expect, type Page } from "@playwright/test";
-import { MIGRATED_E2E_SEGMENTS } from "../../fixtures/migratedPages";
+import { MIGRATED_E2E_PAGES, type MigratedPage } from "../../fixtures/migratedPages";
 import { ADMIN_STORAGE_PATH } from "../../constants";
-import { dismissFeedbackPopup } from "../../helpers/navigation";
+import { clickSidebarLink, dismissFeedbackPopup, expectUiRoute, sidebarLink } from "../../helpers/navigation";
+import { proxyIsPremium } from "../../helpers/premium";
 
-/**
- * App Router migration smoke as a user journey: start where the proxy lands you,
- * click a migrated page in the sidebar, confirm it routed and rendered, reload it
- * (the check a wrong server_root_path breaks), bounce to a legacy page and back,
- * and, once two pages are migrated, navigate directly between two migrated pages.
- *
- * Driven by MIGRATED_E2E_SEGMENTS, so it grows as pages are migrated. Set
- * SERVER_ROOT_PATH (e.g. "/litellm") to exercise the non-root mount; leave it
- * unset for the default mount. Boot the proxy with the matching value first.
- */
-const ROOT = process.env.SERVER_ROOT_PATH ?? "";
+const ROOT = (process.env.SERVER_ROOT_PATH ?? "").replace(/\/+$/, "");
+const apiKeys = MIGRATED_E2E_PAGES["api-keys"];
 
-const esc = (s: string) => s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
-const pathRe = (segment: string) => new RegExp(`${esc(ROOT)}/ui/${esc(segment)}/?($|\\?)`);
-// Scope nav lookups to the sidebar (a `complementary` landmark). The top bar
-// now renders a breadcrumb whose current-page item is also a "Virtual Keys"
-// link, so an unscoped locator would match two elements.
-const sidebar = (page: Page) => page.getByRole("complementary");
-const virtualKeysLink = (page: Page) => sidebar(page).getByRole("link", { name: "Virtual Keys", exact: true });
-
-/** The dashboard shell is present (sidebar rendered); page didn't 404 / crash. */
-async function expectRendered(page: Page) {
-  await expect(virtualKeysLink(page)).toBeVisible({ timeout: 20_000 });
+async function expectContent(page: Page, destination: MigratedPage): Promise<void> {
+  await expect(sidebarLink(page, apiKeys.linkName)).toBeVisible({ timeout: 20_000 });
+  const main = page.getByRole("main");
+  if (destination.unlicensedText && !proxyIsPremium()) {
+    await expect(main.getByText(destination.unlicensedText, { exact: true })).toBeVisible();
+    return;
+  }
+  const content = destination.content;
+  const landmark =
+    "role" in content
+      ? main.getByRole(content.role, { name: content.name, exact: true })
+      : main.getByText(content.text, { exact: true });
+  await expect(landmark).toBeVisible();
 }
 
-/**
- * Click a migrated page's sidebar link. Migrated items render as <a href=".../ui/<segment>/">;
- * nested ones live under collapsible groups whose children only render while the
- * group is open, so expand collapsed groups until the link is clickable.
- */
-async function clickSidebar(page: Page, segment: string) {
-  const link = sidebar(page).locator(`a[href$="/ui/${segment}"], a[href$="/ui/${segment}/"]`).first();
-  const collapsedGroups = sidebar(page).getByRole("button", { expanded: false });
-  for (let i = 0; i < 8 && !(await link.isVisible().catch(() => false)); i++) {
-    const stillCollapsed = await collapsedGroups.count();
-    if (stillCollapsed === 0) break;
-    await collapsedGroups.first().click();
-    await expect(collapsedGroups).toHaveCount(stillCollapsed - 1);
-  }
-  await link.click();
+async function navigateToDestination(page: Page, destination: MigratedPage): Promise<void> {
+  await clickSidebarLink(page, destination.linkName, destination.group);
+  await expectUiRoute(page, destination.segment);
+  await dismissFeedbackPopup(page);
+  await expectContent(page, destination);
 }
 
 test.use({ storageState: ADMIN_STORAGE_PATH });
 
 test.describe("App Router migrated pages", () => {
-  for (const segment of MIGRATED_E2E_SEGMENTS) {
-    test(`${segment}: sidebar nav, reload, and round-trip via the api-keys landing`, async ({ page }) => {
+  for (const destination of Object.values(MIGRATED_E2E_PAGES)) {
+    test(`${destination.segment}: sidebar nav, reload, and round-trip via the api-keys landing`, async ({ page }) => {
       const pageErrors: string[] = [];
-      page.on("pageerror", (e) => pageErrors.push(String(e)));
+      page.on("pageerror", (error) => pageErrors.push(String(error)));
 
-      // 1. Start where the proxy lands us.
-      await page.goto(`${ROOT}/ui/`);
+      const landing = await page.goto(`${ROOT}/ui/`);
+      expect(landing?.ok(), "dashboard document loads successfully").toBe(true);
       await dismissFeedbackPopup(page);
-      await expectRendered(page);
+      await expectContent(page, apiKeys);
 
-      // 2. Click the migrated page in the sidebar -> path route + rendered.
-      await clickSidebar(page, segment);
-      await expect(page).toHaveURL(pathRe(segment));
-      await expectRendered(page);
-      // 3. Reload the path route directly; a wrong server_root_path 404s here.
-      await page.reload();
+      await navigateToDestination(page, destination);
+
+      const reloaded = await page.reload();
+      expect(reloaded?.ok(), `${destination.segment} document loads on reload`).toBe(true);
+      await expectUiRoute(page, destination.segment);
       await dismissFeedbackPopup(page);
-      await expect(page).toHaveURL(pathRe(segment));
-      await expectRendered(page);
-      // 4. Click the Virtual Keys sidebar link to the api-keys landing (now a path route), then back.
-      await virtualKeysLink(page).click();
-      await expect(page).toHaveURL(pathRe("api-keys"));
-      await dismissFeedbackPopup(page);
-      await expectRendered(page);
-      // 5. Click back to the migrated page.
-      await clickSidebar(page, segment);
-      await expect(page).toHaveURL(pathRe(segment));
-      await expectRendered(page);
-      expect(pageErrors, `page errors during ${segment} journey`).toEqual([]);
+      await expectContent(page, destination);
+
+      await navigateToDestination(page, apiKeys);
+      await navigateToDestination(page, destination);
+      expect(pageErrors, `page errors during ${destination.segment} journey`).toEqual([]);
     });
   }
 
   test("navigates directly between two migrated pages", async ({ page }) => {
-    test.skip(MIGRATED_E2E_SEGMENTS.length < 2, "needs >= 2 migrated pages");
-    const [first, second] = MIGRATED_E2E_SEGMENTS;
     const pageErrors: string[] = [];
-    page.on("pageerror", (e) => pageErrors.push(String(e)));
+    page.on("pageerror", (error) => pageErrors.push(String(error)));
 
-    await page.goto(`${ROOT}/ui/`);
+    const landing = await page.goto(`${ROOT}/ui/`);
+    expect(landing?.ok(), "dashboard document loads successfully").toBe(true);
     await dismissFeedbackPopup(page);
+    await expectContent(page, apiKeys);
 
-    await clickSidebar(page, first);
-    await expect(page).toHaveURL(pathRe(first));
-    await expectRendered(page);
-    await clickSidebar(page, second);
-    await expect(page).toHaveURL(pathRe(second));
-    await expectRendered(page);
-    // Back to the first migrated page.
-    await clickSidebar(page, first);
-    await expect(page).toHaveURL(pathRe(first));
-    await expectRendered(page);
-
-    expect(pageErrors, "page errors during migrated -> migrated nav").toEqual([]);
+    for (const destination of [apiKeys, MIGRATED_E2E_PAGES.models, apiKeys]) {
+      await navigateToDestination(page, destination);
+    }
+    expect(pageErrors, "page errors during migrated page navigation").toEqual([]);
   });
 });
diff --git a/tests/e2e/ui/tests/navigation/sidebar.spec.ts b/tests/e2e/ui/tests/navigation/sidebar.spec.ts
index b220dc09ae2..eaa4985c659 100644
--- a/tests/e2e/ui/tests/navigation/sidebar.spec.ts
+++ b/tests/e2e/ui/tests/navigation/sidebar.spec.ts
@@ -3,7 +3,13 @@ import { Role } from "../../fixtures/roles";
 import { ADMIN_STORAGE_PATH } from "../../constants";
 import { Page } from "../../fixtures/pages";
 import { menuLabelToPage } from "../../fixtures/menuMappings";
-import { navigateToPage } from "../../helpers/navigation";
+import {
+  clickSidebarLink,
+  dismissFeedbackPopup,
+  expectUiRoute,
+  navigateToPage,
+  sidebarLink,
+} from "../../helpers/navigation";
 import { MIGRATED_E2E_PAGES } from "../../fixtures/migratedPages";
 import type { Page as PlaywrightPage } from "@playwright/test";
 
@@ -11,7 +17,7 @@ const sidebarButtons = {
   [Role.ProxyAdmin]: [
     "Virtual Keys",
     "Playground",
-    "Models",
+    "Models + Endpoints",
     "Usage",
     "Teams",
     "Internal Users",
@@ -22,9 +28,9 @@ const sidebarButtons = {
 
 /** Migrated pages live at a path route; legacy pages keep the ?page= query param. */
 async function expectPageUrl(page: PlaywrightPage, pageKey: string): Promise<void> {
-  const migratedSegment = MIGRATED_E2E_PAGES[pageKey];
-  if (migratedSegment) {
-    await expect(page).toHaveURL(new RegExp(`/ui/${migratedSegment}/?($|\\?)`));
+  const migratedPage = MIGRATED_E2E_PAGES[pageKey];
+  if (migratedPage) {
+    await expectUiRoute(page, migratedPage.segment);
   } else {
     await expect(page).toHaveURL(new RegExp(`[?&]page=${pageKey}(&|$)`));
   }
@@ -51,12 +57,7 @@ for (const { role, storage } of roles) {
           throw new Error(`No page mapping found for menu label: ${buttonLabel}`);
         }
 
-        // Sidebar items are links inside the `complementary` landmark; scoping
-        // there avoids the top-bar breadcrumb, which also links the page name.
-        const tab = page.getByRole("complementary").getByRole("link", { name: buttonLabel });
-        await expect(tab).toBeVisible();
-
-        await tab.click();
+        await clickSidebarLink(page, buttonLabel);
 
         await expectPageUrl(page, expectedPage);
       }
@@ -81,5 +82,41 @@ for (const { role, storage } of roles) {
       await navigateToPage(page, Page.LlmPlayground);
       await expectPageUrl(page, Page.LlmPlayground);
     });
+
+    for (const format of ["without trailing slash", "absolute with query and fragment", "relative"] as const) {
+      test(`sidebar locator tolerates hrefs ${format}`, async ({ page }) => {
+        await page.goto("/ui/");
+        await dismissFeedbackPopup(page);
+        const link = sidebarLink(page, "Models + Endpoints");
+        await expect(link).toBeVisible();
+        const destination = new URL("/ui/models-and-endpoints/", page.url());
+        const href =
+          format === "without trailing slash"
+            ? destination.pathname.replace(/\/$/, "")
+            : format === "relative"
+              ? "./models-and-endpoints/"
+              : `${destination.href}?source=navigation-smoke#overview`;
+
+        await link.evaluate((element, value) => element.setAttribute("href", value), href);
+        await expect(link).toHaveAttribute("href", href);
+        await clickSidebarLink(page, "Models + Endpoints");
+
+        await expectUiRoute(page, "models-and-endpoints");
+        await expect(
+          page.getByRole("main").getByRole("heading", { name: "Model Management", exact: true }),
+        ).toBeVisible();
+      });
+    }
+
+    test("route assertion tolerates a query string and fragment on a deep link", async ({ page }) => {
+      const response = await page.goto("/ui/models-and-endpoints/?source=navigation-smoke#overview");
+      expect(response?.ok()).toBe(true);
+      await expectUiRoute(page, "models-and-endpoints");
+      await expect(
+        page.getByRole("main").getByRole("heading", { name: "Model Management", exact: true }),
+      ).toBeVisible();
+      expect(new URL(page.url()).search).toBe("?source=navigation-smoke");
+      expect(new URL(page.url()).hash).toBe("#overview");
+    });
   });
 }

From a72041b7572fad9e357e67dc4dfb913822d428ce Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 01:03:32 -0700
Subject: [PATCH 56/79] test(router): pin the retry skip list across attempts
 in a model group

---
 tests/test_litellm/test_router.py | 48 +++++++++++++++++++++++++++++++
 1 file changed, 48 insertions(+)

diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py
index 05ac672691e..a7d4e66bf73 100644
--- a/tests/test_litellm/test_router.py
+++ b/tests/test_litellm/test_router.py
@@ -39,6 +39,7 @@ from litellm.router import (
     _anthropic_stream_should_drop_pre_content_ping,
     _is_retriable_anthropic_status,
 )
+from litellm.router_strategy import simple_shuffle
 from litellm.types.router import DeploymentTypedDict
 
 
@@ -13423,6 +13424,53 @@ async def test_router_retry_policy_400_keeps_upstream_error_when_tags_narrow_the
     assert untagged.call_count == 0
 
 
+@pytest.mark.asyncio
+async def test_router_retry_policy_400_never_returns_to_a_deployment_that_already_refused(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    monkeypatch.setattr(simple_shuffle.random, "choice", lambda deployments: deployments[0])
+    router = litellm.Router(
+        model_list=[
+            _retry_skip_deployment("first-refuser", "first-refuser", litellm_params={"weight": 1}),
+            _retry_skip_deployment("second-refuser", "second-refuser", litellm_params={"weight": 0}),
+            _retry_skip_deployment("accepting", "accepting", litellm_params={"weight": 0}),
+        ],
+        num_retries=3,
+        retry_policy={"BadRequestErrorRetries": 3},
+        disable_cooldowns=True,
+    )
+
+    with respx.mock as respx_mock:
+        first = respx_mock.post("https://first-refuser.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        second = respx_mock.post("https://second-refuser.local/v1/chat/completions").mock(
+            return_value=httpx.Response(400, json={"error": _UPSTREAM_400})
+        )
+        accepting = respx_mock.post("https://accepting.local/v1/chat/completions").mock(
+            return_value=httpx.Response(
+                200,
+                json={
+                    "id": "chatcmpl-lit-7036",
+                    "object": "chat.completion",
+                    "created": 1,
+                    "model": "gpt-5.6",
+                    "choices": [
+                        {"index": 0, "message": {"role": "assistant", "content": "hi back"}, "finish_reason": "stop"}
+                    ],
+                    "usage": {"prompt_tokens": 1, "completion_tokens": 2, "total_tokens": 3},
+                },
+            )
+        )
+        response = await router.acompletion(model="gpt-5.6", messages=[{"role": "user", "content": "hi"}])
+
+    assert first.call_count == 1
+    assert second.call_count == 1
+    assert accepting.call_count == 1
+    assert response.choices[0].message.content == "hi back"
+
+
 def _make_failure_logging_obj():
     return LiteLLMLogging(
         model="gpt-5.6",

From 5ca9e2605081d26503786cdd116e7fc8245c8903 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 01:09:36 -0700
Subject: [PATCH 57/79] fix(ci): leave check-run names the sweep cannot resolve
 out of the comparison

A job name holding an expression the sweep could not resolve was compared as
if it were the published name. Two jobs whose names differ per matrix value or
per caller input were reported as a collision, and a matrix that was itself an
expression collapsed onto the bare job id and did the same.

Model what a job publishes as known names beside the reasons the rest stay
unknown. Anything the sweep cannot work out contributes no name and is
reported as a note instead of guessed at. An expression over contexts that are
fixed for the whole run still compares, so two jobs sharing one of those are
still caught.
---
 .../check_workflow_job_name_collisions.py     | 173 ++++++++++++++----
 .../test_workflow_job_name_collisions.py      | 151 ++++++++++++++-
 2 files changed, 280 insertions(+), 44 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 4eef3df1c11..1e81acc4f90 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -20,16 +20,24 @@ with the other rows', so two shard lists that overlap collide even though their
 templates read differently. Each expression is evaluated per combination over the
 pieces a job name can hold: string literals, `matrix.<key>`, `format()`, `==` and
 `!=`, and the `<cond> && <a> || <b>` idiom, which is how the shards reach their
-real `<shard> / Run tests` names rather than staying opaque. An expression
-nothing resolves stays in the string, so two jobs carrying the same unresolved
-template still compare equal and their collision is still caught.
+real `<shard> / Run tests` names rather than staying opaque.
+
+Whatever the sweep cannot work out is left out of the comparison and reported
+instead of guessed, because a guess that lands wrong fails a workflow GitHub
+would have published perfectly well. A name left holding an expression over
+`inputs`, `needs` or `matrix` reads differently from every job that runs it, so
+it is one of those; an expression over `github` and the other contexts fixed for
+the whole run reads the same everywhere, so two jobs carrying it still clash and
+the check still says so. A matrix that is itself an expression or that lists
+values which are not scalars, and a call this sweep cannot follow, go in the same
+bucket. The cost is that a real clash hiding behind one of them goes unseen,
+which leaves a merge no worse off than before this check existed, where the
+opposite direction would block work that was fine.
 
 A job calling a local reusable workflow publishes one check run per job of the
 callee, named `<caller> / <callee>` and chained through however many levels of
 local calls it takes, which is why a caller's name never collides with a plain
-job that happens to match it. A call this sweep cannot follow, to a workflow
-outside the repo or to one that is not there, publishes nothing rather than a
-name GitHub never posts. A file under `.github/workflows/` that does not read as
+job that happens to match it. A file under `.github/workflows/` that does not read as
 a workflow at all is reported rather than skipped, since skipping it silently
 would hide every job it holds.
 """
@@ -54,6 +62,8 @@ MATRIX_REF: Final = re.compile(r"^matrix\.(?P<key>[\w-]+)$")
 LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
 FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
 COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
+CONTEXT_REF: Final = re.compile(r"\b(?P<root>[a-z][\w-]*)\s*\.")
+RUN_FIXED_CONTEXTS: Final = frozenset({"github", "vars", "env", "runner", "secrets"})
 NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
 NO_CALLERS: Final[frozenset[str]] = frozenset()
 SCALAR: Final = (str, int, float)
@@ -66,6 +76,19 @@ class Unreadable:
     reason: str
 
 
+@dataclass(frozen=True, slots=True)
+class Opaque:
+    reason: str
+
+
+@dataclass(frozen=True, slots=True)
+class Names:
+    """The check-run names a job publishes, beside the reasons the rest of them stay unknown."""
+
+    known: tuple[str, ...] = ()
+    unknown: tuple[str, ...] = ()
+
+
 class Job(BaseModel):
     name: object = None
     uses: str | None = None
@@ -107,12 +130,23 @@ def publishes_check_runs(raw_on: object) -> bool:
     return events(raw_on) != frozenset({"workflow_call"})
 
 
-def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, ...]], ...]:
-    return tuple(
-        (str(key), tuple(scalar_text(value) for value in values if isinstance(value, SCALAR)))
-        for key, values in matrix.items()
-        if str(key) not in MATRIX_DIRECTIVES and isinstance(values, Sequence) and not isinstance(values, str)
+def scalar_list(value: object) -> tuple[str, ...] | Opaque:
+    """One matrix key's values, or why the combinations it produces cannot be worked out."""
+    if not isinstance(value, Sequence) or isinstance(value, str):
+        return Opaque("a matrix key holds something other than a list of values")
+    if any(not isinstance(item, SCALAR) for item in value):
+        return Opaque("a matrix key lists values that are not plain scalars")
+    return tuple(scalar_text(item) for item in value)
+
+
+def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, ...]], ...] | Opaque:
+    listed: Final = tuple(
+        (str(key), scalar_list(values)) for key, values in matrix.items() if str(key) not in MATRIX_DIRECTIVES
     )
+    opaque: Final = next((values for _, values in listed if isinstance(values, Opaque)), None)
+    if opaque is not None:
+        return opaque
+    return tuple((key, values) for key, values in listed if not isinstance(values, Opaque))
 
 
 def directive_rows(matrix: Mapping[str, object], directive: str) -> tuple[Mapping[str, str], ...]:
@@ -150,17 +184,20 @@ def crossed_values(listed: Sequence[tuple[str, tuple[str, ...]]]) -> tuple[Mappi
     )
 
 
-def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...]:
+def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...] | Opaque:
     """One mapping per job the matrix produces, `exclude` applied before `include` as GitHub does."""
     matrix: Final = job.strategy.get("matrix")
-    if not isinstance(matrix, Mapping):
+    if matrix is None:
         return ()
+    if not isinstance(matrix, Mapping):
+        return Opaque("the matrix itself comes from an expression")
+    listed: Final = listed_values(matrix)
+    if isinstance(listed, Opaque):
+        return listed
     rows: Final = directive_rows(matrix, "include")
     dropped: Final = directive_rows(matrix, "exclude")
     kept: Final = tuple(
-        combination
-        for combination in crossed_values(listed_values(matrix))
-        if not any(drops(row, combination) for row in dropped)
+        combination for combination in crossed_values(listed) if not any(drops(row, combination) for row in dropped)
     )
     standalone: Final = tuple(row for row in rows if not any(extends(row, combination) for combination in kept))
     return (*(extended(combination, rows) for combination in kept), *standalone)
@@ -243,11 +280,33 @@ def rendered(template: str, values: Mapping[str, str]) -> str:
     return EXPRESSION.sub(lambda span: resolved_span(span, values), template)
 
 
-def expand(template: str, job: Job) -> tuple[str, ...]:
+def stable(name: str) -> bool:
+    """An expression over the run's own contexts reads the same from every job, so two of them still clash."""
+    return all(
+        reference.group("root") in RUN_FIXED_CONTEXTS
+        for span in EXPRESSION.finditer(name)
+        for reference in CONTEXT_REF.finditer(span.group("body"))
+    )
+
+
+def comparable(name: str) -> bool:
+    """A leftover expression over `inputs`, `needs` or `matrix` names a different check run per job."""
+    return EXPRESSION.search(name) is None or stable(name)
+
+
+def settled(names: Sequence[str]) -> Names:
+    return Names(
+        tuple(name for name in names if comparable(name)),
+        tuple(f"its name stays `{name}`" for name in names if not comparable(name)),
+    )
+
+
+def expand(template: str, job: Job) -> Names:
     combinations: Final = matrix_combinations(job)
-    if not combinations:
-        return (rendered(template, NO_MATRIX),)
-    return tuple(dict.fromkeys(rendered(template, values) for values in combinations))
+    if isinstance(combinations, Opaque):
+        return Names((), (combinations.reason,))
+    over: Final = combinations or (NO_MATRIX,)
+    return settled(tuple(dict.fromkeys(rendered(template, values) for values in over)))
 
 
 def suffixed(job_id: str, combination: Mapping[str, str]) -> str:
@@ -255,11 +314,14 @@ def suffixed(job_id: str, combination: Mapping[str, str]) -> str:
     return f"{job_id} ({', '.join(combination.values())})" if combination else job_id
 
 
-def published_names(job_id: str, job: Job) -> tuple[str, ...]:
+def published_names(job_id: str, job: Job) -> Names:
     if job.name is not None:
         return expand(scalar_text(job.name), job)
-    suffixes: Final = tuple(dict.fromkeys(suffixed(job_id, values) for values in matrix_combinations(job)))
-    return suffixes or (job_id,)
+    combinations: Final = matrix_combinations(job)
+    if isinstance(combinations, Opaque):
+        return Names((), (combinations.reason,))
+    suffixes: Final = tuple(dict.fromkeys(suffixed(job_id, values) for values in combinations))
+    return Names(suffixes or (job_id,))
 
 
 def callee_path(job: Job) -> str | None:
@@ -268,22 +330,40 @@ def callee_path(job: Job) -> str | None:
     return job.uses[len(LOCAL_CALL_PREFIX) :].split("@")[0]
 
 
-def job_names(
-    job_id: str, job: Job, workflows: Mapping[str, Workflow], callers: frozenset[str] = NO_CALLERS
-) -> tuple[str, ...]:
+def joined(groups: Sequence[Names]) -> Names:
+    return Names(
+        tuple(name for group in groups for name in group.known),
+        tuple(reason for group in groups for reason in group.unknown),
+    )
+
+
+def call_blocker(job: Job, workflows: Mapping[str, Workflow], callers: frozenset[str]) -> str | None:
+    path: Final = callee_path(job)
+    if path is None:
+        return "it calls a reusable workflow outside this repository"
+    if path in callers:
+        return f"its call to {path} loops back on itself"
+    return None if path in workflows else f"it calls {path}, which this checkout does not hold"
+
+
+def job_names(job_id: str, job: Job, workflows: Mapping[str, Workflow], callers: frozenset[str] = NO_CALLERS) -> Names:
     prefixes: Final = published_names(job_id, job)
     if job.uses is None:
         return prefixes
-    path: Final = callee_path(job)
-    callee: Final = workflows.get(path or "")
-    if path is None or callee is None or path in callers:
-        return ()
-    suffixes: Final = tuple(
-        name
-        for callee_id, callee_job in callee.jobs.items()
-        for name in job_names(callee_id, callee_job, workflows, callers | {path})
+    blocker: Final = call_blocker(job, workflows, callers)
+    if blocker is not None:
+        return Names((), (*prefixes.unknown, blocker))
+    path: Final = callee_path(job) or ""
+    suffixes: Final = joined(
+        tuple(
+            job_names(callee_id, callee_job, workflows, callers | {path})
+            for callee_id, callee_job in workflows[path].jobs.items()
+        )
+    )
+    return Names(
+        tuple(f"{prefix} / {suffix}" for prefix in prefixes.known for suffix in suffixes.known),
+        (*prefixes.unknown, *suffixes.unknown),
     )
-    return tuple(f"{prefix} / {suffix}" for prefix in prefixes for suffix in suffixes)
 
 
 def readable(sources: Mapping[str, str]) -> Mapping[str, tuple[Workflow, object]]:
@@ -300,15 +380,29 @@ def unreadable(sources: Mapping[str, str]) -> tuple[str, ...]:
     )
 
 
-def published(sources: Mapping[str, str]) -> Iterator[tuple[str, str]]:
+def scanned_jobs(sources: Mapping[str, str]) -> Iterator[tuple[str, str, Names]]:
     parsed: Final = readable(sources)
     workflows: Final = {rel: workflow for rel, (workflow, _) in parsed.items()}
     for rel, (workflow, raw_on) in parsed.items():
         if not publishes_check_runs(raw_on):
             continue
         for job_id, job in workflow.jobs.items():
-            for name in job_names(job_id, job, workflows):
-                yield name, f"{rel} job `{job_id}`"
+            yield rel, job_id, job_names(job_id, job, workflows)
+
+
+def published(sources: Mapping[str, str]) -> Iterator[tuple[str, str]]:
+    for rel, job_id, names in scanned_jobs(sources):
+        for name in names.known:
+            yield name, f"{rel} job `{job_id}`"
+
+
+def blind_spots(sources: Mapping[str, str]) -> tuple[str, ...]:
+    """Jobs whose published names GitHub decides at run time, which no offline sweep can compare."""
+    return tuple(
+        f"{rel} job `{job_id}` publishes a name this check cannot work out because {reason}."
+        for rel, job_id, names in scanned_jobs(sources)
+        for reason in sorted(names.unknown)
+    )
 
 
 def owners_by_name(sources: Mapping[str, str]) -> Iterator[tuple[str, tuple[str, ...]]]:
@@ -338,6 +432,9 @@ def report(header: str, problems: Sequence[str]) -> None:
 def exit_code(sources: Mapping[str, str]) -> int:
     unread: Final = unreadable(sources)
     found: Final = collisions(sources)
+    blind: Final = blind_spots(sources)
+    if blind:
+        print("NOTE: names left out of the comparison:\n  - " + "\n  - ".join(blind))
     report("Some workflows could not be read", unread)
     report("Check-run names are not unique", found)
     if unread or found:
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 3dd8a3d4088..39e61ff48d1 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -2,6 +2,7 @@ from typing import Final
 
 from check_workflow_job_name_collisions import (
     Unreadable,
+    blind_spots,
     callee_path,
     collisions,
     exit_code,
@@ -301,9 +302,9 @@ def test_a_matrix_list_supplies_values_the_same_way_include_rows_do() -> None:
     assert "`Analyze (go)` is published by 2 jobs" in found[0]
 
 
-def test_two_jobs_sharing_one_unresolvable_template_still_collide() -> None:
+def test_two_jobs_sharing_a_template_over_the_run_itself_still_collide() -> None:
     template: Final = (
-        "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ matrix.shard }}}}\n    runs-on: ubuntu-latest\n"
+        "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ github.event_name }}}}-build\n    runs-on: ubuntu-latest\n"
     )
     sources: Final = {
         "a.yml": template.format(job="one"),
@@ -316,6 +317,19 @@ def test_two_jobs_sharing_one_unresolvable_template_still_collide() -> None:
     assert "is published by 2 jobs" in found[0]
 
 
+def test_two_jobs_sharing_a_template_that_reads_per_job_are_not_called_a_collision() -> None:
+    template: Final = (
+        "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ matrix.shard }}}}\n    runs-on: ubuntu-latest\n"
+    )
+    sources: Final = {
+        "a.yml": template.format(job="one"),
+        "b.yml": template.format(job="two"),
+    }
+
+    assert collisions(sources) == ()
+    assert len(blind_spots(sources)) == 2
+
+
 def test_a_file_that_is_not_a_workflow_is_reported_rather_than_skipped() -> None:
     sources: Final = {
         "notes.yml": "just a string\n",
@@ -437,13 +451,14 @@ def test_a_boolean_matrix_value_renders_the_way_github_writes_it() -> None:
     assert names == frozenset({"cache true", "cache false"})
 
 
-def test_a_format_call_its_arguments_cannot_fill_leaves_the_name_unresolved() -> None:
-    names: Final = frozenset(name for name, _ in published({"unit.yml": UNFILLABLE_FORMAT}))
+def test_a_format_call_its_arguments_cannot_fill_publishes_nothing_to_compare() -> None:
+    sources: Final = {"unit.yml": UNFILLABLE_FORMAT}
 
-    assert names == frozenset({"${{ format('{0} {1}', matrix.shard) }}"})
+    assert frozenset(name for name, _ in published(sources)) == frozenset()
+    assert "its name stays" in blind_spots(sources)[0]
 
 
-def test_a_call_to_a_workflow_outside_the_repo_publishes_no_name() -> None:
+def test_a_call_to_a_workflow_outside_the_repo_is_reported_rather_than_guessed() -> None:
     sources: Final = {
         "a.yml": "on: pull_request\njobs:\n  unit:\n    uses: BerriAI/other/.github/workflows/base.yml@main\n",
         "b.yml": "on: pull_request\njobs:\n  unit:\n    runs-on: ubuntu-latest\n",
@@ -451,6 +466,7 @@ def test_a_call_to_a_workflow_outside_the_repo_publishes_no_name() -> None:
 
     assert frozenset(name for name, _ in published(sources)) == frozenset({"unit"})
     assert collisions(sources) == ()
+    assert "outside this repository" in blind_spots(sources)[0]
 
 
 def test_a_chain_of_local_reusable_calls_publishes_every_level_of_the_chain() -> None:
@@ -498,3 +514,126 @@ def test_the_check_passes_when_every_file_reads_and_every_name_is_unique() -> No
     }
 
     assert exit_code(sources) == 0
+
+
+def test_two_callers_of_one_reusable_workflow_named_from_its_inputs_do_not_collide() -> None:
+    sources: Final = {
+        ".github/workflows/callee.yml": (
+            "on:\n  workflow_call:\njobs:\n  run:\n    name: ${{ inputs.suite }}\n    runs-on: ubuntu-latest\n"
+        ),
+        "caller.yml": (
+            "on: pull_request\njobs:\n"
+            "  alpha:\n    name: A\n    uses: ./.github/workflows/callee.yml\n    with:\n      suite: alpha\n"
+            "  beta:\n    name: A\n    uses: ./.github/workflows/callee.yml\n    with:\n      suite: beta\n"
+        ),
+    }
+
+    assert collisions(sources) == ()
+    assert len(blind_spots(sources)) == 2
+
+
+def test_a_matrix_that_is_itself_an_expression_never_collapses_onto_the_bare_job_id() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  build:\n    strategy:\n"
+            "      matrix: ${{ fromJson(needs.plan.outputs.matrix) }}\n    runs-on: ubuntu-latest\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"build"})
+    assert collisions(sources) == ()
+    assert "the matrix itself comes from an expression" in blind_spots(sources)[0]
+
+
+def test_a_matrix_listing_objects_never_collapses_onto_the_bare_job_id() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  build:\n    strategy:\n      matrix:\n        target:\n"
+            "          - os: ubuntu\n          - os: windows\n    runs-on: ubuntu-latest\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"build"})
+    assert collisions(sources) == ()
+    assert "not plain scalars" in blind_spots(sources)[0]
+
+
+def test_a_call_to_a_workflow_file_the_checkout_does_not_hold_is_reported() -> None:
+    sources: Final = {"a.yml": "on: pull_request\njobs:\n  unit:\n    uses: ./.github/workflows/gone.yml\n"}
+
+    assert collisions(sources) == ()
+    assert "which this checkout does not hold" in blind_spots(sources)[0]
+
+
+def test_reusable_workflows_calling_each_other_in_a_loop_are_reported_not_followed() -> None:
+    sources: Final = {
+        ".github/workflows/a.yml": (
+            "on:\n  workflow_call:\njobs:\n  call:\n    name: A\n    uses: ./.github/workflows/b.yml\n"
+        ),
+        ".github/workflows/b.yml": (
+            "on:\n  workflow_call:\njobs:\n  call:\n    name: B\n    uses: ./.github/workflows/a.yml\n"
+        ),
+        "top.yml": "on: pull_request\njobs:\n  top:\n    name: Top\n    uses: ./.github/workflows/a.yml\n",
+    }
+
+    assert collisions(sources) == ()
+    assert any("loops back on itself" in spot for spot in blind_spots(sources))
+
+
+def test_a_caller_still_publishes_the_callee_jobs_it_can_read() -> None:
+    sources: Final = {
+        ".github/workflows/callee.yml": (
+            "on:\n  workflow_call:\njobs:\n"
+            "  lint:\n    name: Lint\n    runs-on: ubuntu-latest\n"
+            "  suite:\n    name: ${{ inputs.suite }}\n    runs-on: ubuntu-latest\n"
+        ),
+        "caller.yml": "on: pull_request\njobs:\n  call:\n    name: A\n    uses: ./.github/workflows/callee.yml\n",
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"A / Lint"})
+    assert len(blind_spots(sources)) == 1
+
+
+def test_a_name_the_check_cannot_work_out_is_reported_without_failing_the_check() -> None:
+    sources: Final = {
+        "a.yml": "on: pull_request\njobs:\n  unit:\n    uses: BerriAI/other/.github/workflows/base.yml@main\n",
+    }
+
+    assert blind_spots(sources) != ()
+    assert exit_code(sources) == 0
+
+
+def test_a_caller_whose_own_name_is_unreadable_publishes_none_of_its_callee_names() -> None:
+    sources: Final = {
+        ".github/workflows/callee.yml": (
+            "on:\n  workflow_call:\njobs:\n  lint:\n    name: Lint\n    runs-on: ubuntu-latest\n"
+        ),
+        "caller.yml": (
+            "on: pull_request\njobs:\n  call:\n    name: ${{ matrix.suite }}\n"
+            "    uses: ./.github/workflows/callee.yml\n"
+        ),
+        "other.yml": "on: pull_request\njobs:\n  plain:\n    name: Lint\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"Lint"})
+    assert collisions(sources) == ()
+    assert "its name stays" in blind_spots(sources)[0]
+
+
+def test_an_include_row_naming_a_listed_key_extends_only_the_combinations_it_matches() -> None:
+    sources: Final = {
+        "unit.yml": (
+            "on: pull_request\njobs:\n  unit:\n"
+            "    name: ${{ matrix.python-version }} ${{ matrix.label }}\n"
+            "    runs-on: ubuntu-latest\n    strategy:\n      matrix:\n"
+            '        python-version: ["3.12", "3.13"]\n'
+            "        include:\n"
+            '          - python-version: "3.12"\n'
+            "            label: fast\n"
+        )
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"3.12 fast"})
+    assert len(blind_spots(sources)) == 1

From 904542a559c1f297c63a5b299557d340c6f3f350 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 01:27:03 -0700
Subject: [PATCH 58/79] fix(ci): stop guessing at names built from contexts the
 sweep cannot read

Three ways the sweep could fail a workflow GitHub would publish fine.

`github.workflow` and `github.job` were counted as fixed for the whole run, so
two jobs naming themselves after the workflow they sit in were reported as a
collision. `runner` and `vars` were wrong the same way. Drop the exception
entirely: a name still holding an expression is one GitHub resolves per job, so
it is nothing to compare, which is what the rest of the module already does.

`format()` was resolved with Python's semantics, so an attribute lookup crashed
the script and a width specifier padded a name GitHub never pads. Fill `{0}`
holes and escaped braces, and treat anything richer as unresolved.

A matrix `include` or `exclude` row holding a value that is not a scalar lost
that key and became an empty row, which excludes every combination. Report the
row instead of quietly reshaping the matrix around it.
---
 .../check_workflow_job_name_collisions.py     | 52 +++++++-------
 .../test_workflow_job_name_collisions.py      | 70 +++++++++++++++++--
 2 files changed, 91 insertions(+), 31 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 1e81acc4f90..9976bd9aa41 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -24,13 +24,12 @@ real `<shard> / Run tests` names rather than staying opaque.
 
 Whatever the sweep cannot work out is left out of the comparison and reported
 instead of guessed, because a guess that lands wrong fails a workflow GitHub
-would have published perfectly well. A name left holding an expression over
-`inputs`, `needs` or `matrix` reads differently from every job that runs it, so
-it is one of those; an expression over `github` and the other contexts fixed for
-the whole run reads the same everywhere, so two jobs carrying it still clash and
-the check still says so. A matrix that is itself an expression or that lists
-values which are not scalars, and a call this sweep cannot follow, go in the same
-bucket. The cost is that a real clash hiding behind one of them goes unseen,
+would have published perfectly well. A name still holding an expression once the
+combination is filled in is one GitHub resolves per job, so it is one of those:
+guessing that two jobs sharing such a template clash would fail workflows over a
+context this sweep cannot read. A matrix that is itself an expression or that
+lists values which are not scalars, an `include` or `exclude` row shaped the same
+way, and a call this sweep cannot follow, go in the same bucket. The cost is that a real clash hiding behind one of them goes unseen,
 which leaves a merge no worse off than before this check existed, where the
 opposite direction would block work that was fine.
 
@@ -62,8 +61,7 @@ MATRIX_REF: Final = re.compile(r"^matrix\.(?P<key>[\w-]+)$")
 LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
 FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
 COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
-CONTEXT_REF: Final = re.compile(r"\b(?P<root>[a-z][\w-]*)\s*\.")
-RUN_FIXED_CONTEXTS: Final = frozenset({"github", "vars", "env", "runner", "secrets"})
+GITHUB_PLACEHOLDER: Final = re.compile(r"\{\{|\}\}|\{\d+\}")
 NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
 NO_CALLERS: Final[frozenset[str]] = frozenset()
 SCALAR: Final = (str, int, float)
@@ -149,15 +147,17 @@ def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, .
     return tuple((key, values) for key, values in listed if not isinstance(values, Opaque))
 
 
-def directive_rows(matrix: Mapping[str, object], directive: str) -> tuple[Mapping[str, str], ...]:
+def directive_rows(matrix: Mapping[str, object], directive: str) -> tuple[Mapping[str, str], ...] | Opaque:
+    """One `include` or `exclude` row, or why the combinations they shape cannot be worked out."""
     rows: Final = matrix.get(directive)
     if not isinstance(rows, Sequence) or isinstance(rows, str):
         return ()
-    return tuple(
-        MappingProxyType({str(key): scalar_text(value) for key, value in row.items() if isinstance(value, SCALAR)})
-        for row in rows
-        if isinstance(row, Mapping)
-    )
+    mappings: Final = tuple(row for row in rows if isinstance(row, Mapping))
+    if len(mappings) != len(rows):
+        return Opaque(f"a matrix `{directive}` row is not a mapping of values")
+    if any(not isinstance(value, SCALAR) for row in mappings for value in row.values()):
+        return Opaque(f"a matrix `{directive}` row holds a value that is not a plain scalar")
+    return tuple(MappingProxyType({str(key): scalar_text(value) for key, value in row.items()}) for row in mappings)
 
 
 def drops(row: Mapping[str, str], combination: Mapping[str, str]) -> bool:
@@ -195,7 +195,11 @@ def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...] | Opaque:
     if isinstance(listed, Opaque):
         return listed
     rows: Final = directive_rows(matrix, "include")
+    if isinstance(rows, Opaque):
+        return rows
     dropped: Final = directive_rows(matrix, "exclude")
+    if isinstance(dropped, Opaque):
+        return dropped
     kept: Final = tuple(
         combination for combination in crossed_values(listed) if not any(drops(row, combination) for row in dropped)
     )
@@ -225,7 +229,10 @@ def split_outside(text: str, token: str) -> tuple[str, ...]:
 
 
 def formatted(template: str, arguments: Sequence[str]) -> str | None:
-    """A `format()` whose placeholders the arguments cannot fill resolves to nothing, never a crash."""
+    """GitHub's `format()` fills `{0}`-style holes and escapes braces, so anything richer resolves to nothing."""
+    residue: Final = GITHUB_PLACEHOLDER.sub("", template)
+    if "{" in residue or "}" in residue:
+        return None
     try:
         return template.format(*arguments)
     except (IndexError, KeyError, ValueError):
@@ -280,18 +287,9 @@ def rendered(template: str, values: Mapping[str, str]) -> str:
     return EXPRESSION.sub(lambda span: resolved_span(span, values), template)
 
 
-def stable(name: str) -> bool:
-    """An expression over the run's own contexts reads the same from every job, so two of them still clash."""
-    return all(
-        reference.group("root") in RUN_FIXED_CONTEXTS
-        for span in EXPRESSION.finditer(name)
-        for reference in CONTEXT_REF.finditer(span.group("body"))
-    )
-
-
 def comparable(name: str) -> bool:
-    """A leftover expression over `inputs`, `needs` or `matrix` names a different check run per job."""
-    return EXPRESSION.search(name) is None or stable(name)
+    """A name still holding an expression is one GitHub resolves per job, so it is nothing to compare."""
+    return EXPRESSION.search(name) is None
 
 
 def settled(names: Sequence[str]) -> Names:
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 39e61ff48d1..44a5adbbb33 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -302,7 +302,7 @@ def test_a_matrix_list_supplies_values_the_same_way_include_rows_do() -> None:
     assert "`Analyze (go)` is published by 2 jobs" in found[0]
 
 
-def test_two_jobs_sharing_a_template_over_the_run_itself_still_collide() -> None:
+def test_two_jobs_sharing_a_template_over_the_run_itself_are_reported_rather_than_guessed() -> None:
     template: Final = (
         "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ github.event_name }}}}-build\n    runs-on: ubuntu-latest\n"
     )
@@ -311,10 +311,72 @@ def test_two_jobs_sharing_a_template_over_the_run_itself_still_collide() -> None
         "b.yml": template.format(job="two"),
     }
 
-    found: Final = collisions(sources)
+    assert collisions(sources) == ()
+    assert len(blind_spots(sources)) == 2
 
-    assert len(found) == 1
-    assert "is published by 2 jobs" in found[0]
+
+def test_a_name_reading_the_workflow_it_sits_in_is_not_called_a_collision() -> None:
+    template: Final = (
+        "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ github.workflow }}}} / build\n    runs-on: ubuntu-latest\n"
+    )
+    sources: Final = {
+        "a.yml": template.format(job="one"),
+        "b.yml": template.format(job="two"),
+    }
+
+    assert collisions(sources) == ()
+    assert exit_code(sources) == 0
+
+
+def test_a_format_call_python_accepts_but_github_does_not_publishes_nothing_to_compare() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  one:\n    name: ${{ format('{0.real}', matrix.shard) }}\n"
+            "    runs-on: ubuntu-latest\n    strategy:\n      matrix:\n        shard: [core]\n"
+        ),
+    }
+
+    assert collisions(sources) == ()
+    assert blind_spots(sources) != ()
+
+
+def test_a_format_call_padding_its_argument_publishes_nothing_to_compare() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  one:\n    name: ${{ format('{0:>8}', matrix.shard) }}\n"
+            "    runs-on: ubuntu-latest\n    strategy:\n      matrix:\n        shard: [core]\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  two:\n    name: '    core'\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+    assert blind_spots(sources) != ()
+
+
+def test_an_exclude_row_that_is_not_a_mapping_is_reported_rather_than_skipped() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        v: [1, 2]\n        exclude:\n          - oops\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  other:\n    name: build (1)\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+    assert blind_spots(sources) != ()
+
+
+def test_an_exclude_row_holding_a_non_scalar_never_drops_every_combination() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        v: [1, 2]\n        exclude:\n          - cfg: {k: 1}\n"
+        ),
+        "b.yml": "on: pull_request\njobs:\n  build:\n    runs-on: ubuntu-latest\n",
+    }
+
+    assert collisions(sources) == ()
+    assert blind_spots(sources) != ()
 
 
 def test_two_jobs_sharing_a_template_that_reads_per_job_are_not_called_a_collision() -> None:

From aa2c41f489eb3eba1de9983d641239b131057daa Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 01:59:03 -0700
Subject: [PATCH 59/79] fix(ci): stop the name-collision check failing legal
 workflows

An expression at `jobs.<id>.strategy` is legal on GitHub, but the model
required a mapping there, so a workflow using one made the whole file
unreadable and turned code-quality red. That job's names are now a blind
spot like any other name the sweep cannot work out offline.

A matrix whose `name:` holds no matrix value publishes that one name once
per combination, which leaves a required context just as ambiguous as two
jobs sharing a name, so it now reports instead of deduping.

A file that does not parse as one YAML document is reported the way the
module already promised, rather than escaping as a traceback.
---
 .../check_workflow_job_name_collisions.py     | 52 ++++++++++++-----
 .../test_workflow_job_name_collisions.py      | 58 +++++++++++++++++++
 2 files changed, 94 insertions(+), 16 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 9976bd9aa41..941039e8340 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -1,10 +1,12 @@
-"""Catch two workflow jobs that publish check runs under the same name.
+"""Catch workflow jobs that publish check runs under the same name.
 
 A ruleset's required status check names a check run and GitHub matches it by that
 name alone. When two jobs publish the same name the required context stops
 mapping to the job that proves it: the commit carries two check runs under one
 name and nothing says which one the ruleset required. Both being green hides the
 clash completely, so the context quietly stops meaning what the ruleset intended.
+One job lands in the same place when its `name:` holds no matrix value, since
+every combination it runs then reports under that one name.
 
 `.github/workflows/auto-close-duplicates.yml` shipped a job id `test` while
 `.github/workflows/test-mcp.yml` already published the required `test` context,
@@ -29,16 +31,17 @@ combination is filled in is one GitHub resolves per job, so it is one of those:
 guessing that two jobs sharing such a template clash would fail workflows over a
 context this sweep cannot read. A matrix that is itself an expression or that
 lists values which are not scalars, an `include` or `exclude` row shaped the same
-way, and a call this sweep cannot follow, go in the same bucket. The cost is that a real clash hiding behind one of them goes unseen,
-which leaves a merge no worse off than before this check existed, where the
-opposite direction would block work that was fine.
+way, a whole `strategy:` that comes from an expression, and a call this sweep
+cannot follow, go in the same bucket. The cost is that a real clash hiding behind
+one of them goes unseen, which leaves a merge no worse off than before this check
+existed, where the opposite direction would block work that was fine.
 
 A job calling a local reusable workflow publishes one check run per job of the
 callee, named `<caller> / <callee>` and chained through however many levels of
 local calls it takes, which is why a caller's name never collides with a plain
-job that happens to match it. A file under `.github/workflows/` that does not read as
-a workflow at all is reported rather than skipped, since skipping it silently
-would hide every job it holds.
+job that happens to match it. A file under `.github/workflows/` that does not
+read as one workflow at all is reported rather than skipped, since skipping it
+silently would hide every job it holds.
 """
 
 import itertools
@@ -90,7 +93,7 @@ class Names:
 class Job(BaseModel):
     name: object = None
     uses: str | None = None
-    strategy: Mapping[str, object] = Field(default_factory=dict)
+    strategy: object = Field(default_factory=dict)
 
 
 class Workflow(BaseModel):
@@ -104,7 +107,10 @@ def scalar_text(value: object) -> str:
 
 def parse(source: str) -> tuple[Workflow, object] | Unreadable:
     """The workflow plus its raw `on:` value, or why the file does not read as one."""
-    parsed: Final = yaml.safe_load(source)
+    try:
+        parsed: Final = yaml.safe_load(source)
+    except yaml.YAMLError:
+        return Unreadable("it does not read as one YAML document")
     if not isinstance(parsed, dict):
         return Unreadable("its top level is not a mapping of workflow keys")
     try:
@@ -186,6 +192,8 @@ def crossed_values(listed: Sequence[tuple[str, tuple[str, ...]]]) -> tuple[Mappi
 
 def matrix_combinations(job: Job) -> tuple[Mapping[str, str], ...] | Opaque:
     """One mapping per job the matrix produces, `exclude` applied before `include` as GitHub does."""
+    if not isinstance(job.strategy, Mapping):
+        return Opaque("its whole `strategy` comes from an expression")
     matrix: Final = job.strategy.get("matrix")
     if matrix is None:
         return ()
@@ -304,7 +312,7 @@ def expand(template: str, job: Job) -> Names:
     if isinstance(combinations, Opaque):
         return Names((), (combinations.reason,))
     over: Final = combinations or (NO_MATRIX,)
-    return settled(tuple(dict.fromkeys(rendered(template, values) for values in over)))
+    return settled(tuple(rendered(template, values) for values in over))
 
 
 def suffixed(job_id: str, combination: Mapping[str, str]) -> str:
@@ -408,13 +416,25 @@ def owners_by_name(sources: Mapping[str, str]) -> Iterator[tuple[str, tuple[str,
         yield name, tuple(owner for _, owner in pairs)
 
 
+def clash(name: str, owners: Sequence[str]) -> str | None:
+    """Why one name is ambiguous, whether two jobs carry it or one job repeats it over its matrix."""
+    jobs: Final = tuple(dict.fromkeys(owners))
+    if len(jobs) > 1:
+        return (
+            f"`{name}` is published by {len(jobs)} jobs: {', '.join(jobs)}. A required status check matching "
+            f"that name cannot say which job proves it; give one of them a distinct `name:` or job id."
+        )
+    if len(owners) > 1:
+        return (
+            f"`{name}` is published {len(owners)} times by {jobs[0]}, once per matrix combination. A required "
+            f"status check matching that name cannot say which run proves it; put a matrix value in its `name:`."
+        )
+    return None
+
+
 def collisions(sources: Mapping[str, str]) -> tuple[str, ...]:
-    return tuple(
-        f"`{name}` is published by {len(owners)} jobs: {', '.join(owners)}. A required status check matching "
-        f"that name cannot say which job proves it; give one of them a distinct `name:` or job id."
-        for name, owners in owners_by_name(sources)
-        if len(owners) > 1
-    )
+    found: Final = tuple(clash(name, owners) for name, owners in owners_by_name(sources))
+    return tuple(message for message in found if message is not None)
 
 
 def workflow_sources() -> Mapping[str, str]:
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index 44a5adbbb33..ab6b1d2965e 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -699,3 +699,61 @@ def test_an_include_row_naming_a_listed_key_extends_only_the_combinations_it_mat
 
     assert frozenset(name for name, _ in published(sources)) == frozenset({"3.12 fast"})
     assert len(blind_spots(sources)) == 1
+
+
+def test_a_job_whose_whole_strategy_is_an_expression_is_reported_rather_than_rejecting_the_file() -> None:
+    sources: Final = {
+        "plan.yml": (
+            "on: pull_request\njobs:\n  plan:\n    name: Plan\n    runs-on: ubuntu-latest\n"
+            "  fan:\n    strategy: ${{ fromJSON(needs.plan.outputs.strategy) }}\n    runs-on: ubuntu-latest\n"
+        )
+    }
+
+    assert unreadable(sources) == ()
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"Plan"})
+    assert "`strategy` comes from an expression" in blind_spots(sources)[0]
+    assert exit_code(sources) == 0
+
+
+def test_one_job_publishing_one_name_for_every_matrix_combination_is_a_collision() -> None:
+    sources: Final = {
+        "unit.yml": (
+            "on: pull_request\njobs:\n  build:\n    name: Run tests\n    runs-on: ubuntu-latest\n"
+            '    strategy:\n      matrix:\n        python-version: ["3.12", "3.13"]\n'
+        )
+    }
+
+    found: Final = collisions(sources)
+    assert len(found) == 1
+    assert "`Run tests` is published 2 times by unit.yml job `build`" in found[0]
+    assert exit_code(sources) == 1
+
+
+def test_a_name_carrying_a_matrix_value_publishes_one_name_per_combination_without_colliding() -> None:
+    sources: Final = {
+        "unit.yml": (
+            "on: pull_request\njobs:\n  build:\n    name: Run tests ${{ matrix.python-version }}\n"
+            '    runs-on: ubuntu-latest\n    strategy:\n      matrix:\n        python-version: ["3.12", "3.13"]\n'
+        )
+    }
+
+    assert frozenset(name for name, _ in published(sources)) == frozenset({"Run tests 3.12", "Run tests 3.13"})
+    assert collisions(sources) == ()
+    assert exit_code(sources) == 0
+
+
+def test_a_file_that_is_not_valid_yaml_is_reported_rather_than_raising() -> None:
+    sources: Final = {"broken.yml": "jobs:\n  build: [\n"}
+
+    assert unreadable(sources) == (
+        "broken.yml sits in the workflows directory but it does not read as one YAML "
+        "document, so none of its jobs were checked.",
+    )
+    assert exit_code(sources) == 1
+
+
+def test_a_file_holding_two_yaml_documents_is_reported_rather_than_raising() -> None:
+    sources: Final = {"two.yml": "on: pull_request\n---\non: push\n"}
+
+    assert len(unreadable(sources)) == 1
+    assert exit_code(sources) == 1

From 5a27e11263b2e267475ab39e287ead25920d43d9 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 02:32:50 -0700
Subject: [PATCH 60/79] fix(ci): compare names one workflow run settles the
 same way

A `name:` whose only leftover expressions read a `github.` property other
than `github.job` is filled in identically for every job of the run that
publishes it, so two jobs of one workflow carrying it land on the same
check run. Those names now compare against the other jobs of their own
file instead of sitting in the blind-spot bucket. They stay out of the
comparison across files, where two workflows can run on different events
---
 .../check_workflow_job_name_collisions.py     | 67 ++++++++++++--
 .../test_workflow_job_name_collisions.py      | 91 ++++++++++++++++++-
 2 files changed, 148 insertions(+), 10 deletions(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 941039e8340..482f8386536 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -27,12 +27,16 @@ real `<shard> / Run tests` names rather than staying opaque.
 Whatever the sweep cannot work out is left out of the comparison and reported
 instead of guessed, because a guess that lands wrong fails a workflow GitHub
 would have published perfectly well. A name still holding an expression once the
-combination is filled in is one GitHub resolves per job, so it is one of those:
-guessing that two jobs sharing such a template clash would fail workflows over a
-context this sweep cannot read. A matrix that is itself an expression or that
-lists values which are not scalars, an `include` or `exclude` row shaped the same
-way, a whole `strategy:` that comes from an expression, and a call this sweep
-cannot follow, go in the same bucket. The cost is that a real clash hiding behind
+combination is filled in is usually one GitHub resolves per job, so it is one of
+those: guessing that two jobs sharing such a template clash would fail workflows
+over a context this sweep cannot read. The exception is a name whose leftover
+expressions all read a `github.` property other than `github.job`, which one run
+fills in the same way for every job in it, so those are compared against the
+other jobs of their own workflow and stay out of the comparison across files,
+where two workflows can run on different events. A matrix that is itself an
+expression or that lists values which are not scalars, an `include` or `exclude`
+row shaped the same way, a whole `strategy:` that comes from an expression, and a
+call this sweep cannot follow, go in the same bucket. The cost is that a real clash hiding behind
 one of them goes unseen, which leaves a merge no worse off than before this check
 existed, where the opposite direction would block work that was fine.
 
@@ -64,6 +68,7 @@ MATRIX_REF: Final = re.compile(r"^matrix\.(?P<key>[\w-]+)$")
 LITERAL: Final = re.compile(r"^'(?P<text>[^']*)'$")
 FORMAT_CALL: Final = re.compile(r"^format\((?P<args>.*)\)$", re.DOTALL)
 COMPARISON: Final = re.compile(r"^(?P<left>.+?)\s*(?P<operator>==|!=)\s*(?P<right>.+)$", re.DOTALL)
+RUN_WIDE: Final = re.compile(r"^github\.(?!job\b)[\w.]+$")
 GITHUB_PLACEHOLDER: Final = re.compile(r"\{\{|\}\}|\{\d+\}")
 NO_MATRIX: Final[Mapping[str, str]] = MappingProxyType({})
 NO_CALLERS: Final[frozenset[str]] = frozenset()
@@ -88,6 +93,7 @@ class Names:
 
     known: tuple[str, ...] = ()
     unknown: tuple[str, ...] = ()
+    local: tuple[str, ...] = ()
 
 
 class Job(BaseModel):
@@ -300,10 +306,17 @@ def comparable(name: str) -> bool:
     return EXPRESSION.search(name) is None
 
 
+def run_wide(name: str) -> bool:
+    """A name whose leftover expressions one workflow run fills in the same way for every job in it."""
+    return all(RUN_WIDE.match(span.group("body").strip()) is not None for span in EXPRESSION.finditer(name))
+
+
 def settled(names: Sequence[str]) -> Names:
+    unresolved: Final = tuple(name for name in names if not comparable(name))
     return Names(
         tuple(name for name in names if comparable(name)),
-        tuple(f"its name stays `{name}`" for name in names if not comparable(name)),
+        tuple(f"its name stays `{name}`" for name in unresolved if not run_wide(name)),
+        tuple(name for name in unresolved if run_wide(name)),
     )
 
 
@@ -340,9 +353,15 @@ def joined(groups: Sequence[Names]) -> Names:
     return Names(
         tuple(name for group in groups for name in group.known),
         tuple(reason for group in groups for reason in group.unknown),
+        tuple(name for group in groups for name in group.local),
     )
 
 
+def tagged(names: Names) -> tuple[tuple[str, bool], ...]:
+    """Each name a job publishes beside whether only its own workflow's run settles it."""
+    return (*((name, False) for name in names.known), *((name, True) for name in names.local))
+
+
 def call_blocker(job: Job, workflows: Mapping[str, Workflow], callers: frozenset[str]) -> str | None:
     path: Final = callee_path(job)
     if path is None:
@@ -366,9 +385,15 @@ def job_names(job_id: str, job: Job, workflows: Mapping[str, Workflow], callers:
             for callee_id, callee_job in workflows[path].jobs.items()
         )
     )
+    composed: Final = tuple(
+        (f"{prefix} / {suffix}", prefix_local or suffix_local)
+        for prefix, prefix_local in tagged(prefixes)
+        for suffix, suffix_local in tagged(suffixes)
+    )
     return Names(
-        tuple(f"{prefix} / {suffix}" for prefix in prefixes.known for suffix in suffixes.known),
+        tuple(name for name, is_local in composed if not is_local),
         (*prefixes.unknown, *suffixes.unknown),
+        tuple(name for name, is_local in composed if is_local),
     )
 
 
@@ -432,9 +457,33 @@ def clash(name: str, owners: Sequence[str]) -> str | None:
     return None
 
 
+def local_published(sources: Mapping[str, str]) -> Iterator[tuple[tuple[str, str], str]]:
+    """Names their own workflow's run settles, keyed by the file whose run settles them."""
+    for rel, job_id, names in scanned_jobs(sources):
+        for name in names.local:
+            yield (rel, name), f"job `{job_id}`"
+
+
+def local_clash(rel: str, name: str, owners: Sequence[str]) -> str | None:
+    """Why one workflow's own run lands several of its jobs on one check run."""
+    if len(owners) < 2:
+        return None
+    jobs: Final = tuple(dict.fromkeys(owners))
+    return (
+        f"`{name}` is published {len(owners)} times inside {rel}, by {', '.join(jobs)}. One run fills that "
+        f"expression in the same way throughout, so they all land on one check run; make the names differ."
+    )
+
+
+def local_clashes(sources: Mapping[str, str]) -> tuple[str, ...]:
+    grouped: Final = itertools.groupby(sorted(local_published(sources)), key=operator.itemgetter(0))
+    found: Final = tuple(local_clash(rel, name, tuple(owner for _, owner in pairs)) for (rel, name), pairs in grouped)
+    return tuple(message for message in found if message is not None)
+
+
 def collisions(sources: Mapping[str, str]) -> tuple[str, ...]:
     found: Final = tuple(clash(name, owners) for name, owners in owners_by_name(sources))
-    return tuple(message for message in found if message is not None)
+    return (*(message for message in found if message is not None), *local_clashes(sources))
 
 
 def workflow_sources() -> Mapping[str, str]:
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index ab6b1d2965e..aeb10985e1b 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -302,7 +302,7 @@ def test_a_matrix_list_supplies_values_the_same_way_include_rows_do() -> None:
     assert "`Analyze (go)` is published by 2 jobs" in found[0]
 
 
-def test_two_jobs_sharing_a_template_over_the_run_itself_are_reported_rather_than_guessed() -> None:
+def test_two_workflows_sharing_a_run_wide_template_are_not_called_a_collision() -> None:
     template: Final = (
         "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ github.event_name }}}}-build\n    runs-on: ubuntu-latest\n"
     )
@@ -311,10 +311,99 @@ def test_two_jobs_sharing_a_template_over_the_run_itself_are_reported_rather_tha
         "b.yml": template.format(job="two"),
     }
 
+    assert collisions(sources) == ()
+    assert blind_spots(sources) == ()
+
+
+def test_two_jobs_of_one_workflow_sharing_a_run_wide_template_are_a_collision() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n"
+            "  one:\n    name: ${{ github.event_name }}-build\n    runs-on: ubuntu-latest\n"
+            "  two:\n    name: ${{ github.event_name }}-build\n    runs-on: ubuntu-latest\n"
+        ),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "is published 2 times inside a.yml, by job `one`, job `two`" in found[0]
+    assert exit_code(sources) == 1
+
+
+def test_a_run_wide_template_carrying_a_matrix_value_does_not_collide_inside_one_workflow() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n"
+            "  one:\n    name: ${{ github.event_name }}-${{ matrix.shard }}\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        shard: [core, extras]\n"
+        ),
+    }
+
+    assert collisions(sources) == ()
+    assert blind_spots(sources) == ()
+
+
+def test_a_run_wide_name_repeated_over_a_matrix_by_one_job_is_a_collision() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n"
+            "  one:\n    name: ${{ github.event_name }}-build\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        shard: [core, extras]\n"
+        ),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "is published 2 times inside a.yml, by job `one`" in found[0]
+
+
+def test_a_name_reading_the_job_it_sits_in_stays_out_of_the_comparison() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n"
+            "  one:\n    name: ${{ github.job }}-build\n    runs-on: ubuntu-latest\n"
+            "  two:\n    name: ${{ github.job }}-build\n    runs-on: ubuntu-latest\n"
+        ),
+    }
+
     assert collisions(sources) == ()
     assert len(blind_spots(sources)) == 2
 
 
+def test_a_run_wide_caller_name_collides_through_the_workflow_it_calls() -> None:
+    sources: Final = {
+        ".github/workflows/a.yml": (
+            "on: pull_request\njobs:\n"
+            "  one:\n    name: ${{ github.event_name }}\n    uses: ./.github/workflows/c.yml\n"
+            "  two:\n    name: ${{ github.event_name }}\n    uses: ./.github/workflows/c.yml\n"
+        ),
+        ".github/workflows/c.yml": "on:\n  workflow_call:\njobs:\n  build:\n    runs-on: ubuntu-latest\n",
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "github.event_name }} / build` is published 2 times inside .github/workflows/a.yml" in found[0]
+
+
+def test_a_run_wide_name_inside_a_called_workflow_collides_under_the_caller() -> None:
+    sources: Final = {
+        ".github/workflows/a.yml": ("on: pull_request\njobs:\n  one:\n    uses: ./.github/workflows/c.yml\n"),
+        ".github/workflows/c.yml": (
+            "on:\n  workflow_call:\njobs:\n"
+            "  build:\n    name: ${{ github.event_name }}\n    runs-on: ubuntu-latest\n"
+            "  lint:\n    name: ${{ github.event_name }}\n    runs-on: ubuntu-latest\n"
+        ),
+    }
+
+    found: Final = collisions(sources)
+
+    assert len(found) == 1
+    assert "`one / ${{ github.event_name }}` is published 2 times inside .github/workflows/a.yml" in found[0]
+
+
 def test_a_name_reading_the_workflow_it_sits_in_is_not_called_a_collision() -> None:
     template: Final = (
         "on: pull_request\njobs:\n  {job}:\n    name: ${{{{ github.workflow }}}} / build\n    runs-on: ubuntu-latest\n"

From 4edf6f4dc530f88afe09d2bb986a4bdc5a962cc0 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 02:59:16 -0700
Subject: [PATCH 61/79] fix(ci): keep an expression matrix directive out of the
 comparison

`include:` or `exclude:` written as `${{ ... }}` read back as a string, and
the sweep treated that as the directive being absent, so it expanded every
combination GitHub would have dropped. A job whose `name:` holds no matrix
value then looked like it repeated one name across combinations that never
run. An absent directive still means no rows; anything that is not a list
of rows now joins the names left out of the comparison
---
 .../check_workflow_job_name_collisions.py     |  4 ++-
 .../test_workflow_job_name_collisions.py      | 32 +++++++++++++++++++
 2 files changed, 35 insertions(+), 1 deletion(-)

diff --git a/tests/code_coverage_tests/check_workflow_job_name_collisions.py b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
index 482f8386536..ae2c1d80c8f 100644
--- a/tests/code_coverage_tests/check_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/check_workflow_job_name_collisions.py
@@ -162,8 +162,10 @@ def listed_values(matrix: Mapping[str, object]) -> tuple[tuple[str, tuple[str, .
 def directive_rows(matrix: Mapping[str, object], directive: str) -> tuple[Mapping[str, str], ...] | Opaque:
     """One `include` or `exclude` row, or why the combinations they shape cannot be worked out."""
     rows: Final = matrix.get(directive)
-    if not isinstance(rows, Sequence) or isinstance(rows, str):
+    if rows is None:
         return ()
+    if not isinstance(rows, Sequence) or isinstance(rows, str):
+        return Opaque(f"a matrix `{directive}` is itself an expression rather than a list of rows")
     mappings: Final = tuple(row for row in rows if isinstance(row, Mapping))
     if len(mappings) != len(rows):
         return Opaque(f"a matrix `{directive}` row is not a mapping of values")
diff --git a/tests/code_coverage_tests/test_workflow_job_name_collisions.py b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
index aeb10985e1b..0f5ba43bd7a 100644
--- a/tests/code_coverage_tests/test_workflow_job_name_collisions.py
+++ b/tests/code_coverage_tests/test_workflow_job_name_collisions.py
@@ -846,3 +846,35 @@ def test_a_file_holding_two_yaml_documents_is_reported_rather_than_raising() ->
 
     assert len(unreadable(sources)) == 1
     assert exit_code(sources) == 1
+
+
+def test_an_exclude_that_is_itself_an_expression_is_reported_rather_than_ignored() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  one:\n    name: build\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        python: ['3.11', '3.12']\n"
+            "        exclude: ${{ fromJson(vars.SKIP) }}\n"
+        ),
+    }
+
+    found: Final = blind_spots(sources)
+
+    assert collisions(sources) == ()
+    assert len(found) == 1
+    assert "a matrix `exclude` is itself an expression" in found[0]
+
+
+def test_an_include_that_is_itself_an_expression_is_reported_rather_than_ignored() -> None:
+    sources: Final = {
+        "a.yml": (
+            "on: pull_request\njobs:\n  one:\n    name: build-${{ matrix.python }}\n    runs-on: ubuntu-latest\n"
+            "    strategy:\n      matrix:\n        python: ['3.11']\n"
+            "        include: ${{ fromJson(vars.EXTRA) }}\n"
+        ),
+    }
+
+    found: Final = blind_spots(sources)
+
+    assert collisions(sources) == ()
+    assert len(found) == 1
+    assert "a matrix `include` is itself an expression" in found[0]

From 168a0055a244acdcf97c330c52e085ab40b1424c Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sun, 6 Sep 2026 12:44:19 -0700
Subject: [PATCH 62/79] chore(lint): stop ratcheting *-budget.json on PR
 branches (#39937)

---
 CLAUDE.md                                    |  2 +-
 scripts/test_quality_gate.py                 | 59 +++-----------------
 tests/test_litellm/test_test_quality_gate.py | 33 ++---------
 3 files changed, 13 insertions(+), 81 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index d9e9e8f1586..a7b9b6b9bc0 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -52,7 +52,7 @@ Don't hesitate to use values in .env to get needed API keys and other secrets, a
 
 Python max line length is 120, not 88
 
-When you fix violations gated by `ruff-strict-budget.json`, `type-discipline-budget.json`, or `basedpyright-code-budget.json`, run `make lint-budget-update` and commit the lowered limits so the ceilings ratchet down instead of leaving stale headroom. It measures the working tree, so it must contain exactly the fixes you're committing
+Never edit or commit `ruff-strict-budget.json`, `type-discipline-budget.json`, `basedpyright-code-budget.json`, or `test-quality-budget.json` on a PR branch, and don't run `make lint-budget-update` there. A scheduled Devin automation lowers the limits on `litellm_internal_staging` in its own PR by exactly what landed since the last ratchet, so concurrent PRs don't fight over the same `"limit"` lines. If your branch already carries a budget edit, drop it before opening the PR
 
 `make check` (f.k.a. `make pre-commit`, which still works identically as an alias) saves its complete output to a log file in .git (overwriting previous logs) and prints that path as its first and last output lines. To inspect a run, read or grep that log instead of re-running the multi-minute checks just to see a different slice
 
diff --git a/scripts/test_quality_gate.py b/scripts/test_quality_gate.py
index 7d34b194f1c..9e22eec29ab 100644
--- a/scripts/test_quality_gate.py
+++ b/scripts/test_quality_gate.py
@@ -10,17 +10,12 @@ base.
 
 Every rule is seeded at exactly its count on the day the gate landed, so the
 suite's existing debt is grandfathered and any net-new violation trips the gate
-immediately. ``--update`` ratchets a limit down by the violations this branch
-fixed relative to its branch point (the merge-base), so the ceilings only ever
-fall. Base counts are measured with the *current* checker, so a rule introduced
-on this branch is counted at the base too and ratchets like every other one.
-
-Only ever falling is not the same as always falling, so the gate enforces the
-second half: a branch that clears violations and leaves the ceiling above its
-new count fails, naming the rules and telling the author to run
-``make lint-budget-update``. Without that, a removed violation could come back
-later under a ceiling nobody lowered. Drift already in the base is never
-blamed, so this fires only on the branch that did the clearing.
+immediately. ``--update`` ratchets a limit down by the violations fixed relative
+to ``--base``, so the ceilings only ever fall. Base counts are measured with the
+*current* checker, so a rule introduced on this branch is counted at the base too
+and ratchets like every other one. The ratchet runs as a scheduled automation
+against litellm_internal_staging, not on PR branches, so concurrent PRs never
+race to edit the same limit.
 
 The deliberate difference from its sibling: this gate has no headroom anywhere.
 Type discipline seeded LIT010/LIT011 at 1.5x to leave room for an in-flight
@@ -144,21 +139,6 @@ def over_ceiling(head: Mapping[str, int], budget: Mapping[str, Mapping[str, int]
     )
 
 
-def unratcheted(
-    head: Mapping[str, int],
-    base: Mapping[str, int],
-    budget: Mapping[str, Mapping[str, int]],
-) -> tuple[Breach, ...]:
-    """Rules this branch cleared without lowering the ceiling behind them. Requires
-    both `head < base`, so drift already in the base is never blamed on this change,
-    and `head < limit`, so a ceiling already at the count is left alone."""
-    return tuple(sorted(
-        Breach(rule, head.get(rule, 0), spec["limit"], head.get(rule, 0) - base.get(rule, 0))
-        for rule, spec in budget.items()
-        if head.get(rule, 0) < base.get(rule, 0) and head.get(rule, 0) < spec["limit"]
-    ))
-
-
 def evaluate(
     head: Mapping[str, int],
     base: Mapping[str, int],
@@ -198,38 +178,15 @@ def introduced(
     return tuple(v for v in violations if v.line in changed.get(v.file, frozenset()))
 
 
-def touches_measured_tree(base_point: str) -> bool:
-    """Whether this branch changed anything that can move a count. A branch that
-    touches neither the test tree nor the checker cannot have cleared a violation,
-    so the base scan is skipped and the gate stays cheap on the common change."""
-    changed: Final = _run(
-        ["git", "diff", "--name-only", base_point, "--", TARGET, str(CHECKER.relative_to(REPO_ROOT))]
-    )
-    return bool(changed.strip())
-
-
 def cmd_check(base: str) -> None:
     budget: Final = json.loads(BUDGET_PATH.read_text())
     head: Final = head_violations()
     head_counts: Final = count_by_rule(head)
-    base_point: Final = resolve_base_point(base)
-    if not over_ceiling(head_counts, budget) and not touches_measured_tree(base_point):
+    if not over_ceiling(head_counts, budget):
         print(f"OK: every TQ rule is within its test-suite ceiling (base {base})")
         return
+    base_point: Final = resolve_base_point(base)
     base_at_point: Final = base_counts(base_point)
-    stale: Final = unratcheted(head_counts, base_at_point, budget)
-    if stale:
-        print(f"FAIL: TQ-rule limits were left above the count this branch reached (base {base}):")
-        for breach in stale:
-            print(
-                f"  {breach.rule}: this branch cleared {-breach.added} down to {breach.total}, "
-                f"but the limit is still {breach.cap}"
-            )
-        print(
-            "Run `make lint-budget-update` and commit the lowered limits, so the "
-            "violations you cleared cannot come back under a ceiling nobody moved."
-        )
-        raise SystemExit(1)
     breaches: Final = evaluate(head_counts, base_at_point, budget)
     if not breaches:
         print(f"OK: every TQ rule is within its test-suite ceiling (base {base})")
diff --git a/tests/test_litellm/test_test_quality_gate.py b/tests/test_litellm/test_test_quality_gate.py
index 8cce6bc735a..a8b38ecdf49 100644
--- a/tests/test_litellm/test_test_quality_gate.py
+++ b/tests/test_litellm/test_test_quality_gate.py
@@ -1,11 +1,9 @@
 """Tests for scripts/test_quality_gate.py.
 
-The gate's whole value is that it blames a change only for what it adds, that a limit
-can never rise, and that a limit cannot stay above a count the branch pushed below it.
-All three live in pure functions, so they are tested directly: `evaluate` for the blame
-rule, `ratcheted_budget` for the one-way ratchet, `unratcheted` for the ceiling a branch
-left behind, and `parse_changed_lines` for the diff scan that turns a breach into
-file:line.
+The gate's whole value is that it blames a change only for what it adds and that a
+limit can never rise. Both live in pure functions, so they are tested directly:
+`evaluate` for the blame rule, `ratcheted_budget` for the one-way ratchet, and
+`parse_changed_lines` for the diff scan that turns a breach into file:line.
 """
 
 import importlib.util
@@ -73,29 +71,6 @@ def test_ratchet_lowers_a_rule_introduced_on_this_branch_like_any_other():
     assert updated["TQ001"]["limit"] == 4
 
 
-def test_a_branch_that_cleared_violations_must_lower_the_ceiling():
-    stale = gate.unratcheted({"TQ001": 6}, {"TQ001": 10}, _BUDGET)
-    assert [(b.rule, b.total, b.cap, b.added) for b in stale] == [("TQ001", 6, 10, -4)]
-
-
-def test_headroom_already_in_the_base_is_not_blamed_on_this_branch():
-    assert gate.unratcheted({"TQ001": 6}, {"TQ001": 6}, _BUDGET) == ()
-
-
-def test_a_branch_that_cleared_down_to_the_ceiling_exactly_is_clean():
-    assert gate.unratcheted({"TQ001": 10}, {"TQ001": 12}, _BUDGET) == ()
-
-
-def test_a_branch_that_added_violations_is_not_a_ratchet_finding():
-    assert gate.unratcheted({"TQ001": 14}, {"TQ001": 10}, _BUDGET) == ()
-
-
-def test_the_ratchet_finding_survives_the_update_that_answers_it():
-    cleared = {"TQ001": 6}
-    updated = gate.ratcheted_budget(_BUDGET, cleared, {"TQ001": 10})
-    assert gate.unratcheted(cleared, {"TQ001": 10}, updated) == ()
-
-
 def test_parse_changed_lines_groups_hunks_under_their_own_file():
     diff = (
         "diff --git a/tests/a.py b/tests/a.py\n"

From 728d0953af94c29959232032c0646bf13f87cfaf Mon Sep 17 00:00:00 2001
From: yujonglee <yujong@berri.ai>
Date: Mon, 7 Sep 2026 09:00:58 -0700
Subject: [PATCH 63/79] ci: simplify Rust checks and remove wheel PR comments
 (#39975)

* ci: limit Rust workflows to Rust directory changes

* ci: run Rust checks when their workflow changes

* ci: report Rust wheels only for successful Rust changes

* ci: keep Rust wheel reports in the workflow summary

* ci: group Rust lint and validation jobs

* ci: keep Rust job names distinct from required lint and test checks

* ci: drop the unused Python setup from the Rust lint job

---------

Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
---
 .../workflows/report-rust-release-wheel.yml   | 129 ------------------
 .github/workflows/test-rust.yml               | 111 +++++++--------
 2 files changed, 50 insertions(+), 190 deletions(-)
 delete mode 100644 .github/workflows/report-rust-release-wheel.yml

diff --git a/.github/workflows/report-rust-release-wheel.yml b/.github/workflows/report-rust-release-wheel.yml
deleted file mode 100644
index 74e6be69604..00000000000
--- a/.github/workflows/report-rust-release-wheel.yml
+++ /dev/null
@@ -1,129 +0,0 @@
-name: Report LiteLLM Rust release wheel
-
-on: # zizmor: ignore[dangerous-triggers] reporter executes no PR code and consumes no PR artifacts or outputs
-  workflow_run:
-    workflows:
-      - LiteLLM Rust
-    types:
-      - completed
-
-permissions: {}
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.workflow_run.pull_requests[0].number || github.event.workflow_run.id }}
-  cancel-in-progress: false
-
-jobs:
-  report-release-wheel:
-    name: report release wheel
-    if: >-
-      github.event.workflow_run.event == 'pull_request' &&
-      github.event.workflow_run.path == '.github/workflows/test-rust.yml' &&
-      github.event.workflow_run.head_repository.full_name == github.repository &&
-      github.event.workflow_run.pull_requests[0].number != null
-    runs-on: ubuntu-latest
-    timeout-minutes: 5
-    permissions:
-      pull-requests: write
-
-    steps:
-      - name: Link release wheel report on PR
-        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7.1.0
-        env:
-          COMMENT_MARKER: "<!-- litellm-release-wheel-size -->"
-        with:
-          script: |
-            const marker = process.env.COMMENT_MARKER;
-            const workflowRun = context.payload.workflow_run;
-            const allowedConclusions = new Set([
-              "action_required",
-              "cancelled",
-              "failure",
-              "neutral",
-              "skipped",
-              "stale",
-              "startup_failure",
-              "success",
-              "timed_out",
-            ]);
-            if (
-              !allowedConclusions.has(workflowRun.conclusion) ||
-              workflowRun.event !== "pull_request" ||
-              workflowRun.path !== ".github/workflows/test-rust.yml" ||
-              workflowRun.head_repository?.full_name !==
-                `${context.repo.owner}/${context.repo.repo}` ||
-              workflowRun.pull_requests?.length !== 1
-            ) {
-              throw new Error("unexpected source workflow");
-            }
-            const pullRequest = workflowRun.pull_requests[0];
-            const pullRequestNumber = pullRequest.number;
-            const headSha = workflowRun.head_sha;
-            const runId = workflowRun.id;
-            if (
-              !Number.isSafeInteger(pullRequestNumber) ||
-              pullRequestNumber <= 0 ||
-              !Number.isSafeInteger(runId) ||
-              runId <= 0 ||
-              !/^[0-9a-f]{40}$/.test(headSha) ||
-              pullRequest.head?.sha !== headSha
-            ) {
-              throw new Error("invalid source workflow metadata");
-            }
-            const runUrl =
-              `${context.serverUrl}/${context.repo.owner}/${context.repo.repo}` +
-              `/actions/runs/${runId}`;
-            const result =
-              workflowRun.conclusion === "success"
-                ? "successfully"
-                : `with \`${workflowRun.conclusion}\``;
-            const body = [
-              marker,
-              "## LiteLLM Rust workflow",
-              "",
-              `Workflow completed ${result} for \`${headSha}\``,
-              "",
-              `[View workflow run](${runUrl})`,
-            ].join("\n");
-            const comments = await github.paginate(github.rest.issues.listComments, {
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: pullRequestNumber,
-              per_page: 100,
-            });
-            const existing = comments.find(
-              (comment) =>
-                comment.user?.login === "github-actions[bot]" &&
-                comment.body?.startsWith(marker),
-            );
-            const currentPullRequest = (
-              await github.rest.pulls.get({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                pull_number: pullRequestNumber,
-              })
-            ).data;
-            if (
-              currentPullRequest.state !== "open" ||
-              currentPullRequest.head.repo?.full_name !==
-                `${context.repo.owner}/${context.repo.repo}` ||
-              currentPullRequest.head.sha !== headSha
-            ) {
-              core.info("source workflow no longer matches the current pull request head");
-              return;
-            }
-            if (existing) {
-              await github.rest.issues.updateComment({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                comment_id: existing.id,
-                body,
-              });
-            } else {
-              await github.rest.issues.createComment({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                issue_number: pullRequestNumber,
-                body,
-              });
-            }
diff --git a/.github/workflows/test-rust.yml b/.github/workflows/test-rust.yml
index 9b8b132df62..4f56e78ddee 100644
--- a/.github/workflows/test-rust.yml
+++ b/.github/workflows/test-rust.yml
@@ -7,6 +7,7 @@ on:
       - ".cargo/**"
       - "pyproject.toml"
       - "rust-toolchain.toml"
+      - ".github/actions/setup-uv-with-retries/**"
       - ".github/scripts/smoke_test_native_wheel.py"
       - ".github/scripts/verify_linux_native_wheel.py"
       - "tests/test_litellm/rust_bridge/native_route_wheel_test.py"
@@ -22,6 +23,7 @@ on:
       - ".cargo/**"
       - "pyproject.toml"
       - "rust-toolchain.toml"
+      - ".github/actions/setup-uv-with-retries/**"
       - ".github/scripts/smoke_test_native_wheel.py"
       - ".github/scripts/verify_linux_native_wheel.py"
       - "tests/test_litellm/rust_bridge/native_route_wheel_test.py"
@@ -34,102 +36,89 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
   cancel-in-progress: true
 
+env:
+  CARGO_TERM_COLOR: always
+
 jobs:
-  rust-checks:
-    name: rustfmt, clippy, test
+  rust-lint:
     runs-on: ubuntu-latest
     timeout-minutes: 10
     defaults:
       run:
         working-directory: litellm-rust
-    env:
-      CARGO_TERM_COLOR: always
-
     steps:
-      - name: Checkout repository
-        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
         with:
           persist-credentials: false
 
-      - name: Set up Rust
-        run: rustup toolchain install
+      - run: rustup toolchain install --no-self-update
 
-      - name: Cache Cargo registry and target
-        uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
+      - run: cargo fmt --check
+
+      - uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
         with:
           path: |
             ~/.cargo/registry
             ~/.cargo/git
             litellm-rust/target
-          key: ${{ runner.os }}-cargo-${{ hashFiles('rust-toolchain.toml', 'litellm-rust/Cargo.lock') }}
+          key: ${{ runner.os }}-cargo-${{ github.job }}-${{ hashFiles('rust-toolchain.toml', '.cargo/**', 'litellm-rust/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-
+            ${{ runner.os }}-cargo-${{ github.job }}-
 
-      - name: Check Rust formatting
-        run: cargo fmt --check
+      - run: cargo clippy --workspace --all-targets --locked -- -D warnings
 
-      - name: Run Clippy
-        run: cargo clippy --workspace --all-targets --locked -- -D warnings
+      - run: cargo clippy -p litellm-core --all-targets --features bedrock-auth --locked -- -D warnings
 
-      - name: Run Clippy with Bedrock auth
-        run: cargo clippy -p litellm-core --all-targets --features bedrock-auth --locked -- -D warnings
+      - run: cargo clippy -p litellm-ai-gateway --all-targets --all-features --locked -- -D warnings
 
-      - name: Run Clippy with all gateway features
-        run: cargo clippy -p litellm-ai-gateway --all-targets --all-features --locked -- -D warnings
-
-      - name: Run Rust tests
-        run: cargo test --workspace --locked
-
-      - name: Run core tests with Bedrock auth
-        run: cargo test -p litellm-core --features bedrock-auth --locked
-
-      # Not --all-features: python-config links libpython, which this job does not install.
-      - name: Run gateway tests with the server feature
-        run: cargo test -p litellm-ai-gateway --features server --locked
-
-  release-wheel:
-    name: release wheel
+  rust-test:
     runs-on: ubuntu-latest
-    timeout-minutes: 20
-    permissions:
-      contents: read
-    env:
-      CARGO_TERM_COLOR: always
+    timeout-minutes: 30
 
     steps:
-      - name: Checkout repository
-        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
         with:
           persist-credentials: false
 
-      - name: Set up Python
-        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
         with:
           python-version: "3.12"
 
-      - name: Set up uv
-        uses: ./.github/actions/setup-uv-with-retries
+      - uses: ./.github/actions/setup-uv-with-retries
         with:
           version: "0.10.9"
 
-      - name: Set up Rust
-        run: rustup toolchain install
+      - run: rustup toolchain install --no-self-update
 
-      - name: Build release wheel
-        run: uv build --wheel --out-dir dist
+      - uses: actions/cache@0057852bfaa89a56745cba8c7296529d2fc39830 # v4.3.0
+        with:
+          path: |
+            ~/.cargo/registry
+            ~/.cargo/git
+            litellm-rust/target
+          key: ${{ runner.os }}-cargo-${{ github.job }}-${{ hashFiles('rust-toolchain.toml', '.cargo/**', 'litellm-rust/Cargo.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-cargo-${{ github.job }}-
 
-      - name: Build panic contract wheel
-        run: >-
+      - run: cargo test --workspace --locked
+        working-directory: litellm-rust
+
+      - run: cargo test -p litellm-core --features bedrock-auth --locked
+        working-directory: litellm-rust
+
+      - run: cargo test -p litellm-ai-gateway --features server --locked
+        working-directory: litellm-rust
+
+      - run: uv build --wheel --out-dir dist
+
+      - run: python .github/scripts/verify_linux_native_wheel.py dist/*.whl
+        env:
+          RELEASE_WHEEL_COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+
+      - run: python tests/test_litellm/rust_bridge/native_route_wheel_test.py dist/*.whl
+
+      - run: >-
           uv build --wheel --out-dir panic-dist
           --config-setting "maturin.build-args=--features panic-test,extension-module"
 
-      - name: Smoke-test native panic unwinding
-        run: python .github/scripts/smoke_test_native_wheel.py panic-dist/*.whl
-
-      - name: Verify stripped native extension
-        env:
-          RELEASE_WHEEL_COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
-        run: python .github/scripts/verify_linux_native_wheel.py dist/*.whl
-
-      - name: Test native route wheel
-        run: python tests/test_litellm/rust_bridge/native_route_wheel_test.py dist/*.whl
+      - run: python .github/scripts/smoke_test_native_wheel.py panic-dist/*.whl

From a1e7293fa921f6894aa0bc2fffb2db3f379342f0 Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Mon, 7 Sep 2026 12:45:16 -0400
Subject: [PATCH 64/79] fix(mcp): apply key and team guardrails to MCP tool
 calls (#39629)

* fix(mcp): apply key and team guardrails to MCP tool calls

Guardrails attached to a virtual key or team were only enforced on LLM
routes. The synthetic request built for MCP tool call guardrail hooks
carried no guardrails in its metadata, so a guardrail with default_on
false never ran on tools/call even when the key explicitly listed it.
Resolve key, team, and project guardrails onto the synthetic request
with the same helper the chat path uses.

* fix(mcp): pass project metadata through without a mutable default

* fix(mcp): mark the request dict parameter mutable-ok with a reason

* test(mcp): explain the premium_user patch and tighten the helper docstring
---
 litellm/proxy/litellm_pre_call_utils.py       | 40 +++++++++------
 litellm/proxy/utils.py                        | 10 +++-
 .../mcp_server/test_mcp_server_manager.py     | 49 +++++++++++++++++++
 tests/test_litellm/proxy/test_proxy_utils.py  | 28 +++++++++++
 4 files changed, 109 insertions(+), 18 deletions(-)

diff --git a/litellm/proxy/litellm_pre_call_utils.py b/litellm/proxy/litellm_pre_call_utils.py
index d026c5510e6..56512570448 100644
--- a/litellm/proxy/litellm_pre_call_utils.py
+++ b/litellm/proxy/litellm_pre_call_utils.py
@@ -2882,6 +2882,28 @@ def _add_guardrails_from_policies_in_metadata(
     )
 
 
+def add_guardrails_from_auth_metadata(
+    user_api_key_dict: UserAPIKeyAuth,
+    data: dict,  # mutable-ok: writes guardrails into the live request dict, same contract as the helpers it wraps
+    metadata_variable_name: str,
+) -> None:
+    """Resolve key, team, and project guardrails, direct and via policies, onto the request metadata."""
+    _add_guardrails_from_key_or_team_metadata(
+        key_metadata=user_api_key_dict.metadata,
+        team_metadata=user_api_key_dict.team_metadata,
+        project_metadata=user_api_key_dict.project_metadata,
+        data=data,
+        metadata_variable_name=metadata_variable_name,
+    )
+    _add_guardrails_from_policies_in_metadata(
+        key_metadata=user_api_key_dict.metadata,
+        team_metadata=user_api_key_dict.team_metadata,
+        project_metadata=user_api_key_dict.project_metadata,
+        data=data,
+        metadata_variable_name=metadata_variable_name,
+    )
+
+
 async def move_guardrails_to_metadata(
     data: dict,
     _metadata_variable_name: str,
@@ -2914,22 +2936,8 @@ async def move_guardrails_to_metadata(
             data.pop("policies", None)
             return
 
-    # Check key/team/project-level guardrails
-    _add_guardrails_from_key_or_team_metadata(
-        key_metadata=user_api_key_dict.metadata,
-        team_metadata=user_api_key_dict.team_metadata,
-        project_metadata=project_metadata,
-        data=data,
-        metadata_variable_name=_metadata_variable_name,
-    )
-
-    #########################################################################################
-    # Add guardrails from policies attached to key/team/project metadata
-    #########################################################################################
-    _add_guardrails_from_policies_in_metadata(
-        key_metadata=user_api_key_dict.metadata,
-        team_metadata=user_api_key_dict.team_metadata,
-        project_metadata=project_metadata,
+    add_guardrails_from_auth_metadata(
+        user_api_key_dict=user_api_key_dict,
         data=data,
         metadata_variable_name=_metadata_variable_name,
     )
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index b4e4dfeae67..f1103ce6a29 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -152,7 +152,7 @@ from litellm.proxy.hooks.parallel_request_limiter_v3 import (
 from litellm.proxy.hooks.sensitive_data_routing import (
     _PROXY_SensitiveDataRoutingHandler,
 )
-from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
+from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup, add_guardrails_from_auth_metadata
 from litellm.proxy.management_helpers.key_settings_audit import with_settings_updated_at
 from litellm.proxy.policy_engine.pipeline_executor import PipelineExecutor
 from litellm.repositories.budget_repository import BudgetRepository
@@ -924,7 +924,13 @@ class ProxyLogging:
             "incoming_bearer_token": kwargs.get("incoming_bearer_token"),
             "metadata": {"headers": kwargs.get("headers") or {}},
         }
-
+        user_api_key_auth: Final = kwargs.get("user_api_key_auth")
+        if isinstance(user_api_key_auth, UserAPIKeyAuth):
+            add_guardrails_from_auth_metadata(
+                user_api_key_dict=user_api_key_auth,
+                data=synthetic_data,
+                metadata_variable_name="metadata",
+            )
         return synthetic_data
 
     def _convert_llm_result_to_mcp_response(self, llm_result, request_obj) -> MCPPreCallResponseObject | None:
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py
index 764e2bb0e99..d19363d3b5f 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server_manager.py
@@ -58,6 +58,11 @@ from litellm.proxy._types import (
 from litellm.types.llms.custom_http import httpxSpecialProvider
 from litellm.types.mcp import MCPAuth, MCPAuthType
 from litellm.types.mcp_server.mcp_server_manager import MCPOAuthMetadata, MCPServer
+from litellm.caching.caching import DualCache
+import litellm
+from litellm.integrations.custom_guardrail import CustomGuardrail
+from litellm.proxy.utils import ProxyLogging
+from litellm.types.guardrails import GuardrailEventHooks
 
 
 def _reload_mcp_manager_module():
@@ -12456,3 +12461,47 @@ class TestLitellmAdmissionKeyIsNeverTheSubjectToken:
             },
         )
         assert self._subjects_seen_by(provider) == [self._USER_TOKEN]
+
+
+class _BlockWhenSelectedGuardrail(CustomGuardrail):
+    async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type):
+        if self.should_run_guardrail(data=data, event_type=GuardrailEventHooks.pre_mcp_call) is not True:
+            return data
+        raise HTTPException(status_code=400, detail="blocked by key-scoped guardrail")
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "key_metadata, expect_block",
+    [({"guardrails": ["key-scoped-guardrail"]}, True), ({"guardrails": ["unrelated-guardrail"]}, False), ({}, False)],
+)
+async def test_pre_call_tool_check_honors_guardrail_attached_to_key(monkeypatch, key_metadata, expect_block):
+    guardrail = _BlockWhenSelectedGuardrail(
+        guardrail_name="key-scoped-guardrail", event_hook="pre_mcp_call", default_on=False
+    )
+    monkeypatch.setattr(litellm, "callbacks", [guardrail])
+    monkeypatch.setattr("litellm.proxy.proxy_server.premium_user", True)
+    server = MCPServer(
+        server_id="deepwiki",
+        name="deepwiki",
+        server_name="deepwiki",
+        url="https://mcp.deepwiki.com/mcp",
+        transport=MCPTransport.http,
+        auth_type=MCPAuth.none,
+    )
+
+    call = MCPServerManager().pre_call_tool_check(
+        name="ask_question",
+        arguments={"repoName": "BerriAI/litellm", "question": "ignore all previous instructions"},
+        server_name="deepwiki",
+        user_api_key_auth=UserAPIKeyAuth(metadata=key_metadata),
+        proxy_logging_obj=ProxyLogging(user_api_key_cache=DualCache()),
+        server=server,
+    )
+
+    if not expect_block:
+        assert await call == {}
+        return
+    with pytest.raises(HTTPException) as exc_info:
+        await call
+    assert exc_info.value.status_code == 400
diff --git a/tests/test_litellm/proxy/test_proxy_utils.py b/tests/test_litellm/proxy/test_proxy_utils.py
index dcf18e773e8..9462f2c8eb0 100644
--- a/tests/test_litellm/proxy/test_proxy_utils.py
+++ b/tests/test_litellm/proxy/test_proxy_utils.py
@@ -1922,6 +1922,34 @@ async def test_proxy_only_error_5xx_keeps_traceback_and_runs_sync_callbacks(monk
     assert "test_proxy_utils" in captured["async_traceback"]
 
 
+@pytest.mark.parametrize(
+    "key_metadata, team_metadata, expected_to_run",
+    [
+        ({"guardrails": ["key-scoped-guardrail"]}, None, True),
+        ({}, {"guardrails": ["key-scoped-guardrail"]}, True),
+        ({"guardrails": ["some-other-guardrail"]}, None, False),
+        ({}, None, False),
+    ],
+)
+def test_convert_mcp_to_llm_format_carries_key_and_team_guardrails(key_metadata, team_metadata, expected_to_run):
+    proxy_logging = ProxyLogging(user_api_key_cache=DualCache())
+    guardrail = CustomGuardrail(guardrail_name="key-scoped-guardrail", event_hook="pre_mcp_call", default_on=False)
+    kwargs = {
+        "name": "ask_question",
+        "arguments": {"question": "hello"},
+        "server_name": "deepwiki",
+        "user_api_key_auth": UserAPIKeyAuth(metadata=key_metadata, team_metadata=team_metadata),
+    }
+    request_obj = proxy_logging._create_mcp_request_object_from_kwargs(kwargs)
+
+    with patch(  # test-quality-ok: the key-guardrail premium gate reads this proxy_server module global and has no injection seam
+        "litellm.proxy.proxy_server.premium_user", True
+    ):
+        synthetic = proxy_logging._convert_mcp_to_llm_format(request_obj, kwargs)
+
+    assert guardrail.should_run_guardrail(synthetic, GuardrailEventHooks.pre_mcp_call) is expected_to_run
+
+
 class _TracebackRecordingLogger(CustomLogger):
     def __init__(self) -> None:
         super().__init__()

From a2b7868a5bdb45030d4b224189eb1b1010ad1c4f Mon Sep 17 00:00:00 2001
From: yuneng-jiang <yuneng@berri.ai>
Date: Mon, 7 Sep 2026 09:53:18 -0700
Subject: [PATCH 65/79] test(ui): pin wire contracts for key, model and MCP
 server forms (#40019)

* test(ui): pin wire contracts for key, model and MCP server forms

Add vitest cases that pin what the key edit, key create, model edit and
MCP server edit forms put on the wire: an edited field reaches the
request with its new value, a cleared field reaches it as an explicit
null, and the dirty-only body is pinned as an expected failure until
each form moves to pickDirty. Model edit also pins the cost-map-derived
model_info fields as an expected failure.

KeyEditView hands a cleared max_budget to KeyInfoView as an empty
string and handleKeyUpdate maps it to null, so the null is pinned at
the /key/update boundary in key_info_view.test.tsx and the KeyEditView
case is an expected failure. buildEditServerPayload passes a cleared
description through as an empty string, so that case is an expected
failure too.

* test(ui): split masked model_info pins and retarget the create tracker

The model_info expected-failure case held three assertions, and it.fails
stops at the first one, so a later revamp that fixed max_input_tokens
while leaving mode leaking would still report an expected failure. Split
it into one case per pinned field group so each flips on its own.

The key create tracker asserted a body of only key_alias, which a create
can never send: key_type, user_id, duration and metadata are always
mounted. Retarget it at the real over-send, which is the Optional
Settings section adding fifteen undefined-valued keys when the user opens
it without filling anything in.
---
 .../editServerPayload.differential.test.ts    | 43 ++++++++++-
 .../src/components/model_info_view.test.tsx   | 72 +++++++++++++++++++
 .../create_key_button.integration.test.tsx    | 31 ++++++++
 .../templates/key_edit_view.test.tsx          | 52 ++++++++++++++
 .../templates/key_info_view.test.tsx          | 10 +++
 5 files changed, 207 insertions(+), 1 deletion(-)

diff --git a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.test.ts
index f9cab181e71..66786fac1c3 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.test.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/mcp-servers/_components/editServerPayload.differential.test.ts
@@ -12,7 +12,7 @@ import {
 } from "@/components/mcp_tools/types";
 import { AUTH_TYPES_REQUIRING_CREDENTIALS } from "./createServerPayload";
 import { TOOL_DISPLAY_NAME_PATTERN, normalizeEnvVars } from "./utils";
-import { buildEditServerPayload, type EditServerUiState } from "./editServerPayload";
+import { buildEditServerPayload, type EditServerFormValues, type EditServerUiState } from "./editServerPayload";
 import { CASES, baseUi } from "./editServerPayload.differential.cases";
 
 // GENERATED by scratchpad/emit_test.py. The body below is machine-extracted from
@@ -317,6 +317,47 @@ describe("buildEditServerPayload matches the pre-extraction handleSave body", ()
   });
 });
 
+const EDIT_FORM_VALUES: EditServerFormValues = {
+  server_name: "srv",
+  alias: "srv_alias",
+  description: "a server",
+  transport: "http",
+  url: "https://example.com/mcp",
+  auth_type: "none",
+  mcp_access_groups: [],
+  extra_headers: [],
+  static_headers: [],
+  env_vars: [],
+  allow_all_keys: false,
+  available_on_public_internet: true,
+};
+
+describe("buildEditServerPayload wire contract", () => {
+  it("carries an edited alias and the server identifier onto the wire", () => {
+    const result = buildEditServerPayload({ ...EDIT_FORM_VALUES, alias: "renamed" }, baseUi);
+
+    expect(result).toMatchObject({ kind: "ok", payload: { server_id: "srv_1", alias: "renamed" } });
+  });
+
+  it.fails(
+    "sends description as an explicit null when the field is cleared (expected to fail until the forms revamp, tri-state PATCH tracker: today the cleared field reaches the wire as an empty string)",
+    () => {
+      const result = buildEditServerPayload({ ...EDIT_FORM_VALUES, description: "" }, baseUi);
+
+      expect(result).toMatchObject({ kind: "ok", payload: { description: null } });
+    },
+  );
+
+  it.fails(
+    "sends only the server identifier and the edited alias (expected to fail until the forms revamp, tri-state PATCH tracker)",
+    () => {
+      const result = buildEditServerPayload({ ...EDIT_FORM_VALUES, alias: "renamed" }, baseUi);
+
+      expect(result).toStrictEqual({ kind: "ok", payload: { server_id: "srv_1", alias: "renamed" } });
+    },
+  );
+});
+
 void ADMIN_CONFIG_CREDENTIAL_KEYS;
 void AUTH_TYPE;
 void AUTH_TYPES_REQUIRING_CREDENTIALS;
diff --git a/ui/litellm-dashboard/src/components/model_info_view.test.tsx b/ui/litellm-dashboard/src/components/model_info_view.test.tsx
index 768183907db..3db9418dfb9 100644
--- a/ui/litellm-dashboard/src/components/model_info_view.test.tsx
+++ b/ui/litellm-dashboard/src/components/model_info_view.test.tsx
@@ -1729,5 +1729,77 @@ describe("ModelInfoView", () => {
         expect(payload.litellm_params.cache_control_injection_points).toEqual([{ location: "message", index: "2" }]);
       });
     });
+
+    const setInputCost = (value: string) => {
+      fireEvent.change(screen.getByPlaceholderText("Enter input cost"), { target: { value } });
+    };
+
+    it("carries an edited input cost and the model identifier onto the wire", async () => {
+      const user = userEvent.setup();
+      await enterEditMode(user);
+      setInputCost("5");
+      const payload = await save(user);
+
+      expect(mockModelPatchUpdateCall.mock.calls[0][2]).toBe("123");
+      expect(payload.litellm_params.input_cost_per_token).toBe(5 / 1_000_000);
+    });
+
+    it.fails(
+      "sends only the edited input cost (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        const user = userEvent.setup();
+        await enterEditMode(user);
+        setInputCost("5");
+        const payload = await save(user);
+
+        expect(payload).toStrictEqual({ litellm_params: { input_cost_per_token: 5 / 1_000_000 } });
+      },
+    );
+
+    const savePayloadAfterCostEditOnResolvedModel = async () => {
+      const resolved = {
+        ...defaultModelData,
+        model_info: {
+          ...defaultModelData.model_info,
+          max_input_tokens: 128_000,
+          mode: "chat",
+          supports_vision: true,
+          supports_function_calling: true,
+        },
+      };
+      mockUseModelsInfo.mockReturnValue({ data: { data: [resolved] }, isLoading: false, error: null });
+      mockModelInfoV1Call.mockResolvedValue({ data: [resolved] });
+      const user = userEvent.setup();
+      await enterEditMode(user);
+      setInputCost("5");
+      return save(user);
+    };
+
+    it.fails(
+      "leaves max_input_tokens off the wire when only the input cost is edited (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        const payload = await savePayloadAfterCostEditOnResolvedModel();
+
+        expect(payload.model_info).not.toHaveProperty("max_input_tokens");
+      },
+    );
+
+    it.fails(
+      "leaves mode off the wire when only the input cost is edited (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        const payload = await savePayloadAfterCostEditOnResolvedModel();
+
+        expect(payload.model_info).not.toHaveProperty("mode");
+      },
+    );
+
+    it.fails(
+      "leaves every supports_ capability off the wire when only the input cost is edited (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        const payload = await savePayloadAfterCostEditOnResolvedModel();
+
+        expect(Object.keys(payload.model_info).filter((key) => key.startsWith("supports_"))).toStrictEqual([]);
+      },
+    );
   });
 });
diff --git a/ui/litellm-dashboard/src/components/organisms/create_key_button.integration.test.tsx b/ui/litellm-dashboard/src/components/organisms/create_key_button.integration.test.tsx
index 045dcfa3ceb..3471ef00eb0 100644
--- a/ui/litellm-dashboard/src/components/organisms/create_key_button.integration.test.tsx
+++ b/ui/litellm-dashboard/src/components/organisms/create_key_button.integration.test.tsx
@@ -544,6 +544,37 @@ describe("CreateKey", () => {
 
       expect((await createdPayload()).metadata).toBe('{"team":"research"}');
     });
+
+    it("carries the typed key alias and the chosen team onto the wire", async () => {
+      state.teams = [{ team_id: "team-1", team_alias: "Team One", models: [] }];
+      await openModal({ teams: state.teams as unknown as Team[] });
+      await nameTheKey("wire-alias");
+      await userEvent.click(await screen.findByLabelText("Team"));
+      await userEvent.click(await screen.findByRole("option", { name: /Team One/ }));
+      await submit();
+
+      expect(await createdPayload()).toMatchObject({ key_alias: "wire-alias", team_id: "team-1" });
+    });
+
+    it("sends team_id as an explicit null when no team is chosen", async () => {
+      await openModal();
+      await nameTheKey();
+      await submit();
+
+      expect(await createdPayload()).toHaveProperty("team_id", null);
+    });
+
+    it.fails(
+      "adds no keys for an Optional Settings section the user opened but never filled (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        await openModal();
+        await nameTheKey();
+        await openSection(/Optional Settings/i);
+        await submit();
+
+        expect(await createdPayload()).toStrictEqual(ALL_CLOSED_PAYLOAD);
+      },
+    );
   });
 
   describe("key ownership", () => {
diff --git a/ui/litellm-dashboard/src/components/templates/key_edit_view.test.tsx b/ui/litellm-dashboard/src/components/templates/key_edit_view.test.tsx
index 8b506329595..4f732083c3e 100644
--- a/ui/litellm-dashboard/src/components/templates/key_edit_view.test.tsx
+++ b/ui/litellm-dashboard/src/components/templates/key_edit_view.test.tsx
@@ -2300,5 +2300,57 @@ describe("KeyEditView", () => {
       });
       expect(onSubmitMock.mock.calls[0][0]).toHaveProperty("tag_rpm_limit", { "test-tag": 7 });
     });
+
+    const setRpmLimit = (value: string) => {
+      fireEvent.change(screen.getByLabelText("RPM Limit"), { target: { value } });
+    };
+
+    it("carries an edited RPM limit and the key identifier onto the wire", async () => {
+      const onSubmitMock = vi.fn().mockResolvedValue(undefined);
+      renderForPayload(onSubmitMock);
+      await screen.findByRole("button", { name: /save changes/i });
+
+      setRpmLimit("25");
+      await userEvent.click(screen.getByRole("button", { name: /save changes/i }));
+
+      await waitFor(() => {
+        expect(onSubmitMock).toHaveBeenCalledTimes(1);
+      });
+      expect(onSubmitMock.mock.calls[0][0]).toMatchObject({ token: "test-token-123", rpm_limit: "25" });
+    });
+
+    it.fails(
+      "sends max_budget as an explicit null when the field is cleared (expected to fail until the forms revamp, tri-state PATCH tracker: today the view hands KeyInfoView an empty string and handleKeyUpdate maps it to null)",
+      async () => {
+        const onSubmitMock = vi.fn().mockResolvedValue(undefined);
+        renderForPayload(onSubmitMock);
+        await screen.findByRole("button", { name: /save changes/i });
+
+        await userEvent.clear(screen.getByLabelText("Max Budget (USD)"));
+        await userEvent.click(screen.getByRole("button", { name: /save changes/i }));
+
+        await waitFor(() => {
+          expect(onSubmitMock).toHaveBeenCalledTimes(1);
+        });
+        expect(onSubmitMock.mock.calls[0][0]).toHaveProperty("max_budget", null);
+      },
+    );
+
+    it.fails(
+      "sends only the key identifier and the edited RPM limit (expected to fail until the forms revamp, tri-state PATCH tracker)",
+      async () => {
+        const onSubmitMock = vi.fn().mockResolvedValue(undefined);
+        renderForPayload(onSubmitMock);
+        await screen.findByRole("button", { name: /save changes/i });
+
+        setRpmLimit("25");
+        await userEvent.click(screen.getByRole("button", { name: /save changes/i }));
+
+        await waitFor(() => {
+          expect(onSubmitMock).toHaveBeenCalledTimes(1);
+        });
+        expect(onSubmitMock.mock.calls[0][0]).toStrictEqual({ token: "test-token-123", rpm_limit: "25" });
+      },
+    );
   });
 });
diff --git a/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx b/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
index 7817d4e7cea..7a27cc41e5e 100644
--- a/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
+++ b/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
@@ -1015,6 +1015,16 @@ describe("KeyInfoView", () => {
 
       expect(keyUpdateCall).toHaveBeenCalledWith(expect.anything(), expect.objectContaining({ policies: [] }));
     });
+
+    it("puts the key identifier and an explicit null max_budget on the wire when the edit view hands over a cleared budget", async () => {
+      await enterEditMode({ ...MOCK_KEY_DATA, user_id: "proxy-admin-user" } as KeyResponse);
+      await editViewMocks.onSubmit!({ token: MOCK_KEY_DATA.token, max_budget: "" });
+
+      expect(keyUpdateCall).toHaveBeenCalledWith(
+        expect.anything(),
+        expect.objectContaining({ key: "test-token-123", max_budget: null }),
+      );
+    });
   });
 
   describe("MCP tool permissions on save", () => {

From f40c3842dc4cebbf1eb02379a492acb1d0c87605 Mon Sep 17 00:00:00 2001
From: Yuneng Jiang <yuneng@berri.ai>
Date: Mon, 7 Sep 2026 10:08:19 -0700
Subject: [PATCH 66/79] fix(router): serialize heuristic tuning quota filters
 for Prisma

---
 litellm/repositories/model_repository.py      |  3 +--
 .../repositories/test_repositories.py         | 22 ++++++++++++++++++-
 2 files changed, 22 insertions(+), 3 deletions(-)

diff --git a/litellm/repositories/model_repository.py b/litellm/repositories/model_repository.py
index 540f197044f..d24eb8ffc62 100644
--- a/litellm/repositories/model_repository.py
+++ b/litellm/repositories/model_repository.py
@@ -4,7 +4,6 @@ Model repository for database operations on LiteLLM_ProxyModelTable.
 
 import json
 from collections.abc import Mapping, Sequence
-from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Protocol
 
 from litellm.models.model import LiteLLM_ProxyModelTable
@@ -109,7 +108,7 @@ class ModelRepository(BaseRepository[LiteLLM_ProxyModelTable]):
     async def find_all_except(self, model_id: str) -> Sequence[LiteLLM_ProxyModelTable]:
         """Find every model except the row currently being updated."""
         records: Final = await self.table.find_many(
-            where=MappingProxyType({"model_id": MappingProxyType({"not": model_id})})
+            where={"model_id": {"not": model_id}}  # mutable-ok: Prisma requires plain dicts for query serialization
         )
         return tuple(self._to_model_list(records))
 
diff --git a/tests/test_litellm/repositories/test_repositories.py b/tests/test_litellm/repositories/test_repositories.py
index 758d379f22c..a890d7ceed0 100644
--- a/tests/test_litellm/repositories/test_repositories.py
+++ b/tests/test_litellm/repositories/test_repositories.py
@@ -4,10 +4,13 @@ Tests for gateway repository layer.
 
 import json
 from datetime import datetime
-from typing import Any, Dict, List, Optional
+from types import SimpleNamespace
+from typing import Any, Dict, Final, List, Optional
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
+from prisma import models as prisma_models
+from prisma.builder import QueryBuilder
 
 from litellm.models.base import DomainModel
 from litellm.models.budget import LiteLLM_BudgetTable
@@ -307,6 +310,23 @@ class TestModelRepository:
         client = MockPrismaClient()
         return ModelRepository(client)
 
+    @pytest.mark.asyncio
+    async def test_find_all_except_serializes_exclusion_for_prisma(self) -> None:
+        find_many: Final = AsyncMock(return_value=[])
+        client: Final = SimpleNamespace(
+            db=SimpleNamespace(litellm_proxymodeltable=SimpleNamespace(find_many=find_many))
+        )
+
+        await ModelRepository(client).find_all_except("current-model")
+
+        find_many.assert_awaited_once()
+        query: Final = QueryBuilder(
+            method="find_many",
+            model=prisma_models.LiteLLM_ProxyModelTable,
+            arguments=find_many.call_args.kwargs,
+        ).build_query()
+        assert 'where: { model_id: { not: "current-model" } }' in " ".join(query.split())
+
     def test_table_is_wrapped_for_config_sync(self, repo):
         from litellm.proxy.common_utils.config_sync_pubsub import (
             _PublishOnWriteActions,

From 217cb126236cb2a73245f7c3ac5525f88dfc0f7c Mon Sep 17 00:00:00 2001
From: yujonglee <yujong@berri.ai>
Date: Mon, 7 Sep 2026 10:43:45 -0700
Subject: [PATCH 67/79] refactor(rust): remove per-request enablement arguments
 (#39928)

* refactor(rust): remove per-request enablement arguments

* fix(rust): remove ignored transcription enablement

* refactor(rust): remove OCR-specific bridge controls
---
 .../litellm_core_utils/get_litellm_params.py  |  9 +--
 litellm/llms/custom_httpx/llm_http_handler.py | 11 +--
 litellm/ocr/main.py                           | 11 +--
 litellm/rust_bridge/chat_completions.py       |  3 +-
 litellm/rust_bridge/configuration.py          | 33 +-------
 litellm/rust_bridge/ocr.py                    | 46 +++--------
 litellm/rust_bridge/transcription.py          |  1 -
 litellm/types/router.py                       |  1 -
 .../test_rust_bridge_messages.py              | 26 +-----
 .../test_get_litellm_params.py                | 29 -------
 .../chat/test_anthropic_chat_handler.py       |  8 +-
 .../chat/test_bedrock_converse_handler.py     | 12 ++-
 .../custom_httpx/test_llm_http_handler.py     | 22 +++--
 tests/test_litellm/ocr/test_rust_bridge.py    | 80 ++++++++-----------
 .../responses/test_rust_bridge_websocket.py   | 21 ++---
 .../rust_bridge/test_chat_completions.py      | 38 ++++-----
 .../rust_bridge/test_configuration.py         | 61 +++-----------
 .../test_audio_transcription_rust_bridge.py   |  4 +-
 tests/test_litellm/test_utils.py              | 46 -----------
 ui/litellm-dashboard/src/lib/http/schema.d.ts |  4 -
 20 files changed, 109 insertions(+), 357 deletions(-)

diff --git a/litellm/litellm_core_utils/get_litellm_params.py b/litellm/litellm_core_utils/get_litellm_params.py
index 389e6f7f501..1fd79db15a6 100644
--- a/litellm/litellm_core_utils/get_litellm_params.py
+++ b/litellm/litellm_core_utils/get_litellm_params.py
@@ -21,13 +21,10 @@ AWS_CREDENTIAL_KWARGS_KEYS: Final = frozenset(
     }
 )
 
-# The per-deployment Rust opt-in.
-RUST_KWARG_KEY: Final = "rust"
-
 # Keys `completion()` forwards from its own kwargs into `get_litellm_params`,
 # which are otherwise invisible to it because that call site passes explicit
 # named arguments rather than `**kwargs`.
-FORWARDED_KWARGS_KEYS: Final = AWS_CREDENTIAL_KWARGS_KEYS | frozenset({RUST_KWARG_KEY})
+FORWARDED_KWARGS_KEYS: Final = AWS_CREDENTIAL_KWARGS_KEYS
 
 # Pre-define optional kwargs keys as frozenset for O(1) lookups
 # These are extracted from kwargs only if present, avoiding unnecessary .get() calls
@@ -58,10 +55,6 @@ OPTIONAL_KWARGS_KEYS: Final = (
             "itpm",
             "otpm",
             "use_xai_oauth",
-            # The per-deployment Rust opt-in. `all_litellm_params` keeps it out
-            # of the provider body; this keeps it *in* litellm_params, which is
-            # where the chat completions handlers read it from.
-            RUST_KWARG_KEY,
         }
     )
     | AWS_CREDENTIAL_KWARGS_KEYS
diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py
index f281c249c72..552b1549db8 100644
--- a/litellm/llms/custom_httpx/llm_http_handler.py
+++ b/litellm/llms/custom_httpx/llm_http_handler.py
@@ -163,13 +163,10 @@ from litellm.utils import (
 
 def _rust_responses_websocket_enabled(
     custom_llm_provider: str | None,
-    litellm_params: GenericLiteLLMParams,
 ) -> bool:
     from litellm.rust_bridge.configuration import rust_enabled
 
-    raw_request_override: Final = litellm_params.get("rust")
-    request_override: Final = raw_request_override if isinstance(raw_request_override, bool) else None
-    return custom_llm_provider == "openai" and rust_enabled(request_override=request_override)
+    return custom_llm_provider == "openai" and rust_enabled()
 
 
 from .http_handler import get_shared_realtime_ssl_context
@@ -2403,9 +2400,7 @@ class BaseLLMHTTPHandler:
             return None
         from litellm.rust_bridge.configuration import rust_enabled
 
-        raw_request_override: Final = litellm_params.get("rust")
-        request_override: Final = raw_request_override if isinstance(raw_request_override, bool) else None
-        if not rust_enabled(request_override=request_override):
+        if not rust_enabled():
             return None
         if has_agentic_hook:
             return None
@@ -6514,7 +6509,7 @@ class BaseLLMHTTPHandler:
 
             @asynccontextmanager
             async def _backend_connection():
-                if _rust_responses_websocket_enabled(custom_llm_provider, litellm_params):
+                if _rust_responses_websocket_enabled(custom_llm_provider):
                     from litellm.rust_bridge import responses_websocket as rust_responses_websocket
 
                     rust_backend: Final = await rust_responses_websocket.connect(
diff --git a/litellm/ocr/main.py b/litellm/ocr/main.py
index 6c68971f8d5..56c2292d00c 100644
--- a/litellm/ocr/main.py
+++ b/litellm/ocr/main.py
@@ -29,6 +29,7 @@ from litellm.llms.base_llm.ocr.transformation import (
 )
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
 from litellm.rust_bridge import ocr as rust_ocr_bridge
+from litellm.rust_bridge.configuration import rust_enabled
 from litellm.types.router import GenericLiteLLMParams
 from litellm.utils import ProviderConfigManager, client
 
@@ -196,12 +197,6 @@ def _rust_ocr_supported(prepared_request: _PreparedOCRRequest) -> bool:
     return prepared_request.custom_llm_provider in _RUST_OCR_PROVIDERS
 
 
-def _rust_ocr_enabled(prepared_request: _PreparedOCRRequest) -> bool:
-    raw_request_override: Final = prepared_request.litellm_params.get("rust")
-    request_override: Final = raw_request_override if isinstance(raw_request_override, bool) else None
-    return rust_ocr_bridge.rust_ocr_enabled(request_override=request_override)
-
-
 def _rust_bridge_optional_params(
     prepared_request: _PreparedOCRRequest,
     resolve_secret: Callable[[str], str | None],
@@ -430,7 +425,7 @@ async def aocr(
         custom_llm_provider = prepared.custom_llm_provider
         completion_kwargs.update({"model": model, "custom_llm_provider": custom_llm_provider})
 
-        if _rust_ocr_supported(prepared) and _rust_ocr_enabled(prepared):
+        if _rust_ocr_supported(prepared) and rust_enabled():
             from litellm.secret_managers.main import get_secret_str
 
             rust_response: Final = await _run_rust_aocr(
@@ -702,7 +697,7 @@ def ocr(
         custom_llm_provider = prepared.custom_llm_provider
         completion_kwargs.update({"model": model, "custom_llm_provider": custom_llm_provider})
 
-        if _rust_ocr_supported(prepared) and _rust_ocr_enabled(prepared):
+        if _rust_ocr_supported(prepared) and rust_enabled():
             from litellm.secret_managers.main import get_secret_str
 
             rust_response: Final = _run_rust_ocr(
diff --git a/litellm/rust_bridge/chat_completions.py b/litellm/rust_bridge/chat_completions.py
index c599667ab17..674bd8847f7 100644
--- a/litellm/rust_bridge/chat_completions.py
+++ b/litellm/rust_bridge/chat_completions.py
@@ -247,8 +247,7 @@ def rust_chat_completions_accepts(
         return False
     if stream:
         return False
-    request_override: Final = litellm_params.get("rust") if litellm_params is not None else None
-    if not rust_enabled(request_override=request_override if isinstance(request_override, bool) else None):
+    if not rust_enabled():
         return False
     if _litellm_metadata_reaches_the_provider(custom_llm_provider, litellm_params):
         verbose_logger.debug("Rust chat completions declined (litellm metadata user_id); using the Python path")
diff --git a/litellm/rust_bridge/configuration.py b/litellm/rust_bridge/configuration.py
index 515ab6edef1..5582027bb5d 100644
--- a/litellm/rust_bridge/configuration.py
+++ b/litellm/rust_bridge/configuration.py
@@ -1,13 +1,11 @@
 from __future__ import annotations
 
 import os
-import warnings
 from typing import Final
 
 DEFAULT_RUST_ENABLED: Final = False
 _TRUE_ENV_VALUES: Final = frozenset({"1", "true", "yes", "on"})
 _GLOBAL_ENV_NAME: Final = "LITELLM_RUST"
-_LEGACY_OCR_ENV_NAME: Final = "LITELLM_USE_RUST_OCR"
 
 
 class _RustConfiguration:
@@ -26,49 +24,24 @@ def _parse_env_bool(value: str | None) -> bool | None:
 
 def resolve_rust_enabled(
     *,
-    request_override: bool | None,
     process_override: bool | None,
     environment_override: bool | None,
-    legacy_environment_override: bool | None = None,
     release_default: bool = DEFAULT_RUST_ENABLED,
 ) -> bool:
-    if request_override is not None:
-        return request_override
     if process_override is not None:
         return process_override
     if environment_override is not None:
         return environment_override
-    if legacy_environment_override is not None:
-        return legacy_environment_override
     return release_default
 
 
-def rust_enabled(*, request_override: bool | None = None) -> bool:
-    if request_override is not None:
-        return request_override
-    process_override: Final = _CONFIGURATION.override
-    if process_override is not None:
-        return process_override
-    global_override: Final = _parse_env_bool(os.getenv(_GLOBAL_ENV_NAME))
-    legacy_override: Final = None if global_override is not None else _parse_env_bool(os.getenv(_LEGACY_OCR_ENV_NAME))
-    if legacy_override is not None:
-        warnings.warn(
-            f"{_LEGACY_OCR_ENV_NAME} is deprecated; use {_GLOBAL_ENV_NAME} instead",
-            DeprecationWarning,
-            stacklevel=2,
-        )
+def rust_enabled() -> bool:
     return resolve_rust_enabled(
-        request_override=None,
-        process_override=None,
-        environment_override=global_override,
-        legacy_environment_override=legacy_override,
+        process_override=_CONFIGURATION.override,
+        environment_override=_parse_env_bool(os.getenv(_GLOBAL_ENV_NAME)),
     )
 
 
-def rust_ocr_enabled(*, request_override: bool | None = None) -> bool:
-    return rust_enabled(request_override=request_override)
-
-
 def reset_rust_configuration() -> None:
     _CONFIGURATION.override = None
 
diff --git a/litellm/rust_bridge/ocr.py b/litellm/rust_bridge/ocr.py
index 86038438f57..b7fdb5a98ef 100644
--- a/litellm/rust_bridge/ocr.py
+++ b/litellm/rust_bridge/ocr.py
@@ -7,12 +7,9 @@ from typing import Final, Protocol, cast  # noqa: TID251  # native extension exp
 
 import httpx
 
-from litellm.rust_bridge import configuration as _configuration
+from litellm.rust_bridge.bindings import NativeBinding
 from litellm.rust_bridge.timeouts import timeout_to_seconds as _timeout_to_seconds
 
-rust_ocr_enabled = _configuration.rust_ocr_enabled
-rust = _configuration.rust
-
 
 class RustOcr(Protocol):
     def __call__(
@@ -44,49 +41,24 @@ class RustAocr(Protocol):
         raise NotImplementedError
 
 
-class _Unset:
-    pass
+def _as_ocr(value: object) -> RustOcr | None:
+    return cast(RustOcr, value) if callable(value) else None
 
 
-_UNSET: Final[_Unset] = _Unset()
+def _as_aocr(value: object) -> RustAocr | None:
+    return cast(RustAocr, value) if callable(value) else None
 
 
-_rust_ocr_impl: RustOcr | None = None
-_rust_aocr_impl: RustAocr | None = None
-
-
-def set_rust_ocr(
-    *,
-    ocr: RustOcr | None | _Unset = _UNSET,
-    aocr: RustAocr | None | _Unset = _UNSET,
-) -> None:
-    global _rust_ocr_impl, _rust_aocr_impl
-    if not isinstance(ocr, _Unset):
-        _rust_ocr_impl = ocr
-    if not isinstance(aocr, _Unset):
-        _rust_aocr_impl = aocr
+_OCR: Final = NativeBinding("ocr", validate=_as_ocr)
+_AOCR: Final = NativeBinding("aocr", validate=_as_aocr)
 
 
 def load_rust_ocr() -> RustOcr | None:
-    if _rust_ocr_impl is not None:
-        return _rust_ocr_impl
-    from litellm.rust_bridge import get_native_bridge
-
-    native_bridge: Final = get_native_bridge()
-    if native_bridge is None:
-        return None
-    return cast(RustOcr, native_bridge.ocr)
+    return _OCR.load()
 
 
 def load_rust_aocr() -> RustAocr | None:
-    if _rust_aocr_impl is not None:
-        return _rust_aocr_impl
-    from litellm.rust_bridge import get_native_bridge
-
-    native_bridge: Final = get_native_bridge()
-    if native_bridge is None:
-        return None
-    return cast(RustAocr, getattr(native_bridge, "aocr", None))
+    return _AOCR.load()
 
 
 def ocr(
diff --git a/litellm/rust_bridge/transcription.py b/litellm/rust_bridge/transcription.py
index 3d71f6f8a50..6c81786accd 100644
--- a/litellm/rust_bridge/transcription.py
+++ b/litellm/rust_bridge/transcription.py
@@ -56,7 +56,6 @@ _STATE: Final = _RustTranscriptionState()
 
 
 def configure_rust_transcription(
-    enabled: bool = True,
     *,
     transcription: RustTranscription | None | _Unset = _UNSET,
     atranscription: RustAtranscription | None | _Unset = _UNSET,
diff --git a/litellm/types/router.py b/litellm/types/router.py
index 2dad22751de..0db482d8a58 100644
--- a/litellm/types/router.py
+++ b/litellm/types/router.py
@@ -307,7 +307,6 @@ class GenericLiteLLMParams(CredentialLiteLLMParams, CustomPricingLiteLLMParams):
     """
 
     custom_llm_provider: str | None = None
-    rust: bool | None = None
     tpm: int | None = None
     rpm: int | None = None
     itpm: int | None = None
diff --git a/tests/test_litellm/anthropic_interface/test_rust_bridge_messages.py b/tests/test_litellm/anthropic_interface/test_rust_bridge_messages.py
index b2cf253d164..a30474245c6 100644
--- a/tests/test_litellm/anthropic_interface/test_rust_bridge_messages.py
+++ b/tests/test_litellm/anthropic_interface/test_rust_bridge_messages.py
@@ -126,16 +126,6 @@ def test_load_rust_messages_returns_injected_impl():
     assert rust_messages.load_rust_messages() is bridge
 
 
-def test_bare_rust_still_toggles_ocr():
-    from litellm.rust_bridge.ocr import rust_ocr_enabled
-
-    litellm.rust(True)
-    assert rust_ocr_enabled() is True
-
-    litellm.rust(False)
-    assert rust_ocr_enabled() is False
-
-
 def test_load_rust_amessages_returns_injected_impl():
     bridge = RecordingAsyncMessages()
     litellm.rust(True)
@@ -214,7 +204,7 @@ async def test_amessages_wrapper_forwards_args():
 def _gate(**overrides):
     kwargs = {
         "custom_llm_provider": "azure_ai",
-        "litellm_params": GenericLiteLLMParams(api_key="sk-azure", rust=True),
+        "litellm_params": GenericLiteLLMParams(api_key="sk-azure"),
         "has_agentic_hook": False,
         "model": "claude-sonnet-4-5",
         "api_key": "sk-azure",
@@ -282,18 +272,6 @@ async def test_gate_uses_process_enable_without_request_override():
     assert bridge.calls[0]["custom_llm_provider"] == "azure_ai"
 
 
-@pytest.mark.asyncio
-async def test_gate_skips_rust_when_flag_false():
-    bridge = ExplodingAsyncMessages()
-    litellm.rust(True)
-    rust_messages.set_rust_messages(amessages=bridge)
-
-    response = await _gate(litellm_params=GenericLiteLLMParams(api_key="sk-azure", rust=False))
-
-    assert response is None
-    assert bridge.calls == 0
-
-
 @pytest.mark.asyncio
 async def test_gate_invokes_rust_for_native_anthropic_provider():
     bridge = RecordingAsyncMessages()
@@ -302,7 +280,7 @@ async def test_gate_invokes_rust_for_native_anthropic_provider():
 
     response = await _gate(
         custom_llm_provider="anthropic",
-        litellm_params=GenericLiteLLMParams(api_key="sk-ant", rust=True),
+        litellm_params=GenericLiteLLMParams(api_key="sk-ant"),
         api_key="sk-ant",
         api_base="https://api.anthropic.com",
         headers={"x-api-key": "sk-ant", "anthropic-version": "2023-06-01"},
diff --git a/tests/test_litellm/litellm_core_utils/test_get_litellm_params.py b/tests/test_litellm/litellm_core_utils/test_get_litellm_params.py
index 956da571d43..fb4cb494bee 100644
--- a/tests/test_litellm/litellm_core_utils/test_get_litellm_params.py
+++ b/tests/test_litellm/litellm_core_utils/test_get_litellm_params.py
@@ -215,32 +215,3 @@ class TestMetadataFallsBackToLitellmMetadata:
         assert result["metadata"] is not litellm_metadata
         result["metadata"].pop("trace_id")
         assert litellm_metadata == {"trace_id": "trace-1"}
-
-
-class TestRustOptIn:
-    """`rust: true` is a litellm param, so it has to reach `litellm_params`.
-
-    `all_litellm_params` keeps it out of the provider body; without it also
-    being carried into `litellm_params` the chat completions handlers cannot
-    see the opt-in and the Rust path is silently never taken.
-    """
-
-    def test_rust_is_an_optional_kwargs_key(self):
-        assert "rust" in _OPTIONAL_KWARGS_KEYS
-
-    def test_rust_is_forwarded_from_completion_kwargs(self):
-        from litellm.litellm_core_utils.get_litellm_params import FORWARDED_KWARGS_KEYS
-
-        assert "rust" in FORWARDED_KWARGS_KEYS
-
-    def test_rust_survives_into_litellm_params(self):
-        params = get_litellm_params(rust=True)
-        assert params["rust"] is True
-
-    def test_rust_is_absent_when_the_deployment_did_not_set_it(self):
-        assert "rust" not in get_litellm_params()
-
-    def test_rust_stays_out_of_the_provider_body(self):
-        from litellm.types.utils import all_litellm_params
-
-        assert "rust" in all_litellm_params
diff --git a/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py b/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py
index 043537f8c1f..b4b173b20c3 100644
--- a/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py
+++ b/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_handler.py
@@ -2256,7 +2256,7 @@ class TestRustChatCompletionsHook:
     def _reset_bridge(self, monkeypatch):
         from litellm.rust_bridge import chat_completions as bridge
 
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         bridge.set_rust_chat_completions(
             chat_completions=None, achat_completions=None, decline=None
         )
@@ -2282,7 +2282,7 @@ class TestRustChatCompletionsHook:
             "logging_obj": MagicMock(),
             "optional_params": {"max_tokens": 16},
             "timeout": 30.0,
-            "litellm_params": {"rust": True},
+            "litellm_params": {},
             "acompletion": False,
             "headers": {},
             "client": None,
@@ -2366,7 +2366,8 @@ class TestRustChatCompletionsHook:
         )
         assert seen["call"][0]["optional_params"]["max_tokens"] == 7
 
-    def test_without_the_opt_in_the_core_is_never_consulted(self):
+    def test_without_the_opt_in_the_core_is_never_consulted(self, monkeypatch):
+        monkeypatch.setenv("LITELLM_RUST", "0")
         from litellm.llms.anthropic.chat.handler import AnthropicChatCompletion
         from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 
@@ -2587,6 +2588,7 @@ class TestRustChatCompletionsHook:
 
     def test_pre_call_logging_still_fires_when_rust_is_not_involved(self, monkeypatch):
         """The suppression must not swallow the log on the ordinary path."""
+        monkeypatch.setenv("LITELLM_RUST", "0")
         from litellm.llms.anthropic.chat.handler import AnthropicChatCompletion
         from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 
diff --git a/tests/test_litellm/llms/bedrock/chat/test_bedrock_converse_handler.py b/tests/test_litellm/llms/bedrock/chat/test_bedrock_converse_handler.py
index c4d6896b17b..f34b8eb1fb9 100644
--- a/tests/test_litellm/llms/bedrock/chat/test_bedrock_converse_handler.py
+++ b/tests/test_litellm/llms/bedrock/chat/test_bedrock_converse_handler.py
@@ -48,7 +48,7 @@ RESOLVED_CREDENTIALS = Credentials(
 
 @pytest.fixture(autouse=True)
 def reset_bridge(monkeypatch):
-    monkeypatch.delenv("LITELLM_RUST", raising=False)
+    monkeypatch.setenv("LITELLM_RUST", "1")
     bridge.set_rust_chat_completions(
         chat_completions=None, achat_completions=None, decline=None
     )
@@ -87,7 +87,7 @@ def _completion_kwargs(**overrides):
         "optional_params": {"maxTokens": 16},
         "acompletion": False,
         "timeout": 30.0,
-        "litellm_params": {"rust": True},
+        "litellm_params": {},
         "extra_headers": None,
         "client": None,
         "api_key": None,
@@ -157,7 +157,8 @@ def test_the_core_receives_the_untranslated_openai_messages():
     ]
 
 
-def test_without_the_opt_in_the_core_is_never_consulted():
+def test_without_the_opt_in_the_core_is_never_consulted(monkeypatch):
+    monkeypatch.setenv("LITELLM_RUST", "0")
     seen = _inject()
     try:
         _run(litellm_params={})
@@ -401,9 +402,10 @@ def test_pre_call_logging_fires_once_when_the_sync_rust_path_declines():
     assert logging_obj.pre_call.call_count == 1
 
 
-def test_the_sync_python_path_still_logs_pre_call_without_the_opt_in():
+def test_the_sync_python_path_still_logs_pre_call_without_the_opt_in(monkeypatch):
     """The suppression must not swallow the log on a request the gate declined,
     so a deployment with no `rust` flag keeps exactly the log it always had."""
+    monkeypatch.setenv("LITELLM_RUST", "0")
     logging_obj = MagicMock()
     response = _run(
         logging_obj=logging_obj,
@@ -491,6 +493,7 @@ def test_bearer_token_auth_serves_when_boto3_resolves_no_sigv4_credentials(monke
     """With only `AWS_BEARER_TOKEN_BEDROCK` configured boto3 resolves no
     credentials at all. Preparing the Rust handoff must not dereference that
     None: the bearer token signs the request on its own."""
+    monkeypatch.setenv("LITELLM_RUST", "0")
     monkeypatch.setenv("AWS_BEARER_TOKEN_BEDROCK", "bedrock-bearer-token")
     client = _sync_client_returning_converse_response()
 
@@ -520,6 +523,7 @@ def test_bearer_token_auth_never_runs_the_sigv4_credential_chain(monkeypatch, co
     """The deployment's AWS profile does not exist, so resolving SigV4 credentials
     raises; a bearer-token deployment must still serve the request, since the
     bearer token alone signs it."""
+    monkeypatch.setenv("LITELLM_RUST", "0")
     if configured_through == "env_var":
         monkeypatch.setenv("AWS_BEARER_TOKEN_BEDROCK", "bedrock-bearer-token")
     else:
diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
index 023e2d8843f..5c6cc4a52da 100644
--- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
+++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py
@@ -2689,20 +2689,18 @@ async def test_generic_http_handler_async_streaming_forwards_provider_response_h
 
 
 @pytest.mark.parametrize(
-    "custom_llm_provider, litellm_params, expected",
-    [
-        ("openai", GenericLiteLLMParams(rust=True), True),
-        ("openai", GenericLiteLLMParams(), False),
-        ("openai", GenericLiteLLMParams(rust=False), False),
-        ("azure", GenericLiteLLMParams(rust=True), False),
-        ("hosted_vllm", GenericLiteLLMParams(rust=True), False),
-        (None, GenericLiteLLMParams(rust=True), False),
-    ],
+    "custom_llm_provider, enabled, expected",
+    [("openai", True, True), ("openai", False, False), ("azure", True, False),
+     ("hosted_vllm", True, False), (None, True, False)],
 )
-def test_the_rust_responses_websocket_needs_both_openai_and_the_rust_flag(
-    custom_llm_provider, litellm_params, expected
+def test_the_rust_responses_websocket_needs_openai_and_process_enablement(
+    custom_llm_provider, enabled, expected, monkeypatch
 ):
-    assert _rust_responses_websocket_enabled(custom_llm_provider, litellm_params) is expected
+    from litellm.rust_bridge import configuration
+
+    configuration.reset_rust_configuration()
+    monkeypatch.setenv("LITELLM_RUST", "1" if enabled else "0")
+    assert _rust_responses_websocket_enabled(custom_llm_provider) is expected
 
 
 def test_a_plain_callback_does_not_advertise_a_pre_call_deployment_hook(monkeypatch):
diff --git a/tests/test_litellm/ocr/test_rust_bridge.py b/tests/test_litellm/ocr/test_rust_bridge.py
index 1c2e07e0d24..d28184b96a6 100644
--- a/tests/test_litellm/ocr/test_rust_bridge.py
+++ b/tests/test_litellm/ocr/test_rust_bridge.py
@@ -17,6 +17,7 @@ from litellm.rust_bridge import configuration
 # explicitly via importlib rather than attribute traversal.
 ocr_main = importlib.import_module("litellm.ocr.main")
 rust_bridge = importlib.import_module("litellm.rust_bridge.ocr")
+rust_bridge_bindings = importlib.import_module("litellm.rust_bridge.bindings")
 rust_bridge_loader = importlib.import_module("litellm.rust_bridge.loader")
 
 MODEL = "mistral/mistral-ocr-latest"
@@ -215,11 +216,13 @@ def build_prepared_request(
 @pytest.fixture(autouse=True)
 def _reset_rust_flag():
     """Keep the global toggle isolated between tests."""
-    rust_bridge.set_rust_ocr(ocr=None, aocr=None)
+    rust_bridge._OCR.reset()
+    rust_bridge._AOCR.reset()
     configuration.reset_rust_configuration()
     rust_bridge_loader._cached_bridge = rust_bridge_loader._BRIDGE_SENTINEL
     yield
-    rust_bridge.set_rust_ocr(ocr=None, aocr=None)
+    rust_bridge._OCR.reset()
+    rust_bridge._AOCR.reset()
     configuration.reset_rust_configuration()
     rust_bridge_loader._cached_bridge = rust_bridge_loader._BRIDGE_SENTINEL
 
@@ -229,7 +232,7 @@ def fake_bridge():
     """Enable the Rust path with an injected recording bridge (no native wheel)."""
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
     return bridge
 
 
@@ -238,34 +241,14 @@ def fake_async_bridge():
     """Enable the async Rust path with an injected recording bridge."""
     bridge = RecordingAsyncBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(aocr=bridge)
+    rust_bridge._AOCR.override(bridge)
     return bridge
 
 
-def test_rust_toggles_flag():
-    assert rust_bridge.rust_ocr_enabled() is False
-    litellm.rust(True)
-    assert rust_bridge.rust_ocr_enabled() is True
-    litellm.rust(False)
-    assert rust_bridge.rust_ocr_enabled() is False
-
-
-def test_env_var_enables_rust_ocr(monkeypatch):
-    monkeypatch.setenv("LITELLM_USE_RUST_OCR", "1")
-    with pytest.warns(DeprecationWarning, match="LITELLM_USE_RUST_OCR is deprecated"):
-        assert rust_bridge.rust_ocr_enabled() is True
-
-
-def test_explicit_false_overrides_process_enable():
-    litellm.rust(True)
-
-    assert ocr_main._rust_ocr_enabled(build_prepared_request(litellm_params={"rust": False})) is False
-
-
 def test_load_rust_ocr_returns_injected_impl():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
     assert rust_bridge.load_rust_ocr() is bridge
 
 
@@ -329,7 +312,7 @@ def test_native_bridge_available_reflects_loader(monkeypatch):
 def test_load_rust_aocr_returns_injected_impl():
     bridge = RecordingAsyncBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(aocr=bridge)
+    rust_bridge._AOCR.override(bridge)
     assert rust_bridge.load_rust_aocr() is bridge
 
 
@@ -338,7 +321,8 @@ def test_toggle_without_ocr_arg_preserves_injected_impl():
     bridge = RecordingBridge()
     async_bridge = RecordingAsyncBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge, aocr=async_bridge)
+    rust_bridge._OCR.override(bridge)
+    rust_bridge._AOCR.override(async_bridge)
 
     litellm.rust(False)
     assert rust_bridge.load_rust_ocr() is bridge
@@ -350,16 +334,18 @@ def test_toggle_without_ocr_arg_preserves_injected_impl():
 
 def test_explicit_ocr_none_clears_injected_impl(monkeypatch):
     monkeypatch.setattr(
-        importlib.import_module("litellm.rust_bridge"),
+        rust_bridge_bindings,
         "get_native_bridge",
         lambda: None,
     )
     bridge = RecordingBridge()
     async_bridge = RecordingAsyncBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge, aocr=async_bridge)
+    rust_bridge._OCR.override(bridge)
+    rust_bridge._AOCR.override(async_bridge)
 
-    rust_bridge.set_rust_ocr(ocr=None, aocr=None)
+    rust_bridge._OCR.override(None)
+    rust_bridge._AOCR.override(None)
     assert rust_bridge.load_rust_ocr() is None
     assert rust_bridge.load_rust_aocr() is None
 
@@ -368,7 +354,7 @@ def test_load_rust_ocr_none_when_extension_absent(monkeypatch):
     """With no injected impl and no compiled wheel, the loader returns None so the
     caller degrades to the Python path instead of raising ImportError."""
     monkeypatch.setattr(
-        importlib.import_module("litellm.rust_bridge"),
+        rust_bridge_bindings,
         "get_native_bridge",
         lambda: None,
     )
@@ -385,7 +371,7 @@ def test_load_rust_ocr_uses_compiled_extension(monkeypatch):
     fake_module.ocr = lambda **kwargs: dict(FAKE_OCR_RESPONSE)  # type: ignore[attr-defined]
     fake_module.aocr = lambda **kwargs: dict(FAKE_OCR_RESPONSE)  # type: ignore[attr-defined]
     monkeypatch.setattr(
-        importlib.import_module("litellm.rust_bridge"),
+        rust_bridge_bindings,
         "get_native_bridge",
         lambda: fake_module,
     )
@@ -406,7 +392,7 @@ def test_bridge_wrapper_forwards_prepared_args_and_wraps_response():
 
     litellm.rust(True)
 
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
     response = rust_bridge.ocr(
         model="mistral-ocr-latest",
         document=DOCUMENT,
@@ -441,7 +427,7 @@ async def test_bridge_wrapper_forwards_prepared_async_args_and_wraps_response():
 
     litellm.rust(True)
 
-    rust_bridge.set_rust_ocr(aocr=bridge)
+    rust_bridge._AOCR.override(bridge)
     response = await rust_bridge.aocr(
         model="mistral-ocr-maas",
         document=DOCUMENT,
@@ -470,7 +456,7 @@ def test_run_rust_ocr_prepares_request_and_wraps_response():
     bridge = RecordingBridge()
     logging_obj = RecordingLogging()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     response = ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(
@@ -503,7 +489,7 @@ def test_run_rust_ocr_prepares_request_and_wraps_response():
 def test_run_rust_ocr_resolves_key_via_secret_manager_when_missing():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(api_key=None, timeout=None),
@@ -516,7 +502,7 @@ def test_run_rust_ocr_resolves_key_via_secret_manager_when_missing():
 def test_run_rust_ocr_prefers_explicit_key_over_resolver():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     def _resolver(name: str) -> str | None:
         raise AssertionError(f"resolver should not be called for {name}")
@@ -536,7 +522,7 @@ def test_run_rust_ocr_uses_provider_api_key_env_var():
     bridge = RecordingBridge()
     resolver_calls = []
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     def _resolver(name):
         resolver_calls.append(name)
@@ -559,7 +545,7 @@ def test_run_rust_ocr_uses_provider_api_key_env_var():
 def test_prepare_rust_ocr_call_forwards_vertex_routing_metadata():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(
@@ -586,7 +572,7 @@ def test_prepare_rust_ocr_call_forwards_vertex_routing_metadata():
 def test_prepare_rust_ocr_call_resolves_vertex_routing_metadata_from_secret_manager():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     def _resolver(name: str) -> str | None:
         return {
@@ -610,7 +596,7 @@ def test_prepare_rust_ocr_call_resolves_vertex_routing_metadata_from_secret_mana
 def test_prepare_rust_ocr_call_resolves_azure_ai_api_base_from_secret_manager():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(
@@ -628,7 +614,7 @@ def test_prepare_rust_ocr_call_resolves_azure_ai_api_base_from_secret_manager():
 def test_prepare_rust_ocr_call_resolves_document_intelligence_endpoint():
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(
@@ -649,7 +635,7 @@ def test_run_rust_ocr_runs_pre_call_logging():
     logging_obj = RecordingLogging()
     bridge = RecordingBridge()
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=bridge)
+    rust_bridge._OCR.override(bridge)
 
     ocr_main._run_rust_ocr(
         prepared_request=build_prepared_request(
@@ -737,7 +723,7 @@ def test_ocr_exception_type_uses_resolved_provider_context(
 
     monkeypatch.setattr(ocr_main.litellm, "exception_type", fake_exception_type)
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(ocr=RaisingBridge())
+    rust_bridge._OCR.override(RaisingBridge())
 
     with pytest.raises(CapturedException):
         litellm.ocr(model=MODEL, document=DOCUMENT, api_key="sk-test")
@@ -783,7 +769,7 @@ async def test_aocr_exception_type_uses_resolved_provider_context(
 
     monkeypatch.setattr(ocr_main.litellm, "exception_type", fake_exception_type)
     litellm.rust(True)
-    rust_bridge.set_rust_ocr(aocr=RaisingAsyncBridge())
+    rust_bridge._AOCR.override(RaisingAsyncBridge())
 
     with pytest.raises(CapturedException):
         await litellm.aocr(model=MODEL, document=DOCUMENT, api_key="sk-test")
@@ -812,9 +798,7 @@ def test_ocr_does_not_route_to_rust_when_disabled():
     """With the flag off, the bridge must not be consulted even if an impl exists."""
     bridge = RecordingBridge()
     litellm.rust(False)
-    rust_bridge.set_rust_ocr(ocr=bridge)
-
-    assert rust_bridge.rust_ocr_enabled() is False
+    rust_bridge._OCR.override(bridge)
     # The impl stays available for injection, but the disabled flag gates usage,
     # so ocr() never reaches the Rust path (asserted via the enabled-path test).
     assert bridge.calls == []
diff --git a/tests/test_litellm/responses/test_rust_bridge_websocket.py b/tests/test_litellm/responses/test_rust_bridge_websocket.py
index 4b446368dbe..74d96bda336 100644
--- a/tests/test_litellm/responses/test_rust_bridge_websocket.py
+++ b/tests/test_litellm/responses/test_rust_bridge_websocket.py
@@ -4,7 +4,6 @@ import pytest
 
 from litellm.llms.custom_httpx.llm_http_handler import _rust_responses_websocket_enabled
 from litellm.rust_bridge import configuration, responses_websocket
-from litellm.types.router import GenericLiteLLMParams
 
 
 class _FakeNativeConnection:
@@ -48,22 +47,12 @@ def reset_responses_websocket():
     configuration.reset_rust_configuration()
 
 
-def test_rust_websocket_bridge_is_disabled_without_flag() -> None:
-    assert not _rust_responses_websocket_enabled("openai", GenericLiteLLMParams())
-    assert not _rust_responses_websocket_enabled("anthropic", GenericLiteLLMParams(rust=True))
-    assert _rust_responses_websocket_enabled("openai", GenericLiteLLMParams(rust=True))
-
-
-def test_explicit_false_overrides_process_enable() -> None:
+def test_rust_websocket_bridge_uses_process_enablement() -> None:
+    configuration.rust(False)
+    assert not _rust_responses_websocket_enabled("openai")
     configuration.rust(True)
-
-    assert not _rust_responses_websocket_enabled("openai", GenericLiteLLMParams(rust=False))
-
-
-def test_process_enable_applies_without_request_override() -> None:
-    configuration.rust(True)
-
-    assert _rust_responses_websocket_enabled("openai", GenericLiteLLMParams())
+    assert _rust_responses_websocket_enabled("openai")
+    assert not _rust_responses_websocket_enabled("anthropic")
 
 
 @pytest.mark.asyncio
diff --git a/tests/test_litellm/rust_bridge/test_chat_completions.py b/tests/test_litellm/rust_bridge/test_chat_completions.py
index 0489f4ff017..b2fd2e6dcc0 100644
--- a/tests/test_litellm/rust_bridge/test_chat_completions.py
+++ b/tests/test_litellm/rust_bridge/test_chat_completions.py
@@ -10,8 +10,8 @@ from __future__ import annotations
 import pytest
 
 import litellm
-from litellm.rust_bridge import chat_completions as bridge
 from litellm.rust_bridge import configuration
+from litellm.rust_bridge import chat_completions as bridge
 from litellm.types.utils import ModelResponse
 
 RUST_RESPONSE = {
@@ -67,10 +67,11 @@ def _hide_native_bridge(monkeypatch):
 
 
 @pytest.fixture(autouse=True)
-def reset_bridge():
+def reset_bridge(monkeypatch):
     """Every test starts with no injected callables, and leaves none behind."""
     bridge.set_rust_chat_completions(chat_completions=None, achat_completions=None, decline=None)
     configuration.reset_rust_configuration()
+    monkeypatch.setenv("LITELLM_RUST", "1")
     yield
     bridge.set_rust_chat_completions(chat_completions=None, achat_completions=None, decline=None)
     configuration.reset_rust_configuration()
@@ -112,7 +113,7 @@ def _accepts(**overrides) -> bool:
         "messages": MESSAGES,
         "optional_params": {"max_tokens": 16},
         "custom_llm_provider": "anthropic",
-        "litellm_params": {"rust": True},
+        "litellm_params": {},
         "stream": None,
     }
     kwargs.update(overrides)
@@ -126,23 +127,16 @@ class TestGate:
         bridge.set_rust_chat_completions(decline=gate)
         assert _accepts(litellm_params={}) is False
         assert _accepts(litellm_params=None) is False
-        assert _accepts(litellm_params={"rust": False}) is False
         assert gate.calls == [], "the gate must not be consulted before opt-in"
 
     def test_accepts_when_the_deployment_opted_in_and_the_core_agrees(self, monkeypatch):
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         gate = _RecordingDecline()
         bridge.set_rust_chat_completions(decline=gate)
         assert _accepts() is True
         assert gate.calls[0]["model"] == "claude-sonnet-4-5"
         assert gate.calls[0]["custom_llm_provider"] == "anthropic"
 
-    def test_explicit_false_overrides_process_enable(self):
-        bridge.set_rust_chat_completions(decline=_RecordingDecline())
-        configuration.rust(True)
-
-        assert _accepts(litellm_params={"rust": False}) is False
-
     def test_process_enable_applies_without_request_override(self):
         bridge.set_rust_chat_completions(decline=_RecordingDecline())
         configuration.rust(True)
@@ -155,7 +149,7 @@ class TestGate:
         assert _accepts(litellm_params={}) is True
 
     def test_declines_streaming_and_providers_off_the_path(self, monkeypatch):
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         gate = _RecordingDecline()
         bridge.set_rust_chat_completions(decline=gate)
         assert _accepts(stream=True) is False
@@ -170,10 +164,10 @@ class TestGate:
         handed `optional_params` only, so accepting here would send the request
         to Anthropic with the abuse-detection attribution silently missing.
         """
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         gate = _RecordingDecline()
         bridge.set_rust_chat_completions(decline=gate)
-        assert _accepts(litellm_params={"rust": True, "metadata": {"user_id": "u-123"}}) is False
+        assert _accepts(litellm_params={"metadata": {"user_id": "u-123"}}) is False
         assert gate.calls == [], "the core must not be consulted for a request it cannot see the key of"
 
         # Bedrock's Converse transform reads no `user_id`, and an Anthropic request
@@ -182,13 +176,13 @@ class TestGate:
             _accepts(
                 custom_llm_provider="bedrock",
                 model="bedrock/us-east-1/anthropic.claude-v2",
-                litellm_params={"rust": True, "metadata": {"user_id": "u-123"}},
+                litellm_params={"metadata": {"user_id": "u-123"}},
             )
             is True
         )
-        assert _accepts(litellm_params={"rust": True, "metadata": {"trace_id": "t-1"}}) is True
-        assert _accepts(litellm_params={"rust": True, "metadata": {"user_id": None}}) is True
-        assert _accepts(litellm_params={"rust": True, "metadata": None}) is True
+        assert _accepts(litellm_params={"metadata": {"trace_id": "t-1"}}) is True
+        assert _accepts(litellm_params={"metadata": {"user_id": None}}) is True
+        assert _accepts(litellm_params={"metadata": None}) is True
 
     def test_declines_a_bedrock_request_while_the_proxy_owns_request_metadata(self, monkeypatch):
         """`AmazonConverseConfig` resolves proxy-owned `requestMetadata` onto the
@@ -196,7 +190,7 @@ class TestGate:
         evicting a caller-supplied one. The core can do neither, so an operator
         who armed `bedrock_request_metadata_fields` keeps the Python path.
         """
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         gate = _RecordingDecline()
         bridge.set_rust_chat_completions(decline=gate)
         bedrock = {
@@ -213,17 +207,17 @@ class TestGate:
         assert _accepts(**bedrock) is True, "the decline follows the operator's opt-in alone"
 
     def test_declines_when_the_core_declines(self, monkeypatch):
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         bridge.set_rust_chat_completions(decline=_RecordingDecline("streaming"))
         assert _accepts() is False
 
     def test_declines_when_the_bridge_is_unavailable(self, monkeypatch):
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
         _hide_native_bridge(monkeypatch)
         assert _accepts() is False
 
     def test_declines_when_the_gate_itself_raises(self, monkeypatch):
-        monkeypatch.delenv("LITELLM_RUST", raising=False)
+        monkeypatch.setenv("LITELLM_RUST", "1")
 
         def exploding(**_kwargs):
             raise RuntimeError("boom")
diff --git a/tests/test_litellm/rust_bridge/test_configuration.py b/tests/test_litellm/rust_bridge/test_configuration.py
index 15f69f95335..aff9d5acac1 100644
--- a/tests/test_litellm/rust_bridge/test_configuration.py
+++ b/tests/test_litellm/rust_bridge/test_configuration.py
@@ -10,7 +10,6 @@ from typing import Final
 import pytest
 
 from litellm.rust_bridge import configuration
-from litellm.rust_bridge import ocr as rust_ocr
 
 
 @pytest.fixture(autouse=True)
@@ -19,42 +18,31 @@ def _isolated_configuration(  # pyright: ignore[reportUnusedFunction]  # pytest
 ) -> Generator[None]:
     configuration.reset_rust_configuration()
     monkeypatch.delenv("LITELLM_RUST", raising=False)
-    monkeypatch.delenv("LITELLM_USE_RUST_OCR", raising=False)
-    rust_ocr.set_rust_ocr(ocr=None, aocr=None)
     yield
     configuration.reset_rust_configuration()
-    rust_ocr.set_rust_ocr(ocr=None, aocr=None)
 
 
 @pytest.mark.parametrize(
-    ("request_override", "process", "environment", "legacy_environment", "release_default", "expected"),
+    ("process", "environment", "release_default", "expected"),
     (
-        (False, True, True, True, True, False),
-        (True, False, False, False, False, True),
-        (None, False, True, True, True, False),
-        (None, True, False, False, False, True),
-        (None, None, False, True, True, False),
-        (None, None, True, False, False, True),
-        (None, None, None, False, True, False),
-        (None, None, None, True, False, True),
-        (None, None, None, None, False, False),
-        (None, None, None, None, True, True),
+        (False, True, True, False),
+        (True, False, False, True),
+        (None, False, True, False),
+        (None, True, False, True),
+        (None, None, False, False),
+        (None, None, True, True),
     ),
 )
 def test_resolution_precedence(
-    request_override: bool | None,
     process: bool | None,
     environment: bool | None,
-    legacy_environment: bool | None,
     release_default: bool,
     expected: bool,
 ) -> None:
     assert (
         configuration.resolve_rust_enabled(
-            request_override=request_override,
             process_override=process,
             environment_override=environment,
-            legacy_environment_override=legacy_environment,
             release_default=release_default,
         )
         is expected
@@ -71,7 +59,6 @@ def test_process_override_wins_over_environment(monkeypatch: pytest.MonkeyPatch)
     configuration.rust(True)
 
     assert configuration.rust_enabled() is True
-    assert configuration.rust_enabled(request_override=False) is False
 
 
 def test_global_environment_accepts_explicit_false(monkeypatch: pytest.MonkeyPatch) -> None:
@@ -83,18 +70,8 @@ def test_global_environment_accepts_explicit_false(monkeypatch: pytest.MonkeyPat
 @pytest.mark.parametrize("value", ("", " ", "sometimes", "2"))
 def test_invalid_environment_value_disables_rust(monkeypatch: pytest.MonkeyPatch, value: str) -> None:
     monkeypatch.setenv("LITELLM_RUST", value)
-    monkeypatch.setenv("LITELLM_USE_RUST_OCR", "1")
 
     assert configuration.rust_enabled() is False
-    assert configuration.rust_ocr_enabled() is False
-
-
-@pytest.mark.parametrize("value", ("", " ", "sometimes", "2"))
-def test_invalid_legacy_environment_value_disables_rust(monkeypatch: pytest.MonkeyPatch, value: str) -> None:
-    monkeypatch.setenv("LITELLM_USE_RUST_OCR", value)
-
-    with pytest.warns(DeprecationWarning, match="LITELLM_USE_RUST_OCR is deprecated"):
-        assert configuration.rust_enabled() is False
 
 
 def test_process_override_and_reset_apply_to_existing_threads(monkeypatch: pytest.MonkeyPatch) -> None:
@@ -104,40 +81,20 @@ def test_process_override_and_reset_apply_to_existing_threads(monkeypatch: pytes
         assert executor.submit(configuration.rust_enabled).result() is True
         configuration.rust(False)
         assert executor.submit(configuration.rust_enabled).result() is False
-        assert executor.submit(configuration.rust_ocr_enabled).result() is False
         configuration.reset_rust_configuration()
         assert executor.submit(configuration.rust_enabled).result() is True
-        assert executor.submit(configuration.rust_ocr_enabled).result() is True
 
 
 def test_explicit_override_precedes_invalid_environment(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("LITELLM_RUST", "sometimes")
 
-    assert configuration.rust_enabled(request_override=False) is False
     configuration.rust(True)
     assert configuration.rust_enabled() is True
 
 
-def test_legacy_ocr_environment_is_deprecated_and_global(monkeypatch: pytest.MonkeyPatch) -> None:
-    monkeypatch.setenv("LITELLM_USE_RUST_OCR", "1")
-
-    with pytest.warns(DeprecationWarning, match="LITELLM_USE_RUST_OCR is deprecated"):
-        assert configuration.rust_enabled() is True
-    with pytest.warns(DeprecationWarning, match="LITELLM_USE_RUST_OCR is deprecated"):
-        assert configuration.rust_ocr_enabled() is True
-
-
-def test_global_environment_precedes_legacy_ocr_environment(monkeypatch: pytest.MonkeyPatch) -> None:
-    monkeypatch.setenv("LITELLM_RUST", "0")
-    monkeypatch.setenv("LITELLM_USE_RUST_OCR", "1")
-
-    assert configuration.rust_enabled() is False
-
-
-@pytest.mark.parametrize("environment_name", ("LITELLM_RUST", "LITELLM_USE_RUST_OCR"))
 @pytest.mark.parametrize(("value", "expected"), (("1", "True"), ("0", "False")))
-def test_environment_controls_startup(environment_name: str, value: str, expected: str) -> None:
-    environment: Final = {**os.environ, environment_name: value}
+def test_environment_controls_startup(value: str, expected: str) -> None:
+    environment: Final = {**os.environ, "LITELLM_RUST": value}
     result: Final = subprocess.run(
         (
             sys.executable,
diff --git a/tests/test_litellm/test_audio_transcription_rust_bridge.py b/tests/test_litellm/test_audio_transcription_rust_bridge.py
index bbeb6c38f78..112464bda22 100644
--- a/tests/test_litellm/test_audio_transcription_rust_bridge.py
+++ b/tests/test_litellm/test_audio_transcription_rust_bridge.py
@@ -44,7 +44,7 @@ class AsyncBridge:
 
 def test_enabled_sync_bridge_receives_audio() -> None:
     bridge = SyncBridge()
-    rust_bridge.configure_rust_transcription(True, transcription=bridge)
+    rust_bridge.configure_rust_transcription(transcription=bridge)
     result = rust_bridge.transcription(
         model="mistral.voxtral-mini-3b-2507",
         audio={"data": "AQI=", "format": "wav", "filename": "audio.wav"},
@@ -61,7 +61,7 @@ def test_enabled_sync_bridge_receives_audio() -> None:
 
 @pytest.mark.asyncio
 async def test_enabled_async_bridge() -> None:
-    rust_bridge.configure_rust_transcription(True, atranscription=AsyncBridge())
+    rust_bridge.configure_rust_transcription(atranscription=AsyncBridge())
     result = await rust_bridge.atranscription(
         model="mistral.voxtral-mini-3b-2507",
         audio={"data": "AQI=", "format": "wav", "filename": "audio.wav"},
diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py
index d686b032ee0..271c84384b4 100644
--- a/tests/test_litellm/test_utils.py
+++ b/tests/test_litellm/test_utils.py
@@ -5239,52 +5239,6 @@ def test_client_side_timeout_marker_never_reaches_the_provider():
     )
 
 
-def test_rust_flag_not_forwarded_as_provider_param():
-    forwarded = get_non_default_completion_params({"rust": True, "temperature": 0.5})
-    assert "rust" not in forwarded
-
-
-def test_completion_does_not_leak_rust_flag_into_provider_request_body():
-    mock_response = MagicMock()
-    mock_response.model_dump.return_value = {
-        "id": "chatcmpl-1",
-        "object": "chat.completion",
-        "created": 1234567890,
-        "model": "gpt-4o-mini",
-        "choices": [
-            {
-                "index": 0,
-                "message": {"role": "assistant", "content": "hi"},
-                "finish_reason": "stop",
-            }
-        ],
-        "usage": {
-            "prompt_tokens": 1,
-            "completion_tokens": 1,
-            "total_tokens": 2,
-        },
-    }
-
-    mock_raw_response = MagicMock()
-    mock_raw_response.headers = {}
-    mock_raw_response.parse.return_value = mock_response
-
-    mock_client = MagicMock()
-    mock_client.chat.completions.with_raw_response.create.return_value = mock_raw_response
-
-    litellm.completion(
-        model="openai/gpt-4o-mini",
-        messages=[{"role": "user", "content": "hi"}],
-        rust=True,
-        api_key="sk-test",
-        client=mock_client,
-    )
-
-    create_kwargs = mock_client.chat.completions.with_raw_response.create.call_args.kwargs
-    assert "rust" not in create_kwargs
-    assert "rust" not in (create_kwargs.get("extra_body") or {})
-
-
 class _RecordingDeploymentFailureLogger(CustomLogger):
     def __init__(self) -> None:
         super().__init__()
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index b1534c19670..108e50b52a1 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -29562,8 +29562,6 @@ export interface components {
             regional_processing_uplift_multiplier_us?: number | null;
             /** Rpm */
             rpm?: number | null;
-            /** Rust */
-            rust?: boolean | null;
             /** S3 Bucket Name */
             s3_bucket_name?: string | null;
             /** S3 Encryption Key Id */
@@ -39729,8 +39727,6 @@ export interface components {
             regional_processing_uplift_multiplier_us?: number | null;
             /** Rpm */
             rpm?: number | null;
-            /** Rust */
-            rust?: boolean | null;
             /** S3 Bucket Name */
             s3_bucket_name?: string | null;
             /** S3 Encryption Key Id */

From 8da735410c7903b0e8c92033e444a76d32a03c6f Mon Sep 17 00:00:00 2001
From: yujonglee <yujong@berri.ai>
Date: Mon, 7 Sep 2026 10:44:22 -0700
Subject: [PATCH 68/79] test(ocr): trace callback lifecycle parity (#40063)

---
 .../crates/ai-gateway/src/ocr/hooks.rs        | 12 +++
 .../crates/ai-gateway/tests/ocr_lifecycle.rs  | 52 +++++++++--
 .../python-bridge/src/function_trace.rs       | 28 ++++--
 .../python-bridge/src/routes/definition.rs    |  9 +-
 .../shared/tracing/native.py                  |  3 +-
 .../strategies/trace_parity/models.py         |  1 +
 .../strategies/trace_parity/sdk/execution.py  | 74 +++++++++++----
 .../strategies/trace_parity/sdk/ocr/case.py   | 91 +++++++++++++++++--
 8 files changed, 220 insertions(+), 50 deletions(-)

diff --git a/litellm-rust/crates/ai-gateway/src/ocr/hooks.rs b/litellm-rust/crates/ai-gateway/src/ocr/hooks.rs
index 446b323db3a..ed41f1ff9e7 100644
--- a/litellm-rust/crates/ai-gateway/src/ocr/hooks.rs
+++ b/litellm-rust/crates/ai-gateway/src/ocr/hooks.rs
@@ -265,6 +265,12 @@ impl CallLifecycleHooks<PreparedOcrRequest, PreparedOcrRequest, Value> for OcrLi
         Box::pin(async move { Ok(request) })
     }
 
+    #[tracing::instrument(
+        name = "success_callback",
+        target = "litellm::function_trace",
+        level = "trace",
+        skip_all
+    )]
     fn async_log_success_event<'a>(
         &'a self,
         context: &'a CallLifecycleContext,
@@ -288,6 +294,12 @@ impl CallLifecycleHooks<PreparedOcrRequest, PreparedOcrRequest, Value> for OcrLi
         })
     }
 
+    #[tracing::instrument(
+        name = "failure_callback",
+        target = "litellm::function_trace",
+        level = "trace",
+        skip_all
+    )]
     fn async_log_failure_event<'a>(
         &'a self,
         context: &'a CallLifecycleContext,
diff --git a/litellm-rust/crates/ai-gateway/tests/ocr_lifecycle.rs b/litellm-rust/crates/ai-gateway/tests/ocr_lifecycle.rs
index c3a89f4394d..60e90ed2a7c 100644
--- a/litellm-rust/crates/ai-gateway/tests/ocr_lifecycle.rs
+++ b/litellm-rust/crates/ai-gateway/tests/ocr_lifecycle.rs
@@ -11,9 +11,13 @@ use litellm_ai_gateway::integrations::custom_logger::{
 use litellm_ai_gateway::integrations::types::RequestMetadata;
 use litellm_ai_gateway::ocr::{OcrRequest, ocr};
 use litellm_core::error::Error;
+#[cfg(feature = "trace-parity")]
+use litellm_core::observability::FunctionTrace;
 use serde_json::{Map, Value, json};
 use tokio::io::{AsyncReadExt, AsyncWriteExt};
 use tokio::net::{TcpListener, TcpStream};
+#[cfg(feature = "trace-parity")]
+use tracing::instrument::WithSubscriber;
 
 async fn read_http_headers(socket: &mut TcpStream) -> String {
     let mut request = Vec::new();
@@ -320,14 +324,17 @@ async fn ocr_lifecycle_runs_pre_during_and_success_hooks() {
         GuardrailEventHook::PreCall,
         GuardrailEventHook::DuringCall,
     ]));
-    let response = ocr(OcrRequest {
+    #[cfg(feature = "trace-parity")]
+    let trace = FunctionTrace::default();
+    let api_base = format!("http://{addr}");
+    let call = ocr(OcrRequest {
         model: "mistral-ocr-latest",
         document: json!({
             "type": "document_url",
             "document_url": "https://example.com/doc.pdf"
         }),
         api_key: Some("sk-test"),
-        api_base: Some(&format!("http://{addr}")),
+        api_base: Some(&api_base),
         custom_llm_provider: Some("mistral"),
         extra_headers: None,
         optional_params: Map::new(),
@@ -339,9 +346,10 @@ async fn ocr_lifecycle_runs_pre_during_and_success_hooks() {
             ..Default::default()
         },
         litellm_call_id: Some("ocr-call-1"),
-    })
-    .await
-    .expect("ocr request succeeds");
+    });
+    #[cfg(feature = "trace-parity")]
+    let call = call.with_subscriber(trace.dispatcher());
+    let response = call.await.expect("ocr request succeeds");
 
     assert_eq!(response["pages"][0]["markdown"], "ok");
     assert_eq!(
@@ -359,6 +367,16 @@ async fn ocr_lifecycle_runs_pre_during_and_success_hooks() {
             error_kind: None,
         }]
     );
+    #[cfg(feature = "trace-parity")]
+    assert_eq!(
+        trace
+            .events()
+            .iter()
+            .filter(|event| event.function.ends_with("_callback"))
+            .map(|event| event.function)
+            .collect::<Vec<_>>(),
+        vec!["success_callback"]
+    );
 
     let request = server.await.expect("server task completes");
     assert!(request.contains(r#""guarded_pre":true"#), "{request}");
@@ -388,14 +406,17 @@ async fn ocr_lifecycle_runs_failure_hook_on_provider_error() {
     });
 
     let logger = Arc::new(RecordingOcrLogger::default());
-    let err = ocr(OcrRequest {
+    #[cfg(feature = "trace-parity")]
+    let trace = FunctionTrace::default();
+    let api_base = format!("http://{addr}");
+    let call = ocr(OcrRequest {
         model: "mistral-ocr-latest",
         document: json!({
             "type": "document_url",
             "document_url": "https://example.com/doc.pdf"
         }),
         api_key: Some("sk-test"),
-        api_base: Some(&format!("http://{addr}")),
+        api_base: Some(&api_base),
         custom_llm_provider: Some("mistral"),
         extra_headers: None,
         optional_params: Map::new(),
@@ -404,9 +425,10 @@ async fn ocr_lifecycle_runs_failure_hook_on_provider_error() {
         guardrails: Vec::new(),
         request_metadata: RequestMetadata::default(),
         litellm_call_id: Some("ocr-call-2"),
-    })
-    .await
-    .expect_err("provider error propagates");
+    });
+    #[cfg(feature = "trace-parity")]
+    let call = call.with_subscriber(trace.dispatcher());
+    let err = call.await.expect_err("provider error propagates");
 
     assert!(matches!(err, Error::Http { status: 500, .. }));
     server.await.expect("server task completes");
@@ -421,6 +443,16 @@ async fn ocr_lifecycle_runs_failure_hook_on_provider_error() {
             error_kind: Some("HttpError".to_string()),
         }]
     );
+    #[cfg(feature = "trace-parity")]
+    assert_eq!(
+        trace
+            .events()
+            .iter()
+            .filter(|event| event.function.ends_with("_callback"))
+            .map(|event| event.function)
+            .collect::<Vec<_>>(),
+        vec!["failure_callback"]
+    );
 }
 
 #[tokio::test]
diff --git a/litellm-rust/crates/python-bridge/src/function_trace.rs b/litellm-rust/crates/python-bridge/src/function_trace.rs
index ea7d9f4993e..bc3c962f7a3 100644
--- a/litellm-rust/crates/python-bridge/src/function_trace.rs
+++ b/litellm-rust/crates/python-bridge/src/function_trace.rs
@@ -1,3 +1,4 @@
+use std::fmt::Display;
 use std::future::Future;
 
 use litellm_core::observability::{FunctionTrace, FunctionTraceEvent};
@@ -6,17 +7,32 @@ use tracing::instrument::WithSubscriber;
 
 #[derive(Serialize)]
 pub(crate) struct TracedResponse<T> {
-    response: T,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    response: Option<T>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    error: Option<String>,
     trace: Vec<FunctionTraceEvent>,
 }
 
 pub(crate) async fn capture<T, E>(
     future: impl Future<Output = Result<T, E>>,
-) -> Result<TracedResponse<T>, E> {
+) -> Result<TracedResponse<T>, E>
+where
+    E: Display,
+{
     let trace = FunctionTrace::default();
-    let response = future.with_subscriber(trace.dispatcher()).await?;
-    Ok(TracedResponse {
-        response,
-        trace: trace.events(),
+    let result = future.with_subscriber(trace.dispatcher()).await;
+    let events = trace.events();
+    Ok(match result {
+        Ok(response) => TracedResponse {
+            response: Some(response),
+            error: None,
+            trace: events,
+        },
+        Err(error) => TracedResponse {
+            response: None,
+            error: Some(error.to_string()),
+            trace: events,
+        },
     })
 }
diff --git a/litellm-rust/crates/python-bridge/src/routes/definition.rs b/litellm-rust/crates/python-bridge/src/routes/definition.rs
index 3285da14d5f..bc51647cbad 100644
--- a/litellm-rust/crates/python-bridge/src/routes/definition.rs
+++ b/litellm-rust/crates/python-bridge/src/routes/definition.rs
@@ -486,10 +486,11 @@ asyncio.run(exercise())
             let code = CString::new(
                 r#"
 result = routes.echo("traced")
-assert result == {
-    "response": "traced",
-    "trace": [{"function": "execute_echo", "depth": 0}],
-}
+assert result["response"] == "traced", result
+assert [event["function"] for event in result["trace"]] == ["execute_echo"], result
+failure = routes.echo("error")
+assert failure["error"] == "invalid request: synthetic error", failure
+assert [event["function"] for event in failure["trace"]] == ["execute_echo"], failure
 "#,
             )
             .expect("Python source should not contain null bytes");
diff --git a/tests/rust-python-harness/shared/tracing/native.py b/tests/rust-python-harness/shared/tracing/native.py
index 4f988f65294..688995cbc4b 100644
--- a/tests/rust-python-harness/shared/tracing/native.py
+++ b/tests/rust-python-harness/shared/tracing/native.py
@@ -19,7 +19,8 @@ class _TraceEventPayload(BaseModel):
 
 class TraceResponsePayload(BaseModel):
     model_config = ConfigDict(strict=True, extra="forbid")
-    response: object
+    response: object = None
+    error: str | None = None
     trace: tuple[_TraceEventPayload, ...] | list[_TraceEventPayload]
 
 
diff --git a/tests/rust-python-harness/strategies/trace_parity/models.py b/tests/rust-python-harness/strategies/trace_parity/models.py
index da560f99730..04659b25382 100644
--- a/tests/rust-python-harness/strategies/trace_parity/models.py
+++ b/tests/rust-python-harness/strategies/trace_parity/models.py
@@ -16,6 +16,7 @@ TraceFailureSource = Literal["python", "rust", "harness"]
 class RouteFixture:
     kwargs: dict[str, object]
     provider_responses: tuple[RecordedHttpResponse, ...]
+    expected_failure: bool = False
 
 
 @dataclass(frozen=True, slots=True)
diff --git a/tests/rust-python-harness/strategies/trace_parity/sdk/execution.py b/tests/rust-python-harness/strategies/trace_parity/sdk/execution.py
index f8d7c55d4e2..eb6c9233565 100644
--- a/tests/rust-python-harness/strategies/trace_parity/sdk/execution.py
+++ b/tests/rust-python-harness/strategies/trace_parity/sdk/execution.py
@@ -2,15 +2,16 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import Awaitable
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Final, Protocol, cast
 
 from ....shared.parity.replay import replay_server
 from ....shared.reporting.models import Surface
-from ....shared.tracing.native import native_trace_events
+from ....shared.tracing.native import TraceResponsePayload, native_trace_events
 from ....shared.tracing.profiler import FunctionTraceEvent, profile_python
 from ....shared.tracing.steps import Engine, pipeline_projection
-from ..models import RouteSpec, TraceExecutionFailure, TraceMode, TraceScenario
+from ..models import RouteFixture, RouteSpec, TraceExecutionFailure, TraceMode, TraceScenario
 from ..reporting import TraceComparisonArtifact
 
 
@@ -18,9 +19,22 @@ class SdkCall(Protocol):
     def __call__(self, **kwargs: object) -> object: ...
 
 
+@dataclass(frozen=True, slots=True)
+class _CollectedTrace:
+    events: tuple[FunctionTraceEvent, ...]
+    error: str | None = None
+
+
 def _invoke(function: SdkCall, kwargs: dict[str, object], *, asynchronous: bool) -> object:
     async def invoke_async() -> object:
-        return await cast(Awaitable[object], function(**kwargs))
+        try:
+            return await cast(Awaitable[object], function(**kwargs))
+        finally:
+            await asyncio.sleep(0)
+            from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
+
+            await asyncio.wait_for(GLOBAL_LOGGING_WORKER.flush(), timeout=10)
+            await GLOBAL_LOGGING_WORKER.stop()
 
     if asynchronous:
         return asyncio.run(invoke_async())
@@ -48,16 +62,30 @@ def _entrypoint(spec: RouteSpec, engine: Engine, *, asynchronous: bool) -> SdkCa
     return cast(SdkCall, getattr(owner, spec.python_entrypoints[int(asynchronous)]))
 
 
+def _python_invocation_error(function: SdkCall, kwargs: dict[str, object], *, asynchronous: bool) -> str | None:
+    try:
+        _invoke(function, kwargs, asynchronous=asynchronous)
+    except Exception as error:
+        return f"{type(error).__name__}: {error}"
+    return None
+
+
 def _collect(
-    function: SdkCall, kwargs: dict[str, object], engine: Engine, *, asynchronous: bool
-) -> tuple[FunctionTraceEvent, ...]:
+    function: SdkCall,
+    fixture: RouteFixture,
+    engine: Engine,
+    *,
+    asynchronous: bool,
+) -> _CollectedTrace:
+    kwargs: Final = fixture.kwargs
     if engine == "rust":
-        return native_trace_events(_invoke(function, kwargs, asynchronous=asynchronous))
+        payload: Final = TraceResponsePayload.model_validate(_invoke(function, kwargs, asynchronous=asynchronous))
+        return _CollectedTrace(native_trace_events(payload), payload.error)
     import litellm
 
     with profile_python(Path(litellm.__file__).parent, threads=True) as profiler:
-        _invoke(function, kwargs, asynchronous=asynchronous)
-    return tuple(profiler.events)
+        error: Final = _python_invocation_error(function, kwargs, asynchronous=asynchronous)
+    return _CollectedTrace(tuple(profiler.events), error)
 
 
 def collect_trace(
@@ -68,22 +96,30 @@ def collect_trace(
         return function
     try:
         with replay_server() as provider:
-            fixture: Final = spec.fixture(engine, provider.url)
-            for response in fixture.provider_responses:
+            base_fixture: Final = spec.fixture(engine, provider.url)
+            for response in base_fixture.provider_responses:
                 provider.enqueue_response(response)
-            kwargs: Final = {
-                **fixture.kwargs,
-                "api_key": "test-key",
-                "api_base": provider.url,
-                **({"timeout_seconds": 5} if engine == "rust" else {"timeout": 5}),
-            }
-            events: Final = _collect(function, kwargs, engine, asynchronous=asynchronous)
+            fixture: Final = RouteFixture(
+                kwargs={
+                    **base_fixture.kwargs,
+                    "api_key": "test-key",
+                    "api_base": provider.url,
+                    **({"timeout_seconds": 5} if engine == "rust" else {"timeout": 5}),
+                },
+                provider_responses=base_fixture.provider_responses,
+                expected_failure=base_fixture.expected_failure,
+            )
+            collected: Final = _collect(function, fixture, engine, asynchronous=asynchronous)
             provider.take_requests(len(fixture.provider_responses))
     except Exception as error:
         return TraceExecutionFailure(engine, f"{type(error).__name__}: {error}")
-    if not events:
+    if fixture.expected_failure and collected.error is None:
+        return TraceExecutionFailure(engine, "call succeeded but the scenario expects failure")
+    if not fixture.expected_failure and collected.error is not None:
+        return TraceExecutionFailure(engine, collected.error)
+    if not collected.events:
         return TraceExecutionFailure(engine, "trace is empty")
-    return events
+    return collected.events
 
 
 def _failure_message(result: tuple[FunctionTraceEvent, ...] | TraceExecutionFailure) -> str | None:
diff --git a/tests/rust-python-harness/strategies/trace_parity/sdk/ocr/case.py b/tests/rust-python-harness/strategies/trace_parity/sdk/ocr/case.py
index fe214f45339..2a4a1b3a152 100644
--- a/tests/rust-python-harness/strategies/trace_parity/sdk/ocr/case.py
+++ b/tests/rust-python-harness/strategies/trace_parity/sdk/ocr/case.py
@@ -19,6 +19,20 @@ COMMON_MAPPINGS: Final = (
     mapping(rust_span="http_request", python_frame=r"AsyncHTTPHandler\.post$|HTTPHandler\.post$"),
 )
 
+SUCCESS_CALLBACK_SYNC_MAPPING: Final = mapping(
+    rust_span="success_callback",
+    python_frame=r"BoundedLoggingThreadPoolExecutor\.submit$",
+)
+SUCCESS_CALLBACK_ASYNC_MAPPING: Final = mapping(
+    rust_span="success_callback",
+    python_frame=r"Logging\.async_success_handler$",
+)
+FAILURE_CALLBACK_MAPPING: Final = mapping(
+    rust_span="failure_callback",
+    python_frame=r"Logging\.(?:async_)?failure_handler$",
+)
+IGNORED_SUCCESS_CALLBACK_MAPPING: Final = mapping(rust_span="success_callback")
+
 SYNC_MAPPINGS: Final = (
     *COMMON_MAPPINGS,
     mapping(rust_span="execute_ocr_provider_call", python_frame=r"BaseLLMHTTPHandler\.ocr$"),
@@ -39,6 +53,21 @@ ASYNC_MAPPINGS: Final = (
     ),
 )
 
+CALLBACK_SUCCESS_SYNC_MAPPINGS: Final = (*SYNC_MAPPINGS, SUCCESS_CALLBACK_SYNC_MAPPING)
+CALLBACK_SUCCESS_ASYNC_MAPPINGS: Final = (*ASYNC_MAPPINGS, SUCCESS_CALLBACK_ASYNC_MAPPING)
+CALLBACK_FAILURE_SYNC_MAPPINGS: Final = (
+    *COMMON_MAPPINGS,
+    mapping(rust_span="execute_ocr_provider_call", python_frame=r"BaseLLMHTTPHandler\.ocr$"),
+    FAILURE_CALLBACK_MAPPING,
+)
+CALLBACK_FAILURE_ASYNC_MAPPINGS: Final = (
+    *COMMON_MAPPINGS,
+    mapping(span="python_ocr_wrapper", python_frame=r"BaseLLMHTTPHandler\.ocr$"),
+    mapping(rust_span="execute_ocr_provider_call", python_frame=r"BaseLLMHTTPHandler\.async_ocr$"),
+    FAILURE_CALLBACK_MAPPING,
+)
+
+
 AZURE_COMMON_MAPPINGS: Final = (
     *COMMON_MAPPINGS[:7],
     mapping(
@@ -99,6 +128,34 @@ def _mistral_fixture(engine: Engine, _base_url: str) -> RouteFixture:
     return _fixture(engine, "mistral/mistral-ocr-latest")
 
 
+def _callback_fixture(engine: Engine, *, failure: bool) -> RouteFixture:
+    fixture: Final = _fixture(engine, "mistral/mistral-ocr-latest")
+    provider_responses: Final = (
+        (
+            RecordedHttpResponse.from_bytes(
+                400,
+                (HttpHeader(name="content-type", value="application/json"),),
+                b'{"message":"trace callback provider failure"}',
+            ),
+        )
+        if failure
+        else fixture.provider_responses
+    )
+    return RouteFixture(
+        kwargs=fixture.kwargs,
+        provider_responses=provider_responses,
+        expected_failure=failure,
+    )
+
+
+def _mistral_callback_success_fixture(engine: Engine, _base_url: str) -> RouteFixture:
+    return _callback_fixture(engine, failure=False)
+
+
+def _mistral_callback_failure_fixture(engine: Engine, _base_url: str) -> RouteFixture:
+    return _callback_fixture(engine, failure=True)
+
+
 def _azure_fixture(engine: Engine, _base_url: str) -> RouteFixture:
     return _fixture(
         engine,
@@ -304,36 +361,50 @@ TRACE_SUITE: Final = TraceSuite(
             name="mistral",
             fixture=_mistral_fixture,
             mappings=COMMON_MAPPINGS,
-            sync_mappings=SYNC_MAPPINGS,
-            async_mappings=ASYNC_MAPPINGS,
+            sync_mappings=(*SYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+            async_mappings=(*ASYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+        ),
+        TraceScenario(
+            name="mistral-callback-success",
+            fixture=_mistral_callback_success_fixture,
+            mappings=COMMON_MAPPINGS,
+            sync_mappings=CALLBACK_SUCCESS_SYNC_MAPPINGS,
+            async_mappings=CALLBACK_SUCCESS_ASYNC_MAPPINGS,
+        ),
+        TraceScenario(
+            name="mistral-callback-failure",
+            fixture=_mistral_callback_failure_fixture,
+            mappings=(*COMMON_MAPPINGS, FAILURE_CALLBACK_MAPPING),
+            sync_mappings=CALLBACK_FAILURE_SYNC_MAPPINGS,
+            async_mappings=CALLBACK_FAILURE_ASYNC_MAPPINGS,
         ),
         TraceScenario(
             name="azure-ai",
             fixture=_azure_fixture,
             mappings=AZURE_COMMON_MAPPINGS,
-            sync_mappings=AZURE_SYNC_MAPPINGS,
-            async_mappings=AZURE_ASYNC_MAPPINGS,
+            sync_mappings=(*AZURE_SYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+            async_mappings=(*AZURE_ASYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
         ),
         TraceScenario(
             name="azure-document-intelligence",
             fixture=_azure_document_intelligence_fixture,
             mappings=DOCUMENT_INTELLIGENCE_COMMON_MAPPINGS,
-            sync_mappings=DOCUMENT_INTELLIGENCE_SYNC_MAPPINGS,
-            async_mappings=DOCUMENT_INTELLIGENCE_ASYNC_MAPPINGS,
+            sync_mappings=(*DOCUMENT_INTELLIGENCE_SYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+            async_mappings=(*DOCUMENT_INTELLIGENCE_ASYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
         ),
         TraceScenario(
             name="vertex-ai",
             fixture=_vertex_fixture,
             mappings=VERTEX_COMMON_MAPPINGS,
-            sync_mappings=VERTEX_SYNC_MAPPINGS,
-            async_mappings=VERTEX_ASYNC_MAPPINGS,
+            sync_mappings=(*VERTEX_SYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+            async_mappings=(*VERTEX_ASYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
         ),
         TraceScenario(
             name="vertex-deepseek",
             fixture=_vertex_deepseek_fixture,
             mappings=DEEPSEEK_COMMON_MAPPINGS,
-            sync_mappings=DEEPSEEK_SYNC_MAPPINGS,
-            async_mappings=DEEPSEEK_ASYNC_MAPPINGS,
+            sync_mappings=(*DEEPSEEK_SYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
+            async_mappings=(*DEEPSEEK_ASYNC_MAPPINGS, IGNORED_SUCCESS_CALLBACK_MAPPING),
         ),
     ),
 )

From 2700ffe8c978229628767a45e2c83b7c1bb0297a Mon Sep 17 00:00:00 2001
From: yucheng-berri <yucheng@berri.ai>
Date: Mon, 7 Sep 2026 11:12:34 -0700
Subject: [PATCH 69/79] fix: capture provider request id in failure logging
 payloads (#40045)

* fix: capture provider request id in failure logging payloads

* fix: include common provider request id headers

* test: remove accidental formatter churn

* test: type provider request id cases

* test: cover azure and google request id headers

* refactor: share provider request id headers from constants

* chore: remove redundant request header comment

* test: cover provider header lookup failures
---
 litellm/constants.py                          |  10 ++
 litellm/litellm_core_utils/litellm_logging.py |  27 +++++
 litellm/types/utils.py                        |   1 +
 .../test_litellm_logging.py                   | 111 +++++++++++++++++-
 4 files changed, 148 insertions(+), 1 deletion(-)

diff --git a/litellm/constants.py b/litellm/constants.py
index ce744e9c58a..d53686e5e5b 100644
--- a/litellm/constants.py
+++ b/litellm/constants.py
@@ -1901,6 +1901,16 @@ HTTP_FRAMING_HEADERS: Final[frozenset[str]] = frozenset(
     }
 )
 
+PROVIDER_REQUEST_ID_HEADERS: Final[tuple[str, ...]] = (
+    "x-amzn-requestid",
+    "x-request-id",
+    "request-id",
+    "x-ms-request-id",
+    "apim-request-id",
+    "x-goog-request-id",
+    "cf-ray",
+)
+
 # Browser-facing security headers that a malicious or misconfigured upstream
 # provider must not be able to set on the proxy's own response.
 BROWSER_SECURITY_HEADERS: Final[frozenset[str]] = frozenset(
diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py
index 60abad17c03..ca2cca5360f 100644
--- a/litellm/litellm_core_utils/litellm_logging.py
+++ b/litellm/litellm_core_utils/litellm_logging.py
@@ -42,6 +42,7 @@ from litellm.caching.caching_handler import LLMCachingHandler
 from litellm.constants import (
     DEFAULT_MOCK_RESPONSE_COMPLETION_TOKEN_COUNT,
     DEFAULT_MOCK_RESPONSE_PROMPT_TOKEN_COUNT,
+    PROVIDER_REQUEST_ID_HEADERS,
     SENTRY_DENYLIST,
     SENTRY_PII_DENYLIST,
 )
@@ -255,6 +256,30 @@ _in_memory_loggers: Final[list[CustomLogger]] = []
 
 _STANDARD_LOGGING_METADATA_KEYS: Final[frozenset[str]] = frozenset(StandardLoggingMetadata.__annotations__.keys())
 
+
+def _get_provider_request_id(original_exception: Exception) -> str | None:
+    try:
+        error_response: Final = getattr(original_exception, "response", None)
+        header_sources: Final = (
+            _get_response_headers(original_exception),
+            getattr(error_response, "headers", None),
+            getattr(original_exception, "litellm_response_headers", None),
+        )
+        return next(
+            (
+                str(value)
+                for expected_header_name in PROVIDER_REQUEST_ID_HEADERS
+                for headers in header_sources
+                if isinstance(headers, Mapping)
+                for header_name, value in headers.items()
+                if isinstance(header_name, str) and header_name.lower() == expected_header_name and value
+            ),
+            None,
+        )
+    except Exception:
+        return None
+
+
 ### GLOBAL VARIABLES ###
 
 # Cache custom pricing keys as frozenset for O(1) lookups instead of looping through 49 keys
@@ -5664,6 +5689,7 @@ class StandardLoggingPayloadSetup:
         rate_limit_category: Final = validate_rate_limit_category(getattr(original_exception, "category", None))
         rate_limit_type: Final = validate_rate_limit_type(getattr(original_exception, "rate_limit_type", None))
         budget_error: Final = original_exception if isinstance(original_exception, BudgetExceededError) else None
+        provider_request_id: Final = _get_provider_request_id(original_exception) if original_exception else None
 
         return StandardLoggingPayloadErrorInformation(
             error_code=error_status,
@@ -5671,6 +5697,7 @@ class StandardLoggingPayloadSetup:
             llm_provider=_llm_provider_in_exception,
             traceback=_redact_string(traceback_info),
             error_message=_redact_string(error_message),
+            error_provider_request_id=provider_request_id,
             error_rate_limit_category=rate_limit_category,
             error_rate_limit_type=rate_limit_type,
             error_budget_entity_type=budget_error.entity_type if budget_error else None,
diff --git a/litellm/types/utils.py b/litellm/types/utils.py
index 3a1422bb190..010ff18d166 100644
--- a/litellm/types/utils.py
+++ b/litellm/types/utils.py
@@ -3064,6 +3064,7 @@ class StandardLoggingPayloadErrorInformation(TypedDict, total=False):
     llm_provider: str | None
     traceback: str | None
     error_message: str | None
+    error_provider_request_id: ReadOnly[str | None]
     # error_rate_limit_category:
     #   For 429 / rate-limit errors, the source of the rate limit. One of the
     #   string values defined by `litellm.exceptions.RateLimitErrorCategory`
diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
index 3748b5a9d64..0fdca755685 100644
--- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
+++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py
@@ -1,7 +1,9 @@
+import asyncio
 import contextlib
+import datetime
 import os
 import sys
-import asyncio
+from typing import Literal
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -6439,6 +6441,113 @@ def test_get_standard_logging_object_payload_survives_logging_obj_without_timing
     assert payload["hidden_params"]["litellm_overhead_time_ms"] is None
 
 
+@pytest.mark.parametrize(
+    ("header_name", "header_source"),
+    (
+        ("x-amzn-RequestId", "response"),
+        ("x-request-id", "response"),
+        ("request-id", "response"),
+        ("x-ms-request-id", "response"),
+        ("apim-request-id", "response"),
+        ("x-goog-request-id", "response"),
+        ("cf-ray", "response"),
+        ("X-Request-Id", "litellm_response_headers"),
+        ("X-MS-Request-ID", "headers"),
+    ),
+)
+def test_failure_standard_logging_payload_captures_provider_request_id(
+    logging_obj: LitellmLogging,
+    header_name: str,
+    header_source: Literal["response", "litellm_response_headers", "headers"],
+):
+    from litellm.litellm_core_utils.litellm_logging import (
+        get_standard_logging_object_payload,
+    )
+
+    request_id = "provider-request-123"
+    request = httpx.Request("POST", "https://provider.example/v1/chat/completions")
+    response = httpx.Response(429, headers={header_name: request_id}, request=request)
+    provider_error = httpx.HTTPStatusError("provider error", request=request, response=response)
+    if header_source == "litellm_response_headers":
+        response.headers.clear()
+        provider_error.litellm_response_headers = {header_name: request_id}
+    elif header_source == "headers":
+        response.headers.clear()
+        provider_error.headers = {header_name: request_id}
+    now = datetime.datetime.now()
+
+    payload = get_standard_logging_object_payload(
+        kwargs={"litellm_call_id": "call-1", "model": "test-model", "messages": []},
+        init_response_obj={},
+        start_time=now,
+        end_time=now,
+        logging_obj=logging_obj,
+        status="failure",
+        original_exception=provider_error,
+    )
+
+    assert payload is not None
+    assert payload["error_information"] is not None
+    assert payload["error_information"]["error_provider_request_id"] == request_id
+
+
+def test_get_error_information_ignores_unsupported_headers() -> None:
+    from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
+
+    request = httpx.Request("POST", "https://provider.example/v1/chat/completions")
+    response = httpx.Response(429, headers={"retry-after": "3"}, request=request)
+    provider_error = httpx.HTTPStatusError("provider error", request=request, response=response)
+
+    error_information = StandardLoggingPayloadSetup.get_error_information(provider_error)
+
+    assert error_information["error_provider_request_id"] is None
+
+
+def test_get_error_information_uses_header_precedence_and_fallback() -> None:
+    from litellm.exceptions import RateLimitError
+    from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
+
+    request = httpx.Request("POST", "https://provider.example/v1/chat/completions")
+    response = httpx.Response(
+        429,
+        headers={"x-request-id": "response-id", "x-amzn-requestid": "amazon-id"},
+        request=request,
+    )
+    provider_error = RateLimitError(
+        message="provider error",
+        llm_provider="test-provider",
+        model="test-model",
+        response=response,
+        headers={"retry-after": "3"},
+    )
+
+    error_information = StandardLoggingPayloadSetup.get_error_information(provider_error)
+
+    assert error_information["error_provider_request_id"] == "amazon-id"
+
+
+def test_get_error_information_ignores_malformed_headers() -> None:
+    from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup
+
+    provider_error = Exception("provider error")
+    provider_error.headers = [("x-request-id", "provider-request-123")]
+
+    error_information = StandardLoggingPayloadSetup.get_error_information(provider_error)
+
+    assert error_information["error_provider_request_id"] is None
+
+
+def test_get_provider_request_id_ignores_header_lookup_errors() -> None:
+    from litellm.litellm_core_utils.litellm_logging import _get_provider_request_id
+
+    class HeaderLookupError(Exception):
+        @property
+        def response(self) -> object:
+            raise RuntimeError("headers unavailable")
+
+    assert _get_provider_request_id(HeaderLookupError("provider error")) is None
+
+
 def test_get_standard_logging_object_payload_failure_status_keeps_overhead_none(logging_obj):
     """A post_call guardrail can fail the request after the upstream call succeeded; the failure
     payload keeps litellm_overhead_time_ms None, matching responses that carry their own _hidden_params."""

From 9acc01efce25bc83769a5f45f0c1de4991edcb3e Mon Sep 17 00:00:00 2001
From: yuneng-jiang <yuneng@berri.ai>
Date: Mon, 7 Sep 2026 11:18:16 -0700
Subject: [PATCH 70/79] test(e2e/ui): cover member role and budget edits,
 member permission delegation, and team guardrail removal (#40042)

* test(e2e/ui): cover member role and budget edits, member permission delegation, and team guardrail removal

Three Playwright specs for the Teams flows enterprise customers hit most, each
owning its fixtures and proving the mutation through a read-back rather than a
toast.

- teamMemberEdit: an admin edits a member's team role and per-member budget,
  and both survive a reload of the Members table
- memberPermissions: a plain member is refused /key/generate for their team,
  a team admin grants it on the Member Permissions tab, and the member then
  creates a team key that serves a real completion
- teamGuardrailRemoval: clearing a team's only guardrail on the Settings tab
  really clears it, and traffic the guardrail refused starts serving again

* test(e2e/ui): make the new team specs safe to run in parallel

Fixture ids came from Date.now(), so two repeats starting in the same
millisecond minted the same user id: one got a 409 and the loser's teardown
deleted the user the other was still signed in as. Ids now carry a random
suffix.

Also move the member-permissions setup inside the cleanup-protected block so a
half-finished setup cannot leak a team, and close both browser contexts the
test opens.
---
 .../proxy-admin/teamGuardrailRemoval.spec.ts  | 145 ++++++++++++++
 .../tests/proxy-admin/teamMemberEdit.spec.ts  | 129 +++++++++++++
 .../team-admin/memberPermissions.spec.ts      | 177 ++++++++++++++++++
 3 files changed, 451 insertions(+)
 create mode 100644 tests/e2e/ui/tests/proxy-admin/teamGuardrailRemoval.spec.ts
 create mode 100644 tests/e2e/ui/tests/proxy-admin/teamMemberEdit.spec.ts
 create mode 100644 tests/e2e/ui/tests/team-admin/memberPermissions.spec.ts

diff --git a/tests/e2e/ui/tests/proxy-admin/teamGuardrailRemoval.spec.ts b/tests/e2e/ui/tests/proxy-admin/teamGuardrailRemoval.spec.ts
new file mode 100644
index 00000000000..f7930378b1e
--- /dev/null
+++ b/tests/e2e/ui/tests/proxy-admin/teamGuardrailRemoval.spec.ts
@@ -0,0 +1,145 @@
+import { test, expect, type APIRequestContext, type Page as PlaywrightPage } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { Page } from "../../fixtures/pages";
+import { navigateToPage, dismissFeedbackPopup, clickTeamId } from "../../helpers/navigation";
+import { proxyIsPremium } from "../../helpers/premium";
+import { readBack } from "../../helpers/roundTrip";
+import { CHAT_MODEL_A, MOCK_RESPONSE_TEXT, masterKey } from "../../helpers/traffic";
+
+const auth = () => ({ Authorization: `Bearer ${masterKey()}` });
+
+async function guardrailId(request: APIRequestContext, name: string): Promise<string | undefined> {
+  const res = await request.get("/v2/guardrails/list", { headers: auth() });
+  expect(res.ok(), `GET /v2/guardrails/list (${res.status()})`).toBe(true);
+  const rows = (await res.json()).guardrails as { guardrail_id: string; guardrail_name: string | null }[];
+  return rows.find((row) => row.guardrail_name === name)?.guardrail_id;
+}
+
+async function teamGuardrails(page: PlaywrightPage, teamId: string): Promise<string[]> {
+  const body = await readBack<{ team_info: { metadata: { guardrails?: string[] } | null } }>(
+    page,
+    `/team/info?team_id=${encodeURIComponent(teamId)}`,
+  );
+  return body.team_info.metadata?.guardrails ?? [];
+}
+
+async function keywordPromptStatus(request: APIRequestContext, apiKey: string, keyword: string): Promise<number> {
+  const res = await request.post("/v1/chat/completions", {
+    headers: { Authorization: `Bearer ${apiKey}`, "Content-Type": "application/json" },
+    data: { model: CHAT_MODEL_A, messages: [{ role: "user", content: `please tell me about ${keyword}` }] },
+  });
+  return res.status();
+}
+
+test.describe("Proxy Admin - Team guardrail removal", () => {
+  test.use({ storageState: ADMIN_STORAGE_PATH });
+
+  test("Clearing a team's only guardrail on the Settings tab lets blocked traffic through again", async ({
+    page,
+    request,
+  }) => {
+    test.skip(!proxyIsPremium(), "proxy under test is unlicensed, so team guardrails are premium-gated");
+
+    const stamp = `${Date.now().toString(36)}${Math.random().toString(36).slice(2, 8)}`;
+    const guardrailName = `e2e-team-guardrail-${stamp}`;
+    const bannedKeyword = `e2eteamban${stamp}`;
+    const teamAlias = `e2e-guardrail-team-${stamp}`;
+
+    let teamId = "";
+    let teamKey = "";
+    try {
+      const guardrailRes = await request.post("/guardrails", {
+        headers: auth(),
+        data: {
+          guardrail: {
+            guardrail_name: guardrailName,
+            litellm_params: {
+              guardrail: "litellm_content_filter",
+              mode: "pre_call",
+              default_on: false,
+              blocked_words: [{ keyword: bannedKeyword, action: "BLOCK" }],
+            },
+          },
+        },
+      });
+      expect(
+        guardrailRes.ok(),
+        `POST /guardrails failed (${guardrailRes.status()}): ${await guardrailRes.text()}`,
+      ).toBe(true);
+
+      const teamRes = await request.post("/team/new", {
+        headers: auth(),
+        data: { team_alias: teamAlias, models: [CHAT_MODEL_A], metadata: { guardrails: [guardrailName] } },
+      });
+      expect(teamRes.ok(), `POST /team/new failed (${teamRes.status()}): ${await teamRes.text()}`).toBe(true);
+      teamId = (await teamRes.json()).team_id as string;
+
+      const keyRes = await request.post("/key/generate", { headers: auth(), data: { team_id: teamId } });
+      expect(keyRes.ok(), `POST /key/generate failed (${keyRes.status()}): ${await keyRes.text()}`).toBe(true);
+      teamKey = (await keyRes.json()).key as string;
+
+      await expect
+        .poll(async () => keywordPromptStatus(request, teamKey, bannedKeyword), {
+          message: "the team's guardrail never started refusing the banned keyword",
+          timeout: 60_000,
+        })
+        .toBe(400);
+
+      await navigateToPage(page, Page.Teams);
+      await dismissFeedbackPopup(page);
+      await clickTeamId(page, teamId);
+      await page.getByRole("tab", { name: "Settings" }).click();
+      await page.getByRole("button", { name: "Edit Settings" }).click();
+
+      const chip = page.locator('[data-slot="combobox-chip"]').filter({ hasText: guardrailName });
+      await expect(chip).toBeVisible({ timeout: 10_000 });
+      await chip.locator('[data-slot="combobox-chip-remove"]').click();
+      await expect(chip).toHaveCount(0, { timeout: 10_000 });
+
+      await page.getByRole("button", { name: "Save Changes" }).click();
+
+      await expect
+        .poll(async () => teamGuardrails(page, teamId), {
+          message: "the team still carries a guardrail in /team/info after the save",
+          timeout: 20_000,
+        })
+        .toEqual([]);
+
+      await page.reload();
+      await page.getByRole("tab", { name: "Settings" }).click();
+      await page.getByRole("button", { name: "Edit Settings" }).click();
+      await expect(page.getByRole("combobox", { name: "Select guardrails" })).toBeVisible({ timeout: 15_000 });
+      await expect(
+        page.locator('[data-slot="combobox-chip"]').filter({ hasText: guardrailName }),
+        "the removed guardrail is gone from the Settings tab after a reload",
+      ).toHaveCount(0);
+
+      await expect
+        .poll(async () => keywordPromptStatus(request, teamKey, bannedKeyword), {
+          message: "the team key is still refused for a keyword whose guardrail was removed",
+          timeout: 60_000,
+        })
+        .toBe(200);
+
+      const served = await request.post("/v1/chat/completions", {
+        headers: { Authorization: `Bearer ${teamKey}`, "Content-Type": "application/json" },
+        data: {
+          model: CHAT_MODEL_A,
+          messages: [{ role: "user", content: `please tell me about ${bannedKeyword}` }],
+        },
+      });
+      expect((await served.json()).choices?.[0]?.message?.content).toContain(MOCK_RESPONSE_TEXT);
+    } finally {
+      if (teamKey) {
+        await request.post("/key/delete", { headers: auth(), data: { keys: [teamKey] } });
+      }
+      if (teamId) {
+        await request.post("/team/delete", { headers: auth(), data: { team_ids: [teamId] } });
+      }
+      const id = await guardrailId(request, guardrailName);
+      if (id) {
+        await request.delete(`/guardrails/${id}`, { headers: auth() });
+      }
+    }
+  });
+});
diff --git a/tests/e2e/ui/tests/proxy-admin/teamMemberEdit.spec.ts b/tests/e2e/ui/tests/proxy-admin/teamMemberEdit.spec.ts
new file mode 100644
index 00000000000..c7325e78f75
--- /dev/null
+++ b/tests/e2e/ui/tests/proxy-admin/teamMemberEdit.spec.ts
@@ -0,0 +1,129 @@
+import { test, expect, type Page as PlaywrightPage } from "@playwright/test";
+import { ADMIN_STORAGE_PATH } from "../../constants";
+import { Page } from "../../fixtures/pages";
+import { navigateToPage, dismissFeedbackPopup, clickTeamId } from "../../helpers/navigation";
+import { readBack } from "../../helpers/roundTrip";
+import { CHAT_MODEL_A, masterKey } from "../../helpers/traffic";
+
+interface TeamInfoResponse {
+  team_info: {
+    models: string[];
+    members_with_roles: { user_id?: string; role?: string }[];
+  };
+  team_memberships: {
+    user_id: string;
+    litellm_budget_table: { max_budget: number | null } | null;
+  }[];
+}
+
+const auth = () => ({ Authorization: `Bearer ${masterKey()}` });
+
+async function teamInfo(page: PlaywrightPage, teamId: string): Promise<TeamInfoResponse> {
+  return readBack<TeamInfoResponse>(page, `/team/info?team_id=${encodeURIComponent(teamId)}`);
+}
+
+function roleOf(info: TeamInfoResponse, userId: string): string | undefined {
+  return info.team_info.members_with_roles.find((member) => member.user_id === userId)?.role;
+}
+
+function budgetOf(info: TeamInfoResponse, userId: string): number | null | undefined {
+  return info.team_memberships.find((membership) => membership.user_id === userId)?.litellm_budget_table?.max_budget;
+}
+
+function otherMembers(info: TeamInfoResponse, userId: string): string[] {
+  return info.team_info.members_with_roles
+    .filter((member) => member.user_id !== userId)
+    .map((member) => `${member.user_id}:${member.role}`)
+    .sort();
+}
+
+test.describe("Proxy Admin - Team member edit", () => {
+  test.use({ storageState: ADMIN_STORAGE_PATH });
+
+  const createdTeams: string[] = [];
+  const createdUsers: string[] = [];
+
+  test.afterEach(async ({ request }) => {
+    for (const teamId of createdTeams.splice(0)) {
+      await request.post("/team/delete", { headers: auth(), data: { team_ids: [teamId] } });
+    }
+    for (const userId of createdUsers.splice(0)) {
+      await request.post("/user/delete", { headers: auth(), data: { user_ids: [userId] } });
+    }
+  });
+
+  test("Editing a member's role and per-member budget persists and survives a reload", async ({ page, request }) => {
+    const stamp = `${Date.now().toString(36)}${Math.random().toString(36).slice(2, 8)}`;
+    const memberId = `e2e-member-edit-${stamp}`;
+    const teamAlias = `e2e-member-edit-team-${stamp}`;
+
+    createdUsers.push(memberId);
+    const created = await request.post("/user/new", {
+      headers: auth(),
+      data: { user_id: memberId, user_role: "internal_user", auto_create_key: false },
+    });
+    expect(created.ok(), `POST /user/new failed (${created.status()}): ${await created.text()}`).toBe(true);
+
+    const teamRes = await request.post("/team/new", {
+      headers: auth(),
+      data: {
+        team_alias: teamAlias,
+        models: [CHAT_MODEL_A],
+        members_with_roles: [{ user_id: memberId, role: "admin" }],
+      },
+    });
+    expect(teamRes.ok(), `POST /team/new failed (${teamRes.status()}): ${await teamRes.text()}`).toBe(true);
+    const teamId = (await teamRes.json()).team_id as string;
+    createdTeams.push(teamId);
+
+    const before = await teamInfo(page, teamId);
+    expect(roleOf(before, memberId), "the member starts out as a team admin").toBe("admin");
+    expect(budgetOf(before, memberId) ?? null, "the member starts out with no per-member budget").toBeNull();
+    expect(
+      otherMembers(before, memberId).length,
+      "the team has another member for the edit to leave alone",
+    ).toBeGreaterThan(0);
+
+    await navigateToPage(page, Page.Teams);
+    await dismissFeedbackPopup(page);
+    await clickTeamId(page, teamId);
+    await page.getByRole("tab", { name: "Members" }).click();
+
+    const memberRow = page.locator("tr", { hasText: memberId }).first();
+    await expect(memberRow).toBeVisible({ timeout: 10_000 });
+    await memberRow.getByTestId("edit-member").click();
+
+    const modal = page.getByRole("dialog", { name: "Edit Member" });
+    await expect(modal).toBeVisible({ timeout: 10_000 });
+
+    await modal.getByLabel(/^Role/).click();
+    await page.getByRole("option", { name: "User", exact: true }).click();
+    await modal.getByLabel(/Team Member Budget \(USD\)/).fill("5");
+    await modal.getByRole("button", { name: "Save Changes" }).click();
+
+    await expect(page.getByText("Team member updated successfully").first()).toBeVisible({ timeout: 10_000 });
+
+    await expect
+      .poll(
+        async () => {
+          const info = await teamInfo(page, teamId);
+          return [roleOf(info, memberId), budgetOf(info, memberId)];
+        },
+        { message: "the member's role and budget never landed in /team/info", timeout: 20_000 },
+      )
+      .toEqual(["user", 5]);
+
+    await page.reload();
+    await page.getByRole("tab", { name: "Members" }).click();
+    const reloadedRow = page.locator("tr", { hasText: memberId }).first();
+    await expect(reloadedRow).toBeVisible({ timeout: 15_000 });
+    await expect(reloadedRow.getByText("user", { exact: true }), "role shown after a reload").toBeVisible();
+    await expect(reloadedRow.getByText("$5.00"), "per-member budget shown after a reload").toBeVisible();
+
+    const after = await teamInfo(page, teamId);
+    expect(after.team_info.models, "model access untouched by a member edit").toEqual(before.team_info.models);
+    expect(otherMembers(after, memberId), "the rest of the roster untouched by a member edit").toEqual(
+      otherMembers(before, memberId),
+    );
+  });
+});
diff --git a/tests/e2e/ui/tests/team-admin/memberPermissions.spec.ts b/tests/e2e/ui/tests/team-admin/memberPermissions.spec.ts
new file mode 100644
index 00000000000..75fb3be9b64
--- /dev/null
+++ b/tests/e2e/ui/tests/team-admin/memberPermissions.spec.ts
@@ -0,0 +1,177 @@
+import { test, expect, type Browser, type BrowserContext, type Page as PlaywrightPage } from "@playwright/test";
+import { Page } from "../../fixtures/pages";
+import { navigateToPage, dismissFeedbackPopup, clickTeamId } from "../../helpers/navigation";
+import { CHAT_MODEL_A, MOCK_RESPONSE_TEXT, masterKey } from "../../helpers/traffic";
+
+const PASSWORD = "E2e-Member-Perms-Pass-1!";
+
+const auth = () => ({ Authorization: `Bearer ${masterKey()}` });
+
+async function sessionKey(page: PlaywrightPage): Promise<string> {
+  const cookie = (await page.context().cookies()).find((candidate) => candidate.name === "token");
+  expect(cookie?.value, "logged-in session carries a token cookie").toBeTruthy();
+  const payload = JSON.parse(Buffer.from(cookie!.value.split(".")[1], "base64url").toString("utf-8")) as {
+    key?: string;
+  };
+  expect(payload.key, "session JWT carries the virtual key the dashboard calls with").toMatch(/^sk-/);
+  return payload.key!;
+}
+
+async function signIn(browser: Browser, email: string): Promise<BrowserContext> {
+  const context = await browser.newContext({ storageState: { cookies: [], origins: [] } });
+  const page = await context.newPage();
+  await page.goto("/ui/login");
+  await page.getByPlaceholder("Enter your username").fill(email);
+  await page.getByPlaceholder("Enter your password").fill(PASSWORD);
+  await page.getByRole("button", { name: "Login", exact: true }).click();
+  await expect(page.locator("a", { hasText: "Virtual Keys" })).toBeVisible({ timeout: 30_000 });
+  await dismissFeedbackPopup(page);
+  return context;
+}
+
+test.describe("Team Admin - Member permissions", () => {
+  test.use({ storageState: { cookies: [], origins: [] } });
+
+  test("Granting /key/generate lets a plain member create a team key that serves traffic", async ({
+    browser,
+    request,
+  }) => {
+    const stamp = `${Date.now().toString(36)}${Math.random().toString(36).slice(2, 8)}`;
+    const adminId = `e2e-perm-admin-${stamp}`;
+    const memberId = `e2e-perm-member-${stamp}`;
+    const adminEmail = `${adminId}@test.local`;
+    const memberEmail = `${memberId}@test.local`;
+    const teamAlias = `e2e-perm-team-${stamp}`;
+
+    const createUser = async (userId: string, email: string): Promise<void> => {
+      const created = await request.post("/user/new", {
+        headers: auth(),
+        data: { user_id: userId, user_email: email, user_role: "internal_user", auto_create_key: false },
+      });
+      expect(created.ok(), `POST /user/new for ${userId} (${created.status()}): ${await created.text()}`).toBe(true);
+      const password = await request.post("/user/update", {
+        headers: auth(),
+        data: { user_id: userId, password: PASSWORD },
+      });
+      expect(password.ok(), `POST /user/update for ${userId} (${password.status()})`).toBe(true);
+    };
+
+    let teamId = "";
+    const createdKeys: string[] = [];
+    const contexts: BrowserContext[] = [];
+    try {
+      await createUser(adminId, adminEmail);
+      await createUser(memberId, memberEmail);
+
+      const teamRes = await request.post("/team/new", {
+        headers: auth(),
+        data: {
+          team_alias: teamAlias,
+          models: [CHAT_MODEL_A],
+          members_with_roles: [
+            { user_id: adminId, role: "admin" },
+            { user_id: memberId, role: "user" },
+          ],
+        },
+      });
+      expect(teamRes.ok(), `POST /team/new failed (${teamRes.status()}): ${await teamRes.text()}`).toBe(true);
+      teamId = (await teamRes.json()).team_id as string;
+
+      const memberContext = await signIn(browser, memberEmail);
+      contexts.push(memberContext);
+      const memberPage = memberContext.pages()[0];
+      const memberSessionKey = await sessionKey(memberPage);
+
+      const refused = await memberPage.request.post("/key/generate", {
+        headers: { Authorization: `Bearer ${memberSessionKey}`, "Content-Type": "application/json" },
+        data: { team_id: teamId, key_alias: `e2e-perm-denied-${stamp}` },
+      });
+      expect(refused.status(), "a plain member cannot mint a team key before the grant").toBe(401);
+      expect(await refused.text()).toContain("/key/generate");
+
+      const adminContext = await signIn(browser, adminEmail);
+      contexts.push(adminContext);
+      const adminPage = adminContext.pages()[0];
+      await navigateToPage(adminPage, Page.Teams);
+      await clickTeamId(adminPage, teamId);
+      await adminPage.getByRole("tab", { name: "Member Permissions" }).click();
+
+      for (const route of ["/key/generate", "/key/update"]) {
+        await adminPage.getByRole("row").filter({ hasText: route }).getByRole("checkbox").check();
+      }
+      await adminPage.getByRole("button", { name: "Save Changes" }).click();
+      await expect(adminPage.getByText("Permissions updated successfully").first()).toBeVisible({ timeout: 10_000 });
+
+      await expect
+        .poll(
+          async () => {
+            const res = await request.get(`/team/permissions_list?team_id=${encodeURIComponent(teamId)}`, {
+              headers: auth(),
+            });
+            if (!res.ok()) return [];
+            return ((await res.json()).team_member_permissions ?? []) as string[];
+          },
+          { message: "the granted permissions never landed in /team/permissions_list", timeout: 20_000 },
+        )
+        .toEqual(expect.arrayContaining(["/key/generate", "/key/update"]));
+
+      const keyAlias = `e2e-perm-key-${stamp}`;
+      await navigateToPage(memberPage, Page.ApiKeys);
+      await memberPage.getByRole("button", { name: /Create New Key/i }).click();
+      await expect(memberPage.getByText("Key Ownership")).toBeVisible({ timeout: 10_000 });
+      await memberPage.getByLabel(/Key Name/).fill(keyAlias);
+
+      const teamSelect = memberPage.getByTestId("team-dropdown").getByRole("combobox");
+      await teamSelect.click();
+      await memberPage.keyboard.type(teamAlias);
+      await memberPage.getByRole("option", { name: teamAlias }).first().click();
+
+      await memberPage.getByRole("combobox", { name: "Select models" }).click();
+      await memberPage.getByRole("option", { name: "All Team Models", exact: true }).click();
+      await memberPage.keyboard.press("Escape");
+
+      await memberPage.getByRole("button", { name: "Create Key", exact: true }).click();
+      const saveDialog = memberPage.getByRole("dialog", { name: "Save your Key" });
+      await expect(saveDialog).toBeVisible({ timeout: 15_000 });
+      const apiKey = (await saveDialog.locator("pre").innerText()).trim();
+      expect(apiKey).toMatch(/^sk-/);
+      createdKeys.push(apiKey);
+      await memberPage.keyboard.press("Escape");
+
+      await expect
+        .poll(
+          async () => {
+            const res = await request.get(
+              `/key/list?team_id=${encodeURIComponent(teamId)}&return_full_object=true&size=100`,
+              { headers: auth() },
+            );
+            if (!res.ok()) return null;
+            const row = ((await res.json()).keys as Record<string, unknown>[]).find(
+              (candidate) => candidate.key_alias === keyAlias,
+            );
+            return row ? [row.user_id, row.team_id] : null;
+          },
+          { message: `key ${keyAlias} never appeared on the team with the member as its owner`, timeout: 20_000 },
+        )
+        .toEqual([memberId, teamId]);
+
+      const served = await request.post("/v1/chat/completions", {
+        headers: { Authorization: `Bearer ${apiKey}`, "Content-Type": "application/json" },
+        data: { model: CHAT_MODEL_A, messages: [{ role: "user", content: `member key ping ${stamp}` }] },
+      });
+      expect(served.status(), "the delegated key is a real key the gateway serves").toBe(200);
+      expect((await served.json()).choices?.[0]?.message?.content).toContain(MOCK_RESPONSE_TEXT);
+    } finally {
+      for (const context of contexts) {
+        await context.close();
+      }
+      for (const key of createdKeys) {
+        await request.post("/key/delete", { headers: auth(), data: { keys: [key] } });
+      }
+      if (teamId) {
+        await request.post("/team/delete", { headers: auth(), data: { team_ids: [teamId] } });
+      }
+      await request.post("/user/delete", { headers: auth(), data: { user_ids: [adminId, memberId] } });
+    }
+  });
+});

From 5f2b4d27d712847c0f4fef5333f24848983686f5 Mon Sep 17 00:00:00 2001
From: yujonglee <yujong@berri.ai>
Date: Mon, 7 Sep 2026 11:23:12 -0700
Subject: [PATCH 71/79] test(ocr): add SDK callback E2E parity (#40061)

* test(ocr): cover SDK callback parity

* test(ocr): compare callback kwargs
---
 litellm/ocr/main.py                           |  74 +++--
 .../shared/parity/compare.py                  |   1 +
 .../shared/parity/models.py                   |  21 ++
 .../shared/parity/runner.py                   |  10 +-
 .../shared/parity/test_parity.py              |  52 ++-
 .../strategies/e2e_parity/__init__.py         |   4 +-
 .../e2e_parity/sdk/ocr/test_sdk_parity.py     | 295 +++++++++++++++++-
 .../unit_tests_mapping/cases/ocr.py           |   4 +
 tests/test_litellm/ocr/test_rust_bridge.py    |  22 ++
 9 files changed, 453 insertions(+), 30 deletions(-)

diff --git a/litellm/ocr/main.py b/litellm/ocr/main.py
index 56c2292d00c..df3f9d2096b 100644
--- a/litellm/ocr/main.py
+++ b/litellm/ocr/main.py
@@ -29,6 +29,7 @@ from litellm.llms.base_llm.ocr.transformation import (
 )
 from litellm.llms.custom_httpx.llm_http_handler import BaseLLMHTTPHandler
 from litellm.rust_bridge import ocr as rust_ocr_bridge
+from litellm.rust_bridge.bindings import native_exception_types
 from litellm.rust_bridge.configuration import rust_enabled
 from litellm.types.router import GenericLiteLLMParams
 from litellm.utils import ProviderConfigManager, client
@@ -281,6 +282,33 @@ def _prepare_rust_ocr_call(
     )
 
 
+def _map_rust_ocr_error(
+    error: Exception,
+    prepared_request: _PreparedOCRRequest,
+    exception_types: tuple[type[BaseException], type[BaseException]] | None,
+) -> Exception:
+    if exception_types is None:
+        return error
+    _, upstream_error = exception_types
+    if not isinstance(error, upstream_error):
+        return error
+    error_args: Final = cast(  # cast-ok: BaseException.args is typed with Any in the standard library stubs
+        tuple[object, ...], error.args
+    )
+    status_value: Final = error_args[0] if error_args else 0
+    message_value: Final = error_args[1] if len(error_args) > 1 else str(error)
+    status: Final = status_value if isinstance(status_value, int) else 0
+    message: Final = message_value if isinstance(message_value, str) else str(message_value)
+    error_factory: Final = cast(  # cast-ok: the legacy provider interface leaves callable parameters untyped
+        Callable[..., Exception], prepared_request.provider_config.get_error_class
+    )
+    return error_factory(
+        error_message=message,
+        status_code=status or 500,
+        headers={},  # mutable-ok: provider error factories require a concrete header dict
+    )
+
+
 def _run_rust_ocr(
     prepared_request: _PreparedOCRRequest,
     resolve_api_key: Callable[[str], str | None],
@@ -291,16 +319,19 @@ def _run_rust_ocr(
         prepared_request=prepared_request,
         resolve_api_key=resolve_api_key,
     )
-    rust_response: Final = rust_ocr_bridge.ocr(
-        model=prepared_request.model,
-        document=prepared_request.document,
-        api_key=prepared.api_key,
-        api_base=prepared.api_base,
-        custom_llm_provider=prepared_request.custom_llm_provider,
-        extra_headers=prepared.headers,
-        optional_params=prepared.optional_params,
-        timeout=prepared_request.effective_timeout,
-    )
+    try:
+        rust_response: Final = rust_ocr_bridge.ocr(
+            model=prepared_request.model,
+            document=prepared_request.document,
+            api_key=prepared.api_key,
+            api_base=prepared.api_base,
+            custom_llm_provider=prepared_request.custom_llm_provider,
+            extra_headers=prepared.headers,
+            optional_params=prepared.optional_params,
+            timeout=prepared_request.effective_timeout,
+        )
+    except Exception as error:
+        raise _map_rust_ocr_error(error, prepared_request, native_exception_types()) from error
     if rust_response is None:
         return None
     return OCRResponse.model_validate(rust_response)
@@ -316,16 +347,19 @@ async def _run_rust_aocr(
         prepared_request=prepared_request,
         resolve_api_key=resolve_api_key,
     )
-    rust_response: Final = await rust_ocr_bridge.aocr(
-        model=prepared_request.model,
-        document=prepared_request.document,
-        api_key=prepared.api_key,
-        api_base=prepared.api_base,
-        custom_llm_provider=prepared_request.custom_llm_provider,
-        extra_headers=prepared.headers,
-        optional_params=prepared.optional_params,
-        timeout=prepared_request.effective_timeout,
-    )
+    try:
+        rust_response: Final = await rust_ocr_bridge.aocr(
+            model=prepared_request.model,
+            document=prepared_request.document,
+            api_key=prepared.api_key,
+            api_base=prepared.api_base,
+            custom_llm_provider=prepared_request.custom_llm_provider,
+            extra_headers=prepared.headers,
+            optional_params=prepared.optional_params,
+            timeout=prepared_request.effective_timeout,
+        )
+    except Exception as error:
+        raise _map_rust_ocr_error(error, prepared_request, native_exception_types()) from error
     if rust_response is None:
         return None
     return OCRResponse.model_validate(rust_response)
diff --git a/tests/rust-python-harness/shared/parity/compare.py b/tests/rust-python-harness/shared/parity/compare.py
index adf85e5c8d7..88239ed042a 100644
--- a/tests/rust-python-harness/shared/parity/compare.py
+++ b/tests/rust-python-harness/shared/parity/compare.py
@@ -78,3 +78,4 @@ def assert_parity(baseline: Execution, candidate: Execution, baseline_user_agent
     validate_harness(baseline, candidate, baseline_user_agent)
     assert_request_parity(baseline.requests, candidate.requests)
     assert_value_parity(baseline.report, candidate.report)
+    assert_value_parity(baseline.callbacks, candidate.callbacks, path="$.callbacks")
diff --git a/tests/rust-python-harness/shared/parity/models.py b/tests/rust-python-harness/shared/parity/models.py
index 898b58d23ee..5612e218824 100644
--- a/tests/rust-python-harness/shared/parity/models.py
+++ b/tests/rust-python-harness/shared/parity/models.py
@@ -37,6 +37,25 @@ class SDKError(BaseModel):
     llm_provider: str | None
 
 
+class CallbackObservation(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    hook: Literal[
+        "log_success_event",
+        "async_log_success_event",
+        "log_failure_event",
+        "async_log_failure_event",
+    ]
+    phase: Literal["success", "failure"]
+    model: str | None
+    call_type: str | None
+    litellm_call_id: str | None
+    metadata: JsonValue
+    kwargs: JsonValue
+    payload: JsonValue
+    error: SDKError | None
+
+
 class SDKJsonChunk(BaseModel):
     model_config = ConfigDict(frozen=True)
 
@@ -119,6 +138,7 @@ class Execution(BaseModel):
 
     requests: tuple[CapturedRequest, ...]
     report: SDKReport
+    callbacks: tuple[CallbackObservation, ...] | None = None
 
 
 class SDKCommand(BaseModel):
@@ -133,6 +153,7 @@ class WorkerSuccess(BaseModel):
 
     status: Literal["ok"] = "ok"
     report: SDKReport
+    callbacks: tuple[CallbackObservation, ...] | None = None
 
 
 class WorkerFailure(BaseModel):
diff --git a/tests/rust-python-harness/shared/parity/runner.py b/tests/rust-python-harness/shared/parity/runner.py
index 43a583382cb..feae0201aee 100644
--- a/tests/rust-python-harness/shared/parity/runner.py
+++ b/tests/rust-python-harness/shared/parity/runner.py
@@ -39,9 +39,7 @@ class SubprocessRunner:
         return (
             sys.executable,
             "-m",
-            ".".join(
-                self.entrypoint.resolve().relative_to(PROJECT_ROOT).with_suffix("").parts
-            ),
+            ".".join(self.entrypoint.resolve().relative_to(PROJECT_ROOT).with_suffix("").parts),
             "--parity-worker",
             provider_url,
         )
@@ -113,7 +111,11 @@ class SubprocessWorker:
             )
         assert isinstance(result, WorkerSuccess)
         try:
-            return Execution(requests=self.provider.take_requests(len(responses)), report=result.report)
+            return Execution(
+                requests=self.provider.take_requests(len(responses)),
+                report=result.report,
+                callbacks=result.callbacks,
+            )
         except AssertionError:
             self.provider.reset()
             raise
diff --git a/tests/rust-python-harness/shared/parity/test_parity.py b/tests/rust-python-harness/shared/parity/test_parity.py
index 83daccdf8ba..7c8355fbfd8 100644
--- a/tests/rust-python-harness/shared/parity/test_parity.py
+++ b/tests/rust-python-harness/shared/parity/test_parity.py
@@ -7,7 +7,7 @@ import pytest
 from pydantic import BaseModel, ConfigDict, JsonValue, PrivateAttr
 
 from .compare import assert_model_parity, assert_parity
-from .models import CapturedRequest, Execution, SDKError, SDKSuccess, sdk_error_report
+from .models import CallbackObservation, CapturedRequest, Execution, SDKError, SDKSuccess, sdk_error_report
 
 SENTINEL: Final = "python-parity-fallback"
 
@@ -69,6 +69,56 @@ def test_parity_rejects_response_difference() -> None:
         assert_parity(python, rust, SENTINEL)
 
 
+def test_parity_distinguishes_unobserved_callbacks_from_zero_events() -> None:
+    python: Final = _execution(user_agent=SENTINEL)
+    rust: Final = _execution(user_agent="litellm-rust").model_copy(update={"callbacks": ()})
+
+    with pytest.raises(AssertionError, match=r"\$\.callbacks"):
+        assert_parity(python, rust, SENTINEL)
+
+
+def test_parity_rejects_callback_payload_difference() -> None:
+    observation: Final = CallbackObservation(
+        hook="log_success_event",
+        phase="success",
+        model="test-model",
+        call_type="ocr",
+        litellm_call_id="test-call",
+        metadata={"profile": "success"},
+        kwargs={"model": "test-model"},
+        payload={"model": "test-model", "pages": []},
+        error=None,
+    )
+    python: Final = _execution(user_agent=SENTINEL).model_copy(update={"callbacks": (observation,)})
+    rust: Final = _execution(user_agent="litellm-rust").model_copy(
+        update={"callbacks": (observation.model_copy(update={"payload": {"model": "changed", "pages": []}}),)}
+    )
+
+    with pytest.raises(AssertionError, match=r"\$\.callbacks"):
+        assert_parity(python, rust, SENTINEL)
+
+
+def test_parity_rejects_callback_kwargs_difference() -> None:
+    observation: Final = CallbackObservation(
+        hook="log_success_event",
+        phase="success",
+        model="test-model",
+        call_type="ocr",
+        litellm_call_id="test-call",
+        metadata={"profile": "success"},
+        kwargs={"model": "test-model"},
+        payload={"model": "test-model", "pages": []},
+        error=None,
+    )
+    python: Final = _execution(user_agent=SENTINEL).model_copy(update={"callbacks": (observation,)})
+    rust: Final = _execution(user_agent="litellm-rust").model_copy(
+        update={"callbacks": (observation.model_copy(update={"kwargs": {"model": "changed"}}),)}
+    )
+
+    with pytest.raises(AssertionError, match=r"\$\.callbacks"):
+        assert_parity(python, rust, SENTINEL)
+
+
 def test_parity_rejects_error_difference() -> None:
     python: Final = Execution(
         requests=(),
diff --git a/tests/rust-python-harness/strategies/e2e_parity/__init__.py b/tests/rust-python-harness/strategies/e2e_parity/__init__.py
index f668e178eef..95346bbd496 100644
--- a/tests/rust-python-harness/strategies/e2e_parity/__init__.py
+++ b/tests/rust-python-harness/strategies/e2e_parity/__init__.py
@@ -18,8 +18,8 @@ CASES: Final[tuple[CaseDefinition, ...]] = (
             coverage=Coverage.PARTIAL,
             module="tests.rust-python-harness.strategies.e2e_parity.sdk.ocr.test_sdk_parity",
             note=(
-                "Recorded sync/async SDK parity; invalid-model provider errors differ, "
-                "and Reducto lacks a Rust contract."
+                "Recorded sync/async SDK parity with focused success/error callback profiles; "
+                "Reducto lacks a Rust contract, and known provider parity gaps remain."
             ),
         ),
         surface="sdk",
diff --git a/tests/rust-python-harness/strategies/e2e_parity/sdk/ocr/test_sdk_parity.py b/tests/rust-python-harness/strategies/e2e_parity/sdk/ocr/test_sdk_parity.py
index e72980752f2..5c4abc78081 100644
--- a/tests/rust-python-harness/strategies/e2e_parity/sdk/ocr/test_sdk_parity.py
+++ b/tests/rust-python-harness/strategies/e2e_parity/sdk/ocr/test_sdk_parity.py
@@ -1,23 +1,31 @@
 from __future__ import annotations
 
 import asyncio
+import datetime
+import queue
 import sys
 import tempfile
+import time
 import traceback
-from collections.abc import Callable, Coroutine, Generator
+from collections.abc import Callable, Coroutine, Generator, Mapping
 from contextlib import contextmanager
 from enum import Enum
 from functools import partial
 from pathlib import Path
 from typing import Annotated, Final, Literal, cast
+from urllib.parse import urlsplit, urlunsplit
 
 from pydantic import BaseModel, ConfigDict, Field, JsonValue, TypeAdapter
 
+from litellm.integrations.custom_logger import CustomLogger
 from litellm.llms.base_llm.ocr.transformation import OCRResponse
 
 from .....shared.parity.compare import assert_parity
 from .....shared.parity.fixtures.store import fixture_id, recorded_fixtures
 from .....shared.parity.models import (
+    JSON_VALUE_ADAPTER,
+    CallbackObservation,
+    Execution,
     SDKCommand,
     SDKError,
     SDKReport,
@@ -39,6 +47,9 @@ from .fixtures.config import configured_fixture_directory
 from .fixtures.models import OcrParityCase, OcrSdkInput
 
 API_KEY: Final = "test-key"
+CALLBACK_DELAY_SECONDS: Final = 0.05
+CALLBACK_DRAIN_TIMEOUT_SECONDS: Final = 10.0
+CALLBACK_TERMINALS: Final[tuple[Literal["success", "failure"], ...]] = ("success", "failure")
 PYTHON_HTTP_SENTINEL: Final = "python-ocr-parity-fallback"
 PYTHON_VARIANT: Final = ExecutionVariant(name="Python", environment=(("LITELLM_RUST", "0"),))
 RUST_VARIANT: Final = ExecutionVariant(name="Rust", environment=(("LITELLM_RUST", "1"),))
@@ -75,10 +86,148 @@ class InvalidOcrWorkerCase(BaseModel):
     case: InvalidOcrCase
 
 
-OcrWorkerCase = Annotated[RecordedOcrWorkerCase | InvalidOcrWorkerCase, Field(discriminator="kind")]
+class CallbackOcrWorkerCase(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    kind: Literal["callback"] = "callback"
+    case: OcrParityCase
+    terminal: Literal["success", "failure"]
+
+
+OcrWorkerCase = Annotated[
+    RecordedOcrWorkerCase | InvalidOcrWorkerCase | CallbackOcrWorkerCase,
+    Field(discriminator="kind"),
+]
 OCR_WORKER_CASE_ADAPTER: Final[TypeAdapter[OcrWorkerCase]] = TypeAdapter(OcrWorkerCase)
 
 
+class RecordingCallback(CustomLogger):
+    def __init__(self) -> None:
+        self.message_logging: Final = True
+        self.turn_off_message_logging: Final = False
+        self._observations: Final[queue.SimpleQueue[CallbackObservation]] = queue.SimpleQueue()
+
+    def _normalized_kwargs(self, value: object, key: str | None = None) -> JsonValue:
+        if value is None or isinstance(value, (bool, int, float)):
+            return value
+        if isinstance(value, str):
+            if key != "api_base":
+                return value
+            parsed: Final = urlsplit(value)
+            return urlunsplit(("", "", parsed.path, parsed.query, parsed.fragment))
+        if isinstance(value, datetime.datetime):
+            return "datetime"
+        if isinstance(value, Exception):
+            return sdk_error_report(value).model_dump(mode="json")
+        if isinstance(value, BaseModel):
+            return self._normalized_kwargs(value.model_dump(mode="json"), key)
+        if isinstance(value, Mapping):
+            if any(not isinstance(map_key, str) for map_key in value):
+                raise TypeError("callback kwarg mappings must use string keys")
+            return {
+                map_key: self._normalized_kwargs(map_value, map_key)
+                for map_key, map_value in value.items()
+            }
+        if isinstance(value, (list, tuple)):
+            return [self._normalized_kwargs(item) for item in value]
+        raise TypeError(f"unsupported callback kwarg type: {type(value)}")
+
+    def _record(
+        self,
+        hook: Literal[
+            "log_success_event",
+            "async_log_success_event",
+            "log_failure_event",
+            "async_log_failure_event",
+        ],
+        phase: Literal["success", "failure"],
+        kwargs: dict[str, object],
+        response_obj: object,
+    ) -> None:
+        raw_litellm_params: Final = kwargs.get("litellm_params")
+        litellm_params: Final[Mapping[str, object]] = (
+            cast(Mapping[str, object], raw_litellm_params) if isinstance(raw_litellm_params, Mapping) else {}
+        )
+        raw_metadata: Final = litellm_params.get("metadata")
+        metadata_mapping: Final[Mapping[str, object]] = (
+            cast(Mapping[str, object], raw_metadata) if isinstance(raw_metadata, Mapping) else {}
+        )
+        metadata: Final = JSON_VALUE_ADAPTER.validate_python(
+            {key: metadata_mapping[key] for key in ("callback_profile", "sdk_route") if key in metadata_mapping}
+        )
+        raw_error: Final = kwargs.get("exception")
+        error: Final = sdk_error_report(raw_error) if isinstance(raw_error, Exception) else None
+        normalized_kwargs: Final = self._normalized_kwargs(kwargs)
+        payload_source: Final = (
+            response_obj.model_dump(mode="json") if isinstance(response_obj, BaseModel) else response_obj
+        )
+        payload: Final = JSON_VALUE_ADAPTER.validate_python(payload_source)
+        raw_model: Final = kwargs.get("model")
+        raw_call_type: Final = kwargs.get("call_type")
+        raw_call_id: Final = kwargs.get("litellm_call_id")
+        self._observations.put(
+            CallbackObservation(
+                hook=hook,
+                phase=phase,
+                model=raw_model if isinstance(raw_model, str) else None,
+                call_type=str(raw_call_type) if raw_call_type is not None else None,
+                litellm_call_id=raw_call_id if isinstance(raw_call_id, str) else None,
+                metadata=metadata,
+                kwargs=normalized_kwargs,
+                payload=payload,
+                error=error,
+            )
+        )
+
+    def log_success_event(
+        self,
+        kwargs: dict[str, object],
+        response_obj: object,
+        start_time: datetime.datetime,
+        end_time: datetime.datetime,
+    ) -> None:
+        del start_time, end_time
+        time.sleep(CALLBACK_DELAY_SECONDS)
+        self._record("log_success_event", "success", kwargs, response_obj)
+
+    async def async_log_success_event(
+        self,
+        kwargs: dict[str, object],
+        response_obj: object,
+        start_time: datetime.datetime,
+        end_time: datetime.datetime,
+    ) -> None:
+        del start_time, end_time
+        await asyncio.sleep(CALLBACK_DELAY_SECONDS)
+        self._record("async_log_success_event", "success", kwargs, response_obj)
+
+    def log_failure_event(
+        self,
+        kwargs: dict[str, object],
+        response_obj: object,
+        start_time: datetime.datetime,
+        end_time: datetime.datetime,
+    ) -> None:
+        del start_time, end_time
+        time.sleep(CALLBACK_DELAY_SECONDS)
+        self._record("log_failure_event", "failure", kwargs, response_obj)
+
+    async def async_log_failure_event(
+        self,
+        kwargs: dict[str, object],
+        response_obj: object,
+        start_time: datetime.datetime,
+        end_time: datetime.datetime,
+    ) -> None:
+        del start_time, end_time
+        await asyncio.sleep(CALLBACK_DELAY_SECONDS)
+        self._record("async_log_failure_event", "failure", kwargs, response_obj)
+
+    def observations(self) -> tuple[CallbackObservation, ...]:
+        observations: Final = tuple(self._observations.get_nowait() for _ in range(self._observations.qsize()))
+        return tuple(sorted(observations, key=lambda observation: observation.hook))
+
+
 INVALID_OCR_CASES: Final = (
     InvalidOcrCase(
         name="unsupported_provider",
@@ -229,6 +378,106 @@ def _execute_invalid_sdk_case(
     return _execute_sdk_call(call_kwargs, route, event_loop)
 
 
+def _callback_call_id(route: SDKRoute, terminal: Literal["success", "failure"]) -> str:
+    return f"ocr-callback-{route.value}-{terminal}"
+
+
+def _callback_metadata(route: SDKRoute, terminal: Literal["success", "failure"]) -> dict[str, str]:
+    return {"callback_profile": terminal, "sdk_route": route.value}
+
+
+def _drain_callback_delivery(route: SDKRoute, event_loop: asyncio.AbstractEventLoop) -> None:
+    if route is SDKRoute.AOCR:
+        from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
+
+        async def drain_async_callbacks() -> None:
+            await asyncio.wait_for(GLOBAL_LOGGING_WORKER.flush(), timeout=CALLBACK_DRAIN_TIMEOUT_SECONDS)
+            await GLOBAL_LOGGING_WORKER.stop()
+
+        event_loop.run_until_complete(drain_async_callbacks())
+
+    from litellm.litellm_core_utils.thread_pool_executor import executor
+
+    executor.shutdown(wait=True, cancel_futures=False)
+
+
+def _execute_callback_sdk_case(
+    case: OcrParityCase,
+    route: SDKRoute,
+    terminal: Literal["success", "failure"],
+    mock_url: str,
+    event_loop: asyncio.AbstractEventLoop,
+) -> WorkerSuccess:
+    callback: Final = RecordingCallback()
+    call_kwargs: Final = {
+        **_call_kwargs(case.litellm_input, mock_url, route),
+        "callbacks": [callback],
+        "litellm_call_id": _callback_call_id(route, terminal),
+        "litellm_trace_id": _callback_call_id(route, terminal),
+        "metadata": _callback_metadata(route, terminal),
+    }
+    report: Final = _execute_sdk_call(call_kwargs, route, event_loop)
+    _drain_callback_delivery(route, event_loop)
+    return WorkerSuccess(report=report, callbacks=callback.observations())
+
+
+def _assert_callback_lifecycle(
+    execution: Execution,
+    route: SDKRoute,
+    terminal: Literal["success", "failure"],
+) -> None:
+    observations: Final = execution.callbacks
+    assert observations is not None, f"{route.value} {terminal} callbacks were not observed"
+    expected_hooks: Final = (
+        (f"log_{terminal}_event",)
+        if route is SDKRoute.OCR
+        else ("async_log_success_event",)
+        if terminal == "success"
+        else ("async_log_failure_event", "log_failure_event")
+    )
+    actual_hooks: Final = tuple(observation.hook for observation in observations)
+    assert actual_hooks == expected_hooks, (
+        f"{route.value} {terminal} expected callback hooks {expected_hooks}, received {actual_hooks}"
+    )
+    expected_call_id: Final = _callback_call_id(route, terminal)
+    expected_metadata: Final = _callback_metadata(route, terminal)
+    for observation in observations:
+        assert observation.phase == terminal
+        assert observation.call_type == route.value
+        assert observation.litellm_call_id == expected_call_id
+        assert observation.metadata == expected_metadata
+        assert observation.model
+        if terminal == "success":
+            assert isinstance(execution.report, SDKSuccess)
+            assert observation.payload == execution.report.response
+            assert observation.error is None
+        else:
+            assert isinstance(execution.report, SDKError)
+            assert observation.payload is None
+            assert observation.error is not None
+            assert observation.error.exception_type
+            assert observation.error.message
+            assert observation.error.status_code is not None
+            assert observation.error.status_code >= 400
+
+
+def _check_callback_ocr_sdk_parity(
+    case: OcrParityCase,
+    route: SDKRoute,
+    terminal: Literal["success", "failure"],
+    case_file: Path,
+    runner: SubprocessRunner,
+) -> None:
+    with execution_worker_pair(runner, PYTHON_VARIANT, RUST_VARIANT) as workers:
+        python_worker, rust_worker = workers
+        python: Final = python_worker.execute(case_file, route.value, case.provider_responses)
+        rust: Final = rust_worker.execute(case_file, route.value, case.provider_responses)
+
+    _assert_callback_lifecycle(python, route, terminal)
+    _assert_callback_lifecycle(rust, route, terminal)
+    assert_parity(python, rust, PYTHON_HTTP_SENTINEL)
+
+
 def _check_recorded_ocr_sdk_parity(
     ocr_fixture: OcrParityCase,
     route: SDKRoute,
@@ -276,6 +525,25 @@ def _write_worker_case(directory: Path, index: int, case: OcrWorkerCase) -> Path
     return case_file
 
 
+def _callback_fixture(
+    fixtures: tuple[OcrParityCase, ...],
+    terminal: Literal["success", "failure"],
+) -> OcrParityCase:
+    matching: Final = tuple(
+        fixture
+        for fixture in fixtures
+        if fixture.litellm_input.contract == "mistral"
+        and (
+            all(response.status_code < 400 for response in fixture.provider_responses)
+            if terminal == "success"
+            else any(response.status_code >= 400 for response in fixture.provider_responses)
+        )
+    )
+    if not matching:
+        raise AssertionError(f"no recorded Mistral OCR {terminal} fixture is available for callback parity")
+    return min(matching, key=lambda fixture: fixture_id(fixture.litellm_input, fixture.litellm_input.model))
+
+
 @contextmanager
 def parity_checks() -> Generator[tuple[E2ECheck, ...]]:
     fixtures: Final = tuple(
@@ -298,6 +566,17 @@ def parity_checks() -> Generator[tuple[E2ECheck, ...]]:
             _write_worker_case(directory, len(recorded_files) + index, InvalidOcrWorkerCase(case=case))
             for index, case in enumerate(INVALID_OCR_CASES)
         )
+        callback_cases: Final[tuple[tuple[Literal["success", "failure"], OcrParityCase], ...]] = tuple(
+            (terminal, _callback_fixture(fixtures, terminal)) for terminal in CALLBACK_TERMINALS
+        )
+        callback_files: Final = tuple(
+            _write_worker_case(
+                directory,
+                len(recorded_files) + len(invalid_files) + index,
+                CallbackOcrWorkerCase(case=case, terminal=terminal),
+            )
+            for index, (terminal, case) in enumerate(callback_cases)
+        )
         with execution_worker_pair(runner, PYTHON_VARIANT, RUST_VARIANT) as workers:
             recorded: Final = tuple(
                 E2ECheck(
@@ -315,7 +594,15 @@ def parity_checks() -> Generator[tuple[E2ECheck, ...]]:
                 for case, case_file in zip(INVALID_OCR_CASES, invalid_files, strict=True)
                 for route in SDKRoute
             )
-            yield (*recorded, *invalid)
+            callbacks: Final = tuple(
+                E2ECheck(
+                    f"callback:{route.value}:{terminal}",
+                    partial(_check_callback_ocr_sdk_parity, case, route, terminal, case_file, runner),
+                )
+                for (terminal, case), case_file in zip(callback_cases, callback_files, strict=True)
+                for route in SDKRoute
+            )
+            yield (*recorded, *invalid, *callbacks)
 
 
 def _execute_worker_command(
@@ -333,6 +620,8 @@ def _execute_worker_command(
                 return WorkerSuccess(report=_execute_sdk_case(recorded.litellm_input, route, mock_url, event_loop))
             case InvalidOcrWorkerCase(case=invalid):
                 return WorkerSuccess(report=_execute_invalid_sdk_case(invalid, route, mock_url, event_loop))
+            case CallbackOcrWorkerCase(case=callback_case, terminal=terminal):
+                return _execute_callback_sdk_case(callback_case, route, terminal, mock_url, event_loop)
     except Exception:
         return WorkerFailure(error=traceback.format_exc())
 
diff --git a/tests/rust-python-harness/strategies/unit_tests_mapping/cases/ocr.py b/tests/rust-python-harness/strategies/unit_tests_mapping/cases/ocr.py
index 3e6c4060134..0e771f0dc17 100644
--- a/tests/rust-python-harness/strategies/unit_tests_mapping/cases/ocr.py
+++ b/tests/rust-python-harness/strategies/unit_tests_mapping/cases/ocr.py
@@ -230,6 +230,10 @@ _HOST_ONLY_BRIDGE_EXCLUSIONS: Final = tuple(
             "test_ocr_exception_type_uses_resolved_provider_context",
             "Python wraps bridge exceptions into public errors.",
         ),
+        (
+            "test_rust_upstream_error_uses_ocr_provider_error_mapping",
+            "Python maps native upstream errors through the selected OCR provider config.",
+        ),
         ("test_aocr_routes_to_async_rust_when_enabled", "Python selects and invokes the async native bridge."),
         ("test_aocr_exception_type_uses_resolved_provider_context", "Python wraps async bridge exceptions."),
         ("test_ocr_forwards_timeout_to_rust", "Python converts and forwards explicit timeouts."),
diff --git a/tests/test_litellm/ocr/test_rust_bridge.py b/tests/test_litellm/ocr/test_rust_bridge.py
index d28184b96a6..c34833221cc 100644
--- a/tests/test_litellm/ocr/test_rust_bridge.py
+++ b/tests/test_litellm/ocr/test_rust_bridge.py
@@ -9,6 +9,7 @@ import httpx
 import pytest
 
 import litellm
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.base_llm.ocr.transformation import OCRResponse
 from litellm.rust_bridge import configuration
 
@@ -39,6 +40,10 @@ class CapturedException(Exception):
     pass
 
 
+class RustUpstreamError(Exception):
+    pass
+
+
 class RecordingBridge:
     """A fake ``RustOcr`` callable that records the args it was handed."""
 
@@ -183,6 +188,9 @@ class FakeOCRConfig:
     ) -> str:
         return f"{api_base or 'https://api.mistral.ai/v1'}/ocr"
 
+    def get_error_class(self, error_message: str, status_code: int, headers: dict[str, str]) -> BaseLLMException:
+        return BaseLLMException(status_code=status_code, message=error_message, headers=headers)
+
 
 def build_prepared_request(
     *,
@@ -486,6 +494,20 @@ def test_run_rust_ocr_prepares_request_and_wraps_response():
     }
 
 
+def test_rust_upstream_error_uses_ocr_provider_error_mapping():
+    error = RustUpstreamError(400, '{"message":"invalid model"}')
+
+    mapped = ocr_main._map_rust_ocr_error(
+        error,
+        build_prepared_request(),
+        (RuntimeError, RustUpstreamError),
+    )
+
+    assert isinstance(mapped, BaseLLMException)
+    assert mapped.status_code == 400
+    assert mapped.message == '{"message":"invalid model"}'
+
+
 def test_run_rust_ocr_resolves_key_via_secret_manager_when_missing():
     bridge = RecordingBridge()
     litellm.rust(True)

From 4b3355bdc6b9807d5d1c817094449df03b6af881 Mon Sep 17 00:00:00 2001
From: yuneng-jiang <yuneng@berri.ai>
Date: Mon, 7 Sep 2026 11:30:46 -0700
Subject: [PATCH 72/79] test(e2e): prove the virtual key lifecycle on every
 gateway replica (#40023)

* test(e2e): prove the virtual key lifecycle on every replica

Walks one virtual key through create, read, partial update, clear, enforce
and delete against a live proxy and database, reading every write back on
every gateway replica.

The management suite already had single write-then-read tests for keys, but
none of them proved that a partial /key/update leaves the untouched fields
alone, that an explicit null clears a field, or that a write is visible on
more than the one gateway that took it.

Adds read_back_everywhere to the shared ProxyClient: it polls a GET path on
every URL in PROXY_REPLICA_URLS until each replica's parsed body satisfies
the caller's predicate, and fails naming the replica that never converged.
The CLEAR sentinel in the e2e models makes an explicit JSON null expressible
in a body the transport otherwise strips of None fields.

Documents /key/update's merge patch semantics on the endpoint docstring.

* test(e2e): prove key revocation and field preservation on every replica

Applies the findings from an adversarial review of the first commit.

The delete step only checked that chat was refused on the gateway that took
the write, so it would have passed while a sibling gateway kept serving the
deleted key. It now serves one call from every replica first, so each has the
key cached and the delete has something to revoke everywhere, then polls every
replica for the refusal.

The file also carried its own poll loop that tested the deadline before
attempting, so it gave up one attempt early and skipped the attempt landing
exactly on the deadline. It now shares the harness helper, which is generic
over the polled value rather than over a parsed body, so the same loop covers
both the info read-back and the chat refusal.

The model the enforcement step registers now carries a unique marker in its
alias, matching every other deployment this suite creates, so concurrent runs
never share one model group.

The docstring sentence claimed an explicit null clears any field. It does not:
the metadata-backed fields merge into stored metadata, where a null is a silent
no-op, and only the key's own columns clear. Regenerating the dashboard types
picks up the corrected text.

* fix(e2e): delete a deployment that never becomes servable

Registering a model posts /model/new and then waits for every replica to list
it. When that wait timed out the deployment already existed in the database but
its id had never been returned, so no caller could delete it and the row
outlived the run. It is now deleted before the failure propagates.

Found by review on the key lifecycle suite, whose module fixture registers a
deployment this way, but every caller of the shared helper had the same
exposure.

* docs(e2e): drop the duplicated notes from the lifecycle docstrings

The delete method restated what the warm-up helper already explains, and the
module restated the merge patch rule that the endpoint and the request model
both document.
---
 .../key_management_endpoints.py               |   5 +
 tests/e2e/coverage_registry/mgmt.yaml         |   2 +
 .../e2e/management/test_key_lifecycle_e2e.py  | 299 ++++++++++++++++++
 tests/e2e/models.py                           |  36 ++-
 tests/e2e/proxy_client.py                     | 152 ++++++++-
 tests/e2e/test_proxy_client.py                | 114 ++++++-
 ui/litellm-dashboard/src/lib/http/schema.d.ts |   5 +
 7 files changed, 592 insertions(+), 21 deletions(-)
 create mode 100644 tests/e2e/management/test_key_lifecycle_e2e.py

diff --git a/litellm/proxy/management_endpoints/key_management_endpoints.py b/litellm/proxy/management_endpoints/key_management_endpoints.py
index ebcfab090b5..8e51e250319 100644
--- a/litellm/proxy/management_endpoints/key_management_endpoints.py
+++ b/litellm/proxy/management_endpoints/key_management_endpoints.py
@@ -2840,6 +2840,11 @@ async def update_key_fn(
     """
     Update an existing API key's parameters.
 
+    The body is a merge patch: a field left out keeps its stored value, and on the key's own columns
+    an explicit null clears it. The metadata-backed fields below are the exception, merging into the
+    stored metadata instead: passing one as null leaves it unchanged, while `metadata` itself
+    replaces the stored metadata wholesale.
+
     Parameters:
     - key: Optional[str] - The key to update. Either key or key_alias must be provided.
     - key_alias: Optional[str] - User-friendly key alias. If key is omitted, also identifies the key to update (must match exactly one key, same as /key/delete's key_aliases)
diff --git a/tests/e2e/coverage_registry/mgmt.yaml b/tests/e2e/coverage_registry/mgmt.yaml
index d571fb36546..860d96a50b4 100644
--- a/tests/e2e/coverage_registry/mgmt.yaml
+++ b/tests/e2e/coverage_registry/mgmt.yaml
@@ -5,6 +5,8 @@
 - {id: mgmt.key.update.persists, module: mgmt, tier: P0, surface: api, assertions: [persists], source: "key_management_endpoints.py:2462", rationale: "Budget/model changes persist"}
 - {id: mgmt.key.update.admin_only, module: mgmt, tier: P0, surface: api, assertions: [admin_only], source: "key_management_endpoints.py:2462", rationale: "Non-admin cannot escalate perms"}
 - {id: mgmt.key.update.happy_path, module: mgmt, tier: P1, surface: ui, assertions: [happy_path], source: "key_management_endpoints.py:2462", rationale: "Key edit through the dashboard"}
+- {id: mgmt.key.update.preserves_unrelated_fields, module: mgmt, tier: P0, surface: api, assertions: [preserves_unrelated_fields], source: "key_management_endpoints.py:2829", rationale: "A partial /key/update changes only the field it names; alias, models, limits, budget window, team and metadata read back unchanged on every gateway replica"}
+- {id: mgmt.key.update.clear_persists, module: mgmt, tier: P0, surface: api, assertions: [clear_persists], source: "key_management_endpoints.py:2829", rationale: "An explicit null on /key/update clears max_budget and budget_duration, and the derived budget_reset_at with it, on every gateway replica"}
 - {id: mgmt.key.delete.persists, module: mgmt, tier: P0, surface: api, assertions: [persists], source: "key_management_endpoints.py:3122", rationale: "Deletion revokes future calls"}
 - {id: mgmt.key.delete.admin_only, module: mgmt, tier: P0, surface: api, assertions: [admin_only], source: "key_management_endpoints.py:3122", rationale: "Non-owner cannot delete"}
 - {id: mgmt.key.info.persists, module: mgmt, tier: P0, surface: api, assertions: [persists], source: "key_management_endpoints.py:3380", rationale: "Info reflects all writes"}
diff --git a/tests/e2e/management/test_key_lifecycle_e2e.py b/tests/e2e/management/test_key_lifecycle_e2e.py
new file mode 100644
index 00000000000..4c8effc4d24
--- /dev/null
+++ b/tests/e2e/management/test_key_lifecycle_e2e.py
@@ -0,0 +1,299 @@
+"""Live e2e: one virtual key walked through its whole lifecycle, read back on every
+gateway replica.
+
+Create, read, partial update, clear, enforce, delete: one method per step, and every
+step creates its own team and key (both deleted on teardown) so a step reruns or skips
+on its own. Writes go through the control plane; read-backs poll every URL in
+PROXY_REPLICA_URLS until each replica converges, because a write that is visible on the
+gateway that took it and stale on its neighbour is exactly the failure this file exists
+to catch. Revocation is the slowest of those: a deleted key stays usable on the other
+replicas until their auth cache entry expires, so the delete step polls each of them
+rather than asserting once.
+"""
+
+from __future__ import annotations
+
+import time
+from collections.abc import Callable, Iterator, Mapping
+from dataclasses import dataclass
+from types import MappingProxyType
+from typing import Final
+
+import pytest
+
+from e2e_config import unique_marker
+from e2e_http import Result, StreamingResponse, Success, UnknownApiError, unwrap
+from lifecycle import ResourceManager
+from management_client import MODEL_ACCESS_DENIED_MARKER, ManagementClient
+from models import (
+    CLEAR,
+    ChatBody,
+    ChatMessage,
+    KeyGenerateBody,
+    KeyGenerateResponse,
+    KeyInfo,
+    KeyInfoParams,
+    KeyInfoResponse,
+    KeyMetadata,
+    KeyUpdateBody,
+    LiteLLMParamsBody,
+    TeamNewBody,
+)
+from proxy_client import Converged, NotConverged, Poller, await_converged, await_converged_everywhere
+from transport import Transport
+
+pytestmark = pytest.mark.e2e
+
+BACKING_MODEL: Final = "gpt-4o-mini"
+DENIED_MODEL: Final = "gpt-5.5"
+MAX_BUDGET: Final = 25.0
+TPM_LIMIT: Final = 313131
+RPM_LIMIT: Final = 323232
+UPDATED_RPM_LIMIT: Final = 424242
+BUDGET_DURATION: Final = "30d"
+
+
+@dataclass(frozen=True, slots=True)
+class CreatedKey:
+    written: KeyGenerateBody
+    response: KeyGenerateResponse
+
+    @property
+    def key(self) -> str:
+        return self.response.key
+
+
+@pytest.fixture(scope="module")
+def mock_deployment(client: ManagementClient) -> Iterator[str]:
+    """A deployment that answers from a canned response, so the enforcement step needs no
+    provider key. The alias carries a unique marker, like every other model this suite
+    registers, so concurrent runs never share one model group."""
+    model_name: Final = f"e2e-key-lifecycle-{unique_marker()}"
+    model_id: Final = client.proxy.create_model(model_name, LiteLLMParamsBody(model=BACKING_MODEL, mock_response="ok"))
+    try:
+        yield model_name
+    finally:
+        client.proxy.delete_model(model_id)
+
+
+def _await[T](client: ManagementClient, poller: Poller[T], converged: Callable[[T], bool], failure: str) -> T:
+    outcome: Final = await_converged(
+        poller,
+        converged=converged,
+        timeout=client.proxy.poll_timeout,
+        interval=client.proxy.poll_interval,
+        now=time.monotonic,
+        sleep=time.sleep,
+    )
+    match outcome:
+        case Converged(result=result):
+            return result
+        case NotConverged(last_result=last):
+            pytest.fail(f"{failure}; last outcome: {last}")
+
+
+def _chat_poller(transport: Transport, key: str, model: str) -> Poller[StreamingResponse]:
+    return lambda: transport.send(
+        "/chat/completions",
+        headers=transport.bearer(key),
+        json=ChatBody(
+            model=model,
+            messages=[ChatMessage(role="user", content=f"say hi {unique_marker()}")],
+            max_tokens=16,
+        ),
+    )
+
+
+def _create_key(client: ManagementClient, resources: ResourceManager, model: str) -> CreatedKey:
+    marker: Final = unique_marker()
+    team_id: Final = client.create_team(TeamNewBody(team_alias=f"e2e-key-lifecycle-team-{marker}"))
+    resources.defer(lambda: client.delete_team(team_id))
+    written: Final = KeyGenerateBody(
+        key_alias=f"e2e-key-lifecycle-{marker}",
+        models=[model],
+        max_budget=MAX_BUDGET,
+        tpm_limit=TPM_LIMIT,
+        rpm_limit=RPM_LIMIT,
+        budget_duration=BUDGET_DURATION,
+        metadata=KeyMetadata(tag=marker),
+        team_id=team_id,
+    )
+    response: Final = unwrap(client.generate_key(written))
+    resources.defer(lambda: client.proxy.delete_key(response.key))
+    return CreatedKey(written=written, response=response)
+
+
+def _key_info_everywhere(
+    client: ManagementClient, key: str, settled: Callable[[KeyInfo], bool]
+) -> Mapping[str, KeyInfo]:
+    def converged(result: Result[KeyInfoResponse]) -> bool:
+        return isinstance(result, Success) and settled(result.data.info)
+
+    reads: Final = client.proxy.read_back_everywhere(
+        "/key/info", params=KeyInfoParams(key=key), response_type=KeyInfoResponse, converged=converged
+    )
+    return MappingProxyType({replica: unwrap(read).info for replica, read in reads.items()})
+
+
+def _is_key_not_found(result: Result[KeyInfoResponse]) -> bool:
+    return isinstance(result, UnknownApiError) and result.status_code == 404
+
+
+def _assert_reads_back(info: KeyInfo, expected: KeyGenerateBody, replica: str) -> None:
+    for field, observed, wanted in (
+        ("key_alias", info.key_alias, expected.key_alias),
+        ("models", info.models, expected.models),
+        ("max_budget", info.max_budget, expected.max_budget),
+        ("tpm_limit", info.tpm_limit, expected.tpm_limit),
+        ("rpm_limit", info.rpm_limit, expected.rpm_limit),
+        ("budget_duration", info.budget_duration, expected.budget_duration),
+        ("team_id", info.team_id, expected.team_id),
+        ("metadata", info.metadata, expected.metadata),
+    ):
+        assert observed == wanted, f"{replica}: /key/info reports {field}={observed!r}, expected {wanted!r}"
+
+
+def _poll_chat_ok(client: ManagementClient, key: str, model: str) -> None:
+    _ = _await(
+        client,
+        _chat_poller(client.proxy.transport, key, model),
+        lambda outcome: outcome.ok,
+        f"chat on {model} never succeeded for the key before the deadline",
+    )
+
+
+def _warm_every_replica(client: ManagementClient, key: str, model: str) -> None:
+    """Serve one call from every replica, so each has the key in its auth cache. Without
+    this the revocation check below would only prove a replica rejects a key it never
+    knew, which is true of any random string."""
+    for replica, transport in client.proxy.replicas.items():
+        _ = _await(
+            client,
+            _chat_poller(transport, key, model),
+            lambda outcome: outcome.ok,
+            f"{replica}: chat on {model} never succeeded for the key before the deadline",
+        )
+
+
+def _assert_chat_rejected_everywhere(client: ManagementClient, key: str, model: str) -> None:
+    outcomes: Final = await_converged_everywhere(
+        {replica: _chat_poller(transport, key, model) for replica, transport in client.proxy.replicas.items()},
+        converged=lambda outcome: outcome.status_code == 401,
+        timeout=client.proxy.poll_timeout,
+        interval=client.proxy.poll_interval,
+        now=time.monotonic,
+        sleep=time.sleep,
+    )
+    for replica, outcome in outcomes.items():
+        assert isinstance(outcome, Converged), (
+            f"{replica}: the deleted key was still accepted on chat after "
+            f"{client.proxy.poll_timeout}s, last status {outcome.last_result.status_code}"
+        )
+
+
+class TestKeyLifecycle:
+    def test_create_echoes_every_field_written(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        created: Final = _create_key(client, resources, mock_deployment)
+
+        response: Final = created.response
+        for field, observed, wanted in (
+            ("key_alias", response.key_alias, created.written.key_alias),
+            ("models", response.models, created.written.models),
+            ("max_budget", response.max_budget, created.written.max_budget),
+            ("tpm_limit", response.tpm_limit, created.written.tpm_limit),
+            ("rpm_limit", response.rpm_limit, created.written.rpm_limit),
+            ("budget_duration", response.budget_duration, created.written.budget_duration),
+            ("team_id", response.team_id, created.written.team_id),
+            ("metadata", response.metadata, created.written.metadata),
+        ):
+            assert observed == wanted, f"/key/generate echoed {field}={observed!r}, sent {wanted!r}"
+
+    def test_read_reflects_the_create_on_every_replica(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        created: Final = _create_key(client, resources, mock_deployment)
+
+        infos: Final = _key_info_everywhere(
+            client, created.key, lambda info: info.key_alias == created.written.key_alias
+        )
+        for replica, info in infos.items():
+            _assert_reads_back(info, created.written, replica)
+            assert info.budget_reset_at is not None, (
+                f"{replica}: /key/info reports no budget_reset_at for budget_duration={BUDGET_DURATION!r}"
+            )
+
+    @pytest.mark.covers("mgmt.key.update.preserves_unrelated_fields")
+    def test_partial_update_changes_only_the_named_field(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        created: Final = _create_key(client, resources, mock_deployment)
+        before: Final = _key_info_everywhere(client, created.key, lambda info: info.rpm_limit == RPM_LIMIT)
+
+        _ = unwrap(client.update_key(KeyUpdateBody(key=created.key, rpm_limit=UPDATED_RPM_LIMIT)))
+
+        after: Final = _key_info_everywhere(client, created.key, lambda info: info.rpm_limit == UPDATED_RPM_LIMIT)
+        for replica, info in after.items():
+            _assert_reads_back(info, created.written.model_copy(update={"rpm_limit": UPDATED_RPM_LIMIT}), replica)
+            assert info.budget_reset_at == before[replica].budget_reset_at, (
+                f"{replica}: budget_reset_at moved from {before[replica].budget_reset_at!r} to "
+                f"{info.budget_reset_at!r} on a /key/update that did not name budget_duration"
+            )
+
+    @pytest.mark.covers("mgmt.key.update.clear_persists")
+    def test_explicit_null_clears_the_budget_and_its_reset_time(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        created: Final = _create_key(client, resources, mock_deployment)
+        _ = _key_info_everywhere(client, created.key, lambda info: info.max_budget == MAX_BUDGET)
+
+        _ = unwrap(client.update_key(KeyUpdateBody(key=created.key, max_budget=CLEAR, budget_duration=CLEAR)))
+
+        cleared: Final = _key_info_everywhere(client, created.key, lambda info: info.max_budget is None)
+        for replica, info in cleared.items():
+            assert info.budget_duration is None, (
+                f"{replica}: budget_duration={info.budget_duration!r} survived an explicit null"
+            )
+            assert info.budget_reset_at is None, (
+                f"{replica}: clearing budget_duration left budget_reset_at={info.budget_reset_at!r}"
+            )
+            _assert_reads_back(
+                info, created.written.model_copy(update={"max_budget": None, "budget_duration": None}), replica
+            )
+
+    def test_key_serves_its_model_and_is_denied_others(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        created: Final = _create_key(client, resources, mock_deployment)
+
+        _poll_chat_ok(client, created.key, mock_deployment)
+
+        denied: Final = client.chat_status(created.key, DENIED_MODEL, f"say hi {unique_marker()}")
+        assert denied.status_code == 403, (
+            f"chat on {DENIED_MODEL!r} outside the key's model list must be denied 403, got "
+            f"{denied.status_code}: {denied.body[:300]}"
+        )
+        assert MODEL_ACCESS_DENIED_MARKER in denied.body, (
+            f"403 body must be a model-access denial, got: {denied.body[:300]}"
+        )
+
+    def test_delete_revokes_info_and_chat_on_every_replica(
+        self, client: ManagementClient, resources: ResourceManager, mock_deployment: str
+    ) -> None:
+        """The teardown's deferred delete fires again on the already-deleted key by
+        design: the deferred cleanup must survive this test failing before the
+        in-body delete, and a repeat /key/delete is a cheap no-op the warn-only
+        teardown absorbs."""
+        created: Final = _create_key(client, resources, mock_deployment)
+        _warm_every_replica(client, created.key, mock_deployment)
+
+        client.delete_key_strict(created.key)
+
+        _ = client.proxy.read_back_everywhere(
+            "/key/info",
+            params=KeyInfoParams(key=created.key),
+            response_type=KeyInfoResponse,
+            converged=_is_key_not_found,
+        )
+        _assert_chat_rejected_everywhere(client, created.key, mock_deployment)
diff --git a/tests/e2e/models.py b/tests/e2e/models.py
index c01687d0b31..fa124bdf6bd 100644
--- a/tests/e2e/models.py
+++ b/tests/e2e/models.py
@@ -8,9 +8,9 @@ from __future__ import annotations
 
 from collections.abc import Sequence
 from datetime import datetime
-from typing import Literal
+from typing import Final, Literal
 
-from pydantic import AliasChoices, BaseModel, ConfigDict, Field, RootModel, model_validator
+from pydantic import AliasChoices, BaseModel, ConfigDict, Field, RootModel, model_serializer, model_validator
 
 # ---------- keys ----------
 
@@ -49,6 +49,7 @@ class KeyMetadata(BaseModel):
     logging: list[KeyLoggingCallback] | None = None
     priority: str | None = None
     batch_enqueued_token_limit: int | None = None
+    tag: str | None = None
 
 
 class ObjectPermission(BaseModel):
@@ -81,6 +82,14 @@ class KeyGenerateBody(BaseModel):
 
 class KeyGenerateResponse(BaseModel):
     key: str
+    key_alias: str | None = None
+    models: list[str] = []
+    max_budget: float | None = None
+    tpm_limit: int | None = None
+    rpm_limit: int | None = None
+    budget_duration: str | None = None
+    team_id: str | None = None
+    metadata: KeyMetadata | None = None
 
 
 class KeyRegenerateBody(BaseModel):
@@ -122,6 +131,7 @@ class KeyInfo(BaseModel):
     blocked: bool | None = None
     spend: float | None = None
     max_budget: float | None = None
+    budget_duration: str | None = None
     budget_reset_at: str | None = None
     budget_id: str | None = None
     litellm_budget_table: LiteLLMBudgetTable | None = None
@@ -919,12 +929,34 @@ class CredentialCreateResponse(BaseModel):
 # ---------- key / team / user / organization management ----------
 
 
+class Cleared(BaseModel):
+    """An explicit JSON null in a merge-patch body. The transport drops `None` fields
+    before sending (`exclude_none`), so `None` means "leave the stored value alone"; a
+    field set to `CLEAR` reaches the wire as `null`, which tells the proxy to clear it."""
+
+    model_config = ConfigDict(frozen=True)
+
+    @model_serializer
+    def _as_null(self) -> None:
+        return None
+
+
+CLEAR: Final = Cleared()
+
+
 class KeyUpdateBody(BaseModel):
+    """POST /key/update is a merge patch: a field left `None` is dropped from the body and
+    keeps its stored value, `CLEAR` sends an explicit null that clears it (`budget_duration`
+    clears `budget_reset_at` with it), and `metadata` replaces the stored metadata wholesale."""
+
     key: str
     models: list[str] | None = None
     key_alias: str | None = None
     tpm_limit: int | None = None
     rpm_limit: int | None = None
+    max_budget: float | Cleared | None = None
+    budget_duration: str | Cleared | None = None
+    metadata: KeyMetadata | None = None
 
 
 class KeyBlockBody(BaseModel):
diff --git a/tests/e2e/proxy_client.py b/tests/e2e/proxy_client.py
index 3b15e57dfcf..520cbfde5a9 100644
--- a/tests/e2e/proxy_client.py
+++ b/tests/e2e/proxy_client.py
@@ -16,6 +16,8 @@ from datetime import datetime
 from types import MappingProxyType
 from typing import Final
 
+from pydantic import BaseModel
+
 from e2e_http import (
     AnthropicHeaders,
     AuthHeaders,
@@ -149,9 +151,7 @@ def await_servable(
     last_result: Result[ModelsListResponse] | None = None
     while True:
         t = now()
-        phase_deadline = (
-            started + timeout if first_seen_at is None else first_seen_at + db_sync_seconds
-        )
+        phase_deadline = started + timeout if first_seen_at is None else first_seen_at + db_sync_seconds
         remaining = phase_deadline - t
         if remaining <= 0:
             if (
@@ -164,9 +164,7 @@ def await_servable(
 
         poll_timeout = min(request_timeout, remaining)
         last_result = list_models(poll_timeout)
-        listed = isinstance(last_result, Success) and any(
-            entry.id == model_name for entry in last_result.data.data
-        )
+        listed = isinstance(last_result, Success) and any(entry.id == model_name for entry in last_result.data.data)
         t = now()
         if not listed:
             first_seen_at = None
@@ -179,9 +177,7 @@ def await_servable(
         elif t - first_seen_at >= db_sync_seconds:
             return Servable()
 
-        phase_deadline = (
-            started + timeout if first_seen_at is None else first_seen_at + db_sync_seconds
-        )
+        phase_deadline = started + timeout if first_seen_at is None else first_seen_at + db_sync_seconds
         wait = min(interval, phase_deadline - now())
         if wait > 0:
             sleep(wait)
@@ -239,6 +235,88 @@ def servable_timeout_message(
     )
 
 
+type Poller[T] = Callable[[], T]
+
+
+@dataclass(frozen=True, slots=True)
+class Converged[T]:
+    result: T
+
+
+@dataclass(frozen=True, slots=True)
+class NotConverged[T]:
+    """The deadline passed without a read satisfying the predicate; `last_result` is
+    the final read, so the caller can tell a stale body from a failed request."""
+
+    last_result: T
+
+
+type ConvergeOutcome[T] = Converged[T] | NotConverged[T]
+
+
+def await_converged[T](
+    poll: Poller[T],
+    *,
+    converged: Callable[[T], bool],
+    timeout: float,
+    interval: float,
+    now: Callable[[], float],
+    sleep: Callable[[float], None],
+) -> ConvergeOutcome[T]:
+    """Poll until a read satisfies `converged` or `timeout` elapses.
+
+    Polls before testing the deadline, so a zero or already-spent budget still gets one
+    attempt, and sleeps only min(interval, time left), so the attempt that lands exactly
+    on the deadline is taken rather than skipped. Clock and sleep are injected."""
+    deadline: Final = now() + timeout
+    while True:
+        result = poll()
+        if converged(result):
+            return Converged(result=result)
+        remaining = deadline - now()
+        if remaining <= 0:
+            return NotConverged(last_result=result)
+        sleep(min(interval, remaining))
+
+
+def await_converged_everywhere[T](
+    pollers: Mapping[str, Poller[T]],
+    *,
+    converged: Callable[[T], bool],
+    timeout: float,
+    interval: float,
+    now: Callable[[], float],
+    sleep: Callable[[float], None],
+) -> Mapping[str, ConvergeOutcome[T]]:
+    """`await_converged` against every replica in turn, each with the full budget, so a
+    replica that lags behind the one a write landed on is polled until it catches up
+    rather than failing on its first stale read."""
+    return MappingProxyType(
+        {
+            replica: await_converged(
+                poll, converged=converged, timeout=timeout, interval=interval, now=now, sleep=sleep
+            )
+            for replica, poll in pollers.items()
+        }
+    )
+
+
+def first_lagging_replica[T](
+    outcomes: Mapping[str, ConvergeOutcome[T]],
+) -> tuple[str, NotConverged[T]] | None:
+    return next(
+        ((replica, outcome) for replica, outcome in outcomes.items() if isinstance(outcome, NotConverged)),
+        None,
+    )
+
+
+def converge_timeout_message(*, what: str, replica: str, timeout: float, last_result: object) -> str:
+    return (
+        f"{what} on {replica} never converged within {timeout}s of the write "
+        f"(control/data-plane propagation issue); last read: {last_result}"
+    )
+
+
 @dataclass(frozen=True, slots=True)
 class ProxyClient:
     transport: Transport
@@ -290,6 +368,52 @@ class ProxyClient:
             )
         ).info
 
+    def read_back_everywhere[R: BaseModel](
+        self,
+        path: str,
+        *,
+        params: BaseModel,
+        response_type: type[R],
+        converged: Callable[[Result[R]], bool],
+    ) -> Mapping[str, Result[R]]:
+        """GET `path` under the master key on every replica in PROXY_REPLICA_URLS (the
+        data-plane URL alone when the stack exports no per-gateway addresses), polling
+        each to poll_timeout until its read satisfies `converged`. Returns that read per
+        replica, or fails naming the first replica that never converged and its last
+        read. Behind a load balancer the single address proves one replica converged,
+        not all of them; only per-gateway addresses make this a fleet-wide proof."""
+        outcomes: Final = await_converged_everywhere(
+            {
+                url: self._body_poller(transport, path, params, response_type)
+                for url, transport in self.replicas.items()
+            },
+            converged=converged,
+            timeout=self.poll_timeout,
+            interval=self.poll_interval,
+            now=time.monotonic,
+            sleep=time.sleep,
+        )
+        lagging: Final = first_lagging_replica(outcomes)
+        if lagging is not None:
+            replica, outcome = lagging
+            raise AssertionError(
+                converge_timeout_message(
+                    what=f"GET {path}",
+                    replica=replica,
+                    timeout=self.poll_timeout,
+                    last_result=outcome.last_result,
+                )
+            )
+        return MappingProxyType(
+            {replica: outcome.result for replica, outcome in outcomes.items() if isinstance(outcome, Converged)}
+        )
+
+    @staticmethod
+    def _body_poller[R: BaseModel](
+        transport: Transport, path: str, params: BaseModel, response_type: type[R]
+    ) -> Poller[Result[R]]:
+        return lambda: transport.get(path, headers=transport.master, params=params, response_type=response_type)
+
     def model_info(self) -> list[ModelInfoEntry]:
         """Every configured deployment with the price the proxy resolved for it
         (config override merged over cost-map defaults)."""
@@ -320,9 +444,7 @@ class ProxyClient:
             response_type=FileListResponse,
         )
 
-    def list_fine_tuning_jobs(
-        self, key: str, params: FineTuningJobsParams
-    ) -> Result[FineTuningJobsResponse]:
+    def list_fine_tuning_jobs(self, key: str, params: FineTuningJobsParams) -> Result[FineTuningJobsResponse]:
         return self.transport.get(
             "/v1/fine_tuning/jobs",
             headers=self.transport.bearer(key),
@@ -375,7 +497,11 @@ class ProxyClient:
             )
         ).model_id
         written_at = time.monotonic()
-        self._await_model_servable(body.model_name, listed_for)
+        try:
+            self._await_model_servable(body.model_name, listed_for)
+        except BaseException:
+            self.delete_model(model_id)
+            raise
         settle_propagation(written_at)
         return model_id
 
diff --git a/tests/e2e/test_proxy_client.py b/tests/e2e/test_proxy_client.py
index a508c97b9fb..2caac58333f 100644
--- a/tests/e2e/test_proxy_client.py
+++ b/tests/e2e/test_proxy_client.py
@@ -1,9 +1,11 @@
-"""Harness coverage for the model barrier that gates on every replica.
+"""Harness coverage for the barriers that gate on every replica.
 
 No proxy needed and no ``e2e`` marker: this pins that a model registered through
 the control plane only counts as servable once every configured replica lists it
-on /v1/models, which is what keeps a two-gateway stack from handing a test a
-model that one gateway has not reloaded yet. The fakes are plain pollers and an
+on /v1/models, and that a management write only counts as read back once every
+replica's read satisfies the caller's predicate, which is what keeps a two-gateway
+stack from handing a test a model or a key that one gateway has not caught up on
+yet. The fakes are plain pollers standing in for each replica's transport plus an
 injected clock, so nothing here monkeypatches anything.
 """
 
@@ -12,18 +14,33 @@ from __future__ import annotations
 from collections.abc import Iterable, Mapping
 from dataclasses import dataclass
 from itertools import chain, repeat
+from types import MappingProxyType
 from typing import Final
 
 import pytest
 
 from e2e_config import parse_replica_urls
-from e2e_http import Success
-from models import ModelListEntry, ModelsListResponse
-from proxy_client import ModelsPoller, NotServableOn, Servable, await_servable_everywhere
+from e2e_http import Result, Success
+from models import KeyInfo, KeyInfoResponse, ModelListEntry, ModelsListResponse
+from proxy_client import (
+    Poller,
+    ConvergeOutcome,
+    Converged,
+    ModelsPoller,
+    NotConverged,
+    NotServableOn,
+    Servable,
+    await_converged_everywhere,
+    await_servable_everywhere,
+    first_lagging_replica,
+    converge_timeout_message,
+)
 
 MODEL: Final = "gpt-under-test"
 TIMEOUT: Final = 10.0
 INTERVAL: Final = 2.0
+RPM_BEFORE_UPDATE: Final = 100
+RPM_AFTER_UPDATE: Final = 200
 
 
 @dataclass
@@ -78,6 +95,91 @@ class TestAwaitServableEverywhere:
         assert _await(pollers) == Servable()
 
 
+def _key_info(rpm_limit: int) -> Success[KeyInfoResponse]:
+    return Success(status_code=200, data=KeyInfoResponse(info=KeyInfo(rpm_limit=rpm_limit)))
+
+
+def _reads(results: Iterable[Result[KeyInfoResponse]]) -> Poller[Result[KeyInfoResponse]]:
+    it: Final = iter(results)
+    return lambda: next(it)
+
+
+def _updated(result: Result[KeyInfoResponse]) -> bool:
+    return isinstance(result, Success) and result.data.info.rpm_limit == RPM_AFTER_UPDATE
+
+
+def _converge(
+    pollers: Mapping[str, Poller[Result[KeyInfoResponse]]], clock: FakeClock
+) -> Mapping[str, ConvergeOutcome[Result[KeyInfoResponse]]]:
+    return await_converged_everywhere(
+        pollers,
+        converged=_updated,
+        timeout=TIMEOUT,
+        interval=INTERVAL,
+        now=clock.now,
+        sleep=clock.sleep,
+    )
+
+
+class TestAwaitConvergedEverywhere:
+    def test_waits_for_the_replica_that_lags_behind_the_write(self) -> None:
+        clock: Final = FakeClock()
+        pollers: Final = MappingProxyType(
+            {
+                "gateway-1": _reads(repeat(_key_info(RPM_AFTER_UPDATE))),
+                "gateway-2": _reads(
+                    chain(repeat(_key_info(RPM_BEFORE_UPDATE), 2), repeat(_key_info(RPM_AFTER_UPDATE)))
+                ),
+            }
+        )
+        outcomes: Final = _converge(pollers, clock)
+        assert outcomes == {
+            "gateway-1": Converged(result=_key_info(RPM_AFTER_UPDATE)),
+            "gateway-2": Converged(result=_key_info(RPM_AFTER_UPDATE)),
+        }
+        assert first_lagging_replica(outcomes) is None
+        assert clock.elapsed == 2 * INTERVAL
+
+    def test_names_the_replica_that_never_converges_with_its_last_read(self) -> None:
+        clock: Final = FakeClock()
+        pollers: Final = MappingProxyType(
+            {
+                "gateway-1": _reads(repeat(_key_info(RPM_AFTER_UPDATE))),
+                "gateway-2": _reads(repeat(_key_info(RPM_BEFORE_UPDATE))),
+            }
+        )
+        outcomes: Final = _converge(pollers, clock)
+        assert first_lagging_replica(outcomes) == (
+            "gateway-2",
+            NotConverged(last_result=_key_info(RPM_BEFORE_UPDATE)),
+        )
+        assert clock.elapsed == TIMEOUT
+        message: Final = converge_timeout_message(
+            what="GET /key/info",
+            replica="gateway-2",
+            timeout=TIMEOUT,
+            last_result=_key_info(RPM_BEFORE_UPDATE),
+        )
+        assert "gateway-2" in message and "/key/info" in message and str(RPM_BEFORE_UPDATE) in message
+
+    def test_each_replica_gets_its_own_full_budget(self) -> None:
+        """A replica that converges late must not eat into the next replica's budget: both
+        need most of the timeout here, so one shared deadline would starve the second."""
+        clock: Final = FakeClock()
+        slow: Final = chain(repeat(_key_info(RPM_BEFORE_UPDATE), 3), repeat(_key_info(RPM_AFTER_UPDATE)))
+        pollers: Final = MappingProxyType(
+            {
+                "gateway-1": _reads(slow),
+                "gateway-2": _reads(
+                    chain(repeat(_key_info(RPM_BEFORE_UPDATE), 3), repeat(_key_info(RPM_AFTER_UPDATE)))
+                ),
+            }
+        )
+        outcomes: Final = _converge(pollers, clock)
+        assert first_lagging_replica(outcomes) is None
+        assert clock.elapsed == 2 * 3 * INTERVAL
+
+
 class TestParseReplicaUrls:
     def test_splits_and_trims_the_gateway_addresses(self) -> None:
         raw: Final = " http://127.0.0.1:4010/, http://127.0.0.1:4011 "
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 108e50b52a1..36fd744efc0 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -8017,6 +8017,11 @@ export interface paths {
          * Update Key Fn
          * @description Update an existing API key's parameters.
          *
+         *     The body is a merge patch: a field left out keeps its stored value, and on the key's own columns
+         *     an explicit null clears it. The metadata-backed fields below are the exception, merging into the
+         *     stored metadata instead: passing one as null leaves it unchanged, while `metadata` itself
+         *     replaces the stored metadata wholesale.
+         *
          *     Parameters:
          *     - key: Optional[str] - The key to update. Either key or key_alias must be provided.
          *     - key_alias: Optional[str] - User-friendly key alias. If key is omitted, also identifies the key to update (must match exactly one key, same as /key/delete's key_aliases)

From 96c698030bdc9978e925e0fe01e93636f9878442 Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 7 Sep 2026 11:38:53 -0700
Subject: [PATCH 73/79] fix(spend_logs): keep partition DDL transactions alive
 for their statement timeout (#40098)

* fix(spend_logs): keep partition DDL transactions alive for their statement timeout

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* style: ruff format changed files

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(lint): avoid dict-literal kwargs and keep cast-ok on the cast line

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(lint): cast at the call site instead of widening PrismaClient.tx

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* test(spend_logs): require partition tx timeout to strictly exceed statement bound

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: jesus <jesus@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../spend_logs_partition_manager.py           | 30 ++++++++--
 .../test_spend_logs_partition_manager.py      | 57 ++++++++++++++-----
 2 files changed, 69 insertions(+), 18 deletions(-)

diff --git a/litellm/proxy/db/db_transaction_queue/spend_logs_partition_manager.py b/litellm/proxy/db/db_transaction_queue/spend_logs_partition_manager.py
index 221c142d9d3..b756bfeb6f6 100644
--- a/litellm/proxy/db/db_transaction_queue/spend_logs_partition_manager.py
+++ b/litellm/proxy/db/db_transaction_queue/spend_logs_partition_manager.py
@@ -16,7 +16,12 @@ keeps the batched-DELETE path, so existing deployments are untouched.
 import re
 from collections.abc import Callable
 from datetime import date, datetime, timedelta, timezone
-from typing import TYPE_CHECKING, Final, TypeAlias
+from typing import (
+    TYPE_CHECKING,
+    Final,
+    TypeAlias,
+    cast,  # noqa: TID251  # db.tx is reached through untyped __getattr__ delegation
+)
 
 from litellm._logging import verbose_proxy_logger
 from litellm.constants import (
@@ -25,6 +30,8 @@ from litellm.constants import (
 )
 
 if TYPE_CHECKING:
+    from prisma.client import TransactionManager
+
     from litellm.proxy.utils import PrismaClient
 
 SPEND_LOGS_TABLE: Final = "LiteLLM_SpendLogs"
@@ -116,6 +123,21 @@ def select_partitions_to_drop(partitions: list[tuple[str, datetime | None]], cut
     return [name for name, upper in partitions if upper is not None and upper <= cutoff]
 
 
+_TX_COMMIT_SLACK: Final = timedelta(seconds=5)
+
+
+def _bounded_tx(prisma_client: "PrismaClient", timeout_ms: int) -> "TransactionManager":
+    """
+    Open an interactive transaction that outlives the statement bound it
+    carries. prisma's default 5s transaction timeout would close it mid
+    lock-wait, after which the engine answers the next call with a 422.
+    """
+    return cast(  # cast-ok: PrismaWrapper delegates tx via __getattr__ (untyped)
+        "TransactionManager",
+        prisma_client.db.tx(timeout=timedelta(milliseconds=timeout_ms) + _TX_COMMIT_SLACK),
+    )
+
+
 class SpendLogsPartitionManager:
     def __init__(
         self,
@@ -137,7 +159,7 @@ class SpendLogsPartitionManager:
         if budget_ms is None:
             return False
         try:
-            async with prisma_client.db.tx() as tx:
+            async with _bounded_tx(prisma_client, budget_ms) as tx:
                 await tx.execute_raw(f"SET LOCAL statement_timeout = {budget_ms}")
                 rows: Final = await tx.query_raw(
                     """
@@ -172,7 +194,7 @@ class SpendLogsPartitionManager:
         wait for the lock and statement_timeout bounds the work itself, so a
         partition this run cannot get is simply left for the next one.
         """
-        async with prisma_client.db.tx() as tx:
+        async with _bounded_tx(prisma_client, timeout_ms) as tx:
             await tx.execute_raw(f"SET LOCAL statement_timeout = {timeout_ms}")
             await tx.execute_raw(f"SET LOCAL lock_timeout = {timeout_ms}")
             await tx.execute_raw(statement)
@@ -209,7 +231,7 @@ class SpendLogsPartitionManager:
     async def _list_partitions(
         self, prisma_client: "PrismaClient", timeout_ms: int
     ) -> list[tuple[str, datetime | None]]:
-        async with prisma_client.db.tx() as tx:
+        async with _bounded_tx(prisma_client, timeout_ms) as tx:
             await tx.execute_raw(f"SET LOCAL statement_timeout = {timeout_ms}")
             rows: Final = await tx.query_raw(
                 """
diff --git a/tests/test_litellm/proxy/db/db_transaction_queue/test_spend_logs_partition_manager.py b/tests/test_litellm/proxy/db/db_transaction_queue/test_spend_logs_partition_manager.py
index 4ea655b8871..609dd13afc2 100644
--- a/tests/test_litellm/proxy/db/db_transaction_queue/test_spend_logs_partition_manager.py
+++ b/tests/test_litellm/proxy/db/db_transaction_queue/test_spend_logs_partition_manager.py
@@ -4,7 +4,7 @@ selection, the non-partitioned no-op safety path, and the drop/ensure SQL flow.
 """
 
 from contextlib import asynccontextmanager
-from datetime import date, datetime, timezone
+from datetime import date, datetime, timedelta, timezone
 from unittest.mock import AsyncMock, MagicMock
 
 import pytest
@@ -19,7 +19,6 @@ from litellm.proxy.db.db_transaction_queue.spend_logs_partition_manager import (
     upcoming_partitions,
 )
 
-
 DDL_TIMEOUT_MS = 30000
 
 
@@ -28,19 +27,23 @@ def _budget(ms: "int | None" = DDL_TIMEOUT_MS):
     return lambda: ms
 
 
-def _wire_tx(db) -> list[str]:
+def _wire_tx(db) -> "tuple[list[str], list[int | timedelta | None]]":
     """
     Model the prisma seam the partition DDL uses.
 
     Every statement this manager issues, DDL and catalog query alike, runs inside
     db.tx() so it can carry SET LOCAL timeouts. Those SET LOCAL statements are
-    collected in the returned list rather than forwarded, so assertions on
-    db.execute_raw and db.query_raw still see only the real statements.
+    collected in the first returned list rather than forwarded, so assertions on
+    db.execute_raw and db.query_raw still see only the real statements. The
+    second list records the interactive-transaction timeout each tx was opened
+    with.
     """
     session_settings: list[str] = []
+    tx_timeouts: list[int | timedelta | None] = []
 
     @asynccontextmanager
-    async def _tx():
+    async def _tx(*, max_wait: "int | timedelta | None" = None, timeout: "int | timedelta | None" = None):
+        tx_timeouts.append(timeout)
         tx = MagicMock()
 
         async def _execute_raw(sql, *args):
@@ -57,7 +60,7 @@ def _wire_tx(db) -> list[str]:
         yield tx
 
     db.tx = _tx
-    return session_settings
+    return session_settings, tx_timeouts
 
 
 def test_period_start_per_interval():
@@ -227,7 +230,7 @@ async def test_partition_ddl_carries_a_statement_and_lock_timeout():
             }
         ]
     )
-    session_settings = _wire_tx(client.db)
+    session_settings, _ = _wire_tx(client.db)
 
     await mgr.ensure_partitions(client, _budget(7000))
     await mgr.drop_partitions_older_than(client, datetime(2026, 6, 5, tzinfo=timezone.utc), _budget(7000))
@@ -249,7 +252,7 @@ async def test_catalog_queries_carry_a_statement_timeout():
     mgr = SpendLogsPartitionManager()
     client = MagicMock()
     client.db.query_raw = AsyncMock(return_value=[])
-    session_settings = _wire_tx(client.db)
+    session_settings, _ = _wire_tx(client.db)
 
     await mgr.is_partitioned(client, _budget(4000))
     assert session_settings == ["SET LOCAL statement_timeout = 4000"], (
@@ -263,6 +266,34 @@ async def test_catalog_queries_carry_a_statement_timeout():
     )
 
 
+@pytest.mark.asyncio
+async def test_partition_transactions_outlive_their_statement_bound():
+    """
+    prisma's interactive transaction has its own timeout, 5s by default, which
+    keeps ticking while a statement waits on the partition lock. A tx shorter
+    than the SET LOCAL bound it carries is closed mid-lock-wait and the engine
+    then answers the next call with a 422, so the partition is silently not
+    created. A tx equal to the bound is closed too: the statement can consume
+    its whole bound waiting on the lock, then still needs to commit.
+    """
+    mgr = SpendLogsPartitionManager()
+    client = MagicMock()
+    client.db.execute_raw = AsyncMock(return_value=0)
+    client.db.query_raw = AsyncMock(return_value=[])
+    _, tx_timeouts = _wire_tx(client.db)
+
+    await mgr.is_partitioned(client, _budget())
+    await mgr.ensure_partitions(client, _budget())
+    await mgr.drop_partitions_older_than(client, datetime(2026, 6, 5, tzinfo=timezone.utc), _budget())
+
+    assert len(tx_timeouts) > 0
+    for tx_timeout in tx_timeouts:
+        assert isinstance(tx_timeout, timedelta)
+        assert tx_timeout > timedelta(milliseconds=DDL_TIMEOUT_MS), (
+            f"tx timeout {tx_timeout} does not outlive its {DDL_TIMEOUT_MS}ms statement bound"
+        )
+
+
 @pytest.mark.asyncio
 async def test_partition_loops_stop_when_the_budget_runs_out_mid_way():
     """
@@ -308,17 +339,15 @@ async def test_partition_maintenance_issues_nothing_when_the_budget_is_already_s
 
 
 def test_unsupported_interval_raises():
-    with pytest.raises(ValueError, match='Unsupported partition interval: year'):
+    with pytest.raises(ValueError, match="Unsupported partition interval: year"):
         period_start(date(2026, 6, 1), "year")
-    with pytest.raises(ValueError, match='Unsupported partition interval: year'):
+    with pytest.raises(ValueError, match="Unsupported partition interval: year"):
         next_period_start(date(2026, 6, 1), "year")
 
 
 def test_parse_partition_upper_bound_unparseable_to_value_is_none():
     """A TO(...) value that is not a valid timestamp must not raise; return None."""
-    assert (
-        parse_partition_upper_bound("FOR VALUES FROM ('x') TO ('not-a-date')") is None
-    )
+    assert parse_partition_upper_bound("FOR VALUES FROM ('x') TO ('not-a-date')") is None
 
 
 @pytest.mark.asyncio

From 55fe4a78943711531bcd35fbc30ddd8ff3f30f3b Mon Sep 17 00:00:00 2001
From: Roman D <roman.dovgan@justeattakeaway.com>
Date: Mon, 7 Sep 2026 19:54:59 +0100
Subject: [PATCH 74/79] feat(proxy): resolve root_path per request from a
 configured prefix list (SERVER_ROOT_PATHS) (#35935)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* feat(proxy): resolve root_path per request from SERVER_ROOT_PATHS

One deployment can encode exactly one client-visible URL path prefix
today: SERVER_ROOT_PATH is a scalar stamped onto the app at startup, so
a pod fronting several ingress prefixes 404s every prefix but one before
any handler runs, and MCP OAuth discovery can emit only one prefix's
URLs (RFC 9728 section 3 exact-match fails for the rest).

Add an opt-in outermost ASGI middleware that matches the request path
against a configured prefix list (SERVER_ROOT_PATHS, comma-separated) on
a segment boundary and sets scope["root_path"] for that request only.
Everything downstream is stock Starlette: route matching strips
root_path so routes stay registered root-relative, and request.base_url
re-includes it, so the discovery documents' resource and the 401
challenges' resource_metadata land under the prefix the client actually
called — with no discovery-builder changes.

LazyFeatureMiddleware now strips the scope root_path (falling back to
the cached SERVER_ROOT_PATH scalar) before feature prefix matching, so
lazily-registered routers — the MCP OAuth discovery router among them —
load under per-request prefixes.

Follow-up to the routing discussion on #35226; composes with, but does
not depend on, #35576.

* fix(proxy): import Sequence from collections.abc (ruff UP035 strict-budget gate)

* review(greptile): trim implementation commentary; fixture-own MCP registry state in tests

Addresses both P2s from the first Greptile pass:
- per_request_root_path_middleware.py (and the related _lazy_features /
  proxy_server comments) cut down to the constraints the code cannot
  express, per repo comment guidance
- the new discovery tests no longer clear/repopulate the shared MCP
  registry inline; a fixture snapshots it, hands the test an empty
  registry, and restores it afterwards so no state leaks between cases

* fix(lint): mutable-ok marker on the prefix accumulator (LIT002 type-discipline gate)

* fix(proxy): tie 401 challenges and get_custom_url to the per-request root_path

The per-request root_path middleware sets scope["root_path"] to the
prefix the client actually called, but the OAuth 401 challenges
(raise_user_oauth_challenge / raise_token_exchange_challenge) still
built their resource_metadata from SERVER_ROOT_PATH. On a pod fronting
several prefixes, the challenge advertised a discovery URL under a
different prefix than the discovery document served — the two
disagreed on where the resource metadata lives, and a strict RFC 9728
client refused the challenge. Route the challenges through a small
ContextVar the middleware populates so they read the same effective
root_path Starlette resolves the request under.

The same accessor fixes get_custom_url: when a request lives under a
SERVER_ROOT_PATHS-matched prefix, request.base_url already carries it,
so appending the SERVER_ROOT_PATH scalar on top produced e.g.
/tenant-a/legacy/sso/callback — a path that does not exist. Reading
the per-request prefix instead (and relying on join_paths's tail-dedup)
keeps SSO login/callback URLs under one prefix — the one the request
actually arrived on.

Fallback: outside a request (module-load-time UI URL builders,
background tasks) the ContextVar is unset and the accessor reads
SERVER_ROOT_PATH, matching get_server_root_path() so scalar-only
deployments are byte-identical.

* fix(mcp): challenge URL under per-request prefix must route, and mock parity

Two follow-ups to the review fix that made the 401 challenge use the
per-request root_path:

1. oauth_protected_resource_path must pick the URL structure that
   actually routes for the mechanism in use:
   - The scalar SERVER_ROOT_PATH deployment registers the well-known
     routes with the prefix INSERTED (via well_known_root_suffix at
     import time), matching RFC 8414 §3. The challenge URL must use the
     same insertion or a client fetching it 404s.
   - The per-request SERVER_ROOT_PATHS deployment can't register routes
     per prefix; PerRequestRootPathMiddleware strips the prefix from
     scope["path"] and the router matches the un-inserted route. The
     URL must place the prefix BEFORE .well-known so the strip leaves a
     matching path.
   The previous fix used the insertion form for both, which 404'd the
   discovery fetch on the per-request path — the discovery doc and the
   challenge would then disagree on where the resource metadata lives,
   the very failure the review flagged. End-to-end verified: the URL
   the challenge advertises routes and the doc's `resource` field
   equals the URL the client originally called (RFC 9728 §3).

2. get_request_root_path now delegates its fallback through
   get_server_root_path() instead of reading the env directly, so every
   existing `monkeypatch.setattr("litellm.proxy.utils.get_server_root_path"`
   test override keeps working. This unstubbed the mock on the /v2/login
   test that failed on the last CI run.

Plus the lint budget: annotate the local accumulator Final, tag the
scope["root_path"] rewrite as an intentional ASGI-contract mutation,
tag the reused `path`/`root_path` rebinds in LazyFeatureMiddleware, and
add reason strings to the two new PLC0415 lazy-import noqas.

* test(mcp): pin the reviewer's expected end-state — challenge URL routes, resource matches called URL

End-to-end regression test that mounts the discoverable router + the
per-request root_path middleware, hits an MCP endpoint that raises
raise_user_oauth_challenge, fetches the resource_metadata URL the
challenge advertises, and checks the returned document's `resource`
equals the URL the client originally called (RFC 9728 §3 exact match).

Covers /tenant-a, /tenant-b, and the unprefixed path on the same app so
a regression on any prefix — challenge URL 404s, or doc emits a
different prefix than the client called — fails at this test rather
than in a strict MCP client's discovery.

---------

Co-authored-by: gym-cmd <186399764+gym-cmd@users.noreply.github.com>
---
 .../mcp_server/mcp_server_manager.py          |  13 +-
 .../outbound_credentials/adapter.py           |  26 +-
 .../proxy/_experimental/mcp_server/server.py  |   8 +-
 litellm/proxy/_lazy_features.py               |  19 +-
 .../per_request_root_path_middleware.py       | 122 +++++++
 litellm/proxy/proxy_server.py                 |  20 ++
 litellm/proxy/utils.py                        |  14 +-
 .../outbound_credentials/test_adapter.py      |  86 ++++-
 .../mcp_server/test_discoverable_endpoints.py | 224 +++++++++++++
 .../test_per_request_root_path_middleware.py  | 302 ++++++++++++++++++
 tests/test_litellm/proxy/test_proxy_server.py |  76 +++++
 .../proxy/utils/helpers/test_url_helpers.py   |  73 +++++
 12 files changed, 948 insertions(+), 35 deletions(-)
 create mode 100644 litellm/proxy/middleware/per_request_root_path_middleware.py
 create mode 100644 tests/test_litellm/proxy/middleware/test_per_request_root_path_middleware.py

diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
index 612596bc803..e7fd650a324 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
@@ -163,7 +163,10 @@ from litellm.proxy.common_utils.user_api_key_cache import get_management_object_
 from litellm.proxy.management_endpoints.sso.id_jag_assertion_capture import (
     id_jag_assertion_capture_gap_at_startup,
 )
-from litellm.proxy.utils import PrismaClient, ProxyLogging, get_server_root_path
+from litellm.proxy.middleware.per_request_root_path_middleware import (
+    get_request_root_path,
+)
+from litellm.proxy.utils import PrismaClient, ProxyLogging
 from litellm.repositories.table_repositories import MCPServerRepository
 from litellm.types.llms.custom_http import httpxSpecialProvider
 from litellm.types.mcp import (
@@ -3858,7 +3861,7 @@ class MCPServerManager:
                 if err.tag == "unauthorized" and isinstance(spec.config, AuthorizationCodeConfig):
                     # authorization_code's missing per-user token -> the per-server browser-OAuth
                     # challenge, built here where the full MCPServer is in hand.
-                    raise_user_oauth_challenge(server, root_path=get_server_root_path())
+                    raise_user_oauth_challenge(server, root_path=get_request_root_path())
                 if err.tag == "unauthorized" and isinstance(spec.config, TokenExchangeConfig):
                     # token_exchange (OBO): a missing/rejected subject token -> the RFC 9728 challenge
                     # pointing at the IdP the client must SSO with to obtain one, rather than an opaque
@@ -3866,7 +3869,7 @@ class MCPServerManager:
                     # Access) threads its claims blob into the challenge for the client to satisfy.
                     raise_token_exchange_challenge(
                         server,
-                        root_path=get_server_root_path(),
+                        root_path=get_request_root_path(),
                         claims=err.unauthorized.claims,
                     )
                 raise_public(err)
@@ -3914,7 +3917,7 @@ class MCPServerManager:
         if spec is None or not isinstance(spec.config, (TokenExchangeConfig, IdJagConfig)):
             return
         if subject_token is None and isinstance(spec.config, TokenExchangeConfig):
-            raise_token_exchange_challenge(resolved_server, root_path=get_server_root_path())
+            raise_token_exchange_challenge(resolved_server, root_path=get_request_root_path())
         match await self._cred_provider.resolve_credentials(to_subject(user_api_key_auth, subject_token), spec):
             case Ok(_):
                 return
@@ -3922,7 +3925,7 @@ class MCPServerManager:
                 if err.tag == "unauthorized" and isinstance(spec.config, TokenExchangeConfig):
                     raise_token_exchange_challenge(
                         resolved_server,
-                        root_path=get_server_root_path(),
+                        root_path=get_request_root_path(),
                         claims=err.unauthorized.claims,
                     )
                 raise_public(err)
diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py
index ea2318bd6f1..77979a15199 100644
--- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py
+++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/adapter.py
@@ -12,6 +12,7 @@ every other mode so the caller defers to v1 (parity-safe); it grows one branch p
 from __future__ import annotations
 
 import base64
+import os
 from typing import TYPE_CHECKING, Final, Literal, NoReturn
 
 from fastapi import HTTPException
@@ -310,13 +311,30 @@ def raise_public(error: CredError) -> NoReturn:
 def oauth_protected_resource_path(root_path: str, server: MCPServer) -> str:
     """The server's RFC 9728 Protected Resource Metadata path, the shared anchor of both challenges.
 
-    ``root_path`` is the proxy's ``SERVER_ROOT_PATH``, resolved by the caller (the imperative shell)
-    so this stays a pure function of its inputs; ``"/"`` and ``""`` both mean no prefix. The path is
-    relative, so it resolves against the caller's own host (correct even behind a reverse proxy).
+    ``root_path`` is the prefix the request was routed under, resolved by the caller (the imperative
+    shell); ``"/"`` and ``""`` both mean no prefix. The path is relative, so it resolves against the
+    caller's own host (correct even behind a reverse proxy).
+
+    URL structure depends on how the prefix is served:
+
+    - The scalar ``SERVER_ROOT_PATH`` deployment registers the well-known routes with the prefix
+      *inserted* into the path (via :func:`well_known_root_suffix` at import time), matching RFC 8414
+      §3 well-known path insertion. When ``root_path`` equals ``SERVER_ROOT_PATH`` the URL must use
+      the same insertion or a client fetching it 404s.
+    - The per-request ``SERVER_ROOT_PATHS`` deployment can't register routes per prefix (the prefix
+      set is dynamic and could contain many entries); the middleware strips the prefix from
+      ``scope["path"]`` and the router matches the un-inserted well-known route. The URL must place
+      the prefix *before* ``.well-known`` so the strip leaves a matching path.
+
+    Picking the wrong form 404s the client's discovery fetch — the discovery document and the 401
+    challenge would then disagree on where the resource metadata lives.
     """
     prefix: Final = "" if root_path == "/" else root_path
     name: Final = server.alias or server.server_name or server.name or server.server_id
-    return f"/.well-known/oauth-protected-resource{prefix}/mcp/{name}"
+    scalar_env: Final = os.getenv("SERVER_ROOT_PATH", "").rstrip("/")
+    if not prefix or (scalar_env and prefix == scalar_env):
+        return f"/.well-known/oauth-protected-resource{prefix}/mcp/{name}"
+    return f"{prefix}/.well-known/oauth-protected-resource/mcp/{name}"
 
 
 def raise_user_oauth_challenge(server: MCPServer, *, root_path: str) -> NoReturn:
diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py
index 26f5d6e7c8c..60a9af89cc3 100644
--- a/litellm/proxy/_experimental/mcp_server/server.py
+++ b/litellm/proxy/_experimental/mcp_server/server.py
@@ -3848,12 +3848,14 @@ if MCP_AVAILABLE:
             # then exchanges. A tool-call-time 401 would be wrapped into a JSON-RPC error and the
             # header lost, so the discovery flow needs this pre-emptive challenge.
             if server and server.auth_type == MCPAuth.oauth2_token_exchange and not oauth2_headers:
-                from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (  # noqa: PLC0415
+                from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (  # noqa: PLC0415  # lazy: adapter pulls MCP subgraph
                     raise_token_exchange_challenge,
                 )
-                from litellm.proxy.utils import get_server_root_path  # noqa: PLC0415
+                from litellm.proxy.middleware.per_request_root_path_middleware import (  # noqa: PLC0415  # lazy: middleware imports proxy utils
+                    get_request_root_path,
+                )
 
-                raise_token_exchange_challenge(server, root_path=get_server_root_path())
+                raise_token_exchange_challenge(server, root_path=get_request_root_path())
 
             # Exchange-backed modes (token_exchange's OBO mint, id_jag's stored-assertion mint): run
             # the exchange here at the transport edge, so a rejected subject raises the RFC 9728
diff --git a/litellm/proxy/_lazy_features.py b/litellm/proxy/_lazy_features.py
index 3f90e6c0a7a..50e0a961a49 100644
--- a/litellm/proxy/_lazy_features.py
+++ b/litellm/proxy/_lazy_features.py
@@ -295,15 +295,18 @@ class LazyFeatureMiddleware:
         # Short-circuit once every feature has loaded.
         if scope["type"] in ("http", "websocket") and len(self._loaded) < len(self._features):
             path = scope.get("path", "")
-            # Strip SERVER_ROOT_PATH so prefix matching works under a server
-            # root path. Without this, requests like /api/v1/policies/... never
-            # match the registered prefixes (/policies/...) and lazy features
-            # stay unloaded — every endpoint under them returns 404. The
+            # Strip the request's root_path so prefix matching works under a
+            # server root path. Without this, requests like /api/v1/policies/...
+            # never match the registered prefixes (/policies/...) and lazy
+            # features stay unloaded — every endpoint under them returns 404.
+            # scope["root_path"] wins over the cached env scalar: FastAPI
+            # stamps SERVER_ROOT_PATH there, and PerRequestRootPathMiddleware
+            # resolves SERVER_ROOT_PATHS prefixes there per request. The
             # `+ "/"` boundary prevents false-positive matches (e.g. /apiv2
-            # against root /api). If the path doesn't start with the prefix
-            # (e.g. a reverse proxy already stripped it), we leave it alone.
-            if self._root_path and path.startswith(self._root_path + "/"):
-                path = path[len(self._root_path) :]
+            # against root /api); a pre-stripped path is left alone.
+            root_path: Final = str(scope.get("root_path", "")).rstrip("/") or self._root_path
+            if root_path and path.startswith(root_path + "/"):
+                path = path[len(root_path) :]  # rebind-ok: local strip after the boundary check above
             for feat in self._features:
                 if feat.module_path in self._loaded:
                     continue
diff --git a/litellm/proxy/middleware/per_request_root_path_middleware.py b/litellm/proxy/middleware/per_request_root_path_middleware.py
new file mode 100644
index 00000000000..d5df91458d2
--- /dev/null
+++ b/litellm/proxy/middleware/per_request_root_path_middleware.py
@@ -0,0 +1,122 @@
+"""Per-request ``root_path`` resolution from ``SERVER_ROOT_PATHS``.
+
+``SERVER_ROOT_PATH`` is a startup scalar, so one deployment serves exactly one
+client-visible URL path prefix; a request under any other prefix 404s before a
+handler runs. When the ingress preserves several prefixes into one pod (e.g.
+``/tenant-a/*`` and ``/tenant-b/*``), the matched prefix becomes that
+request's ``scope["root_path"]`` instead: Starlette strips it during route
+matching and rebuilds it into ``request.base_url``, so every emitted URL —
+the MCP OAuth discovery ``resource`` (RFC 9728 §3) and the 401 challenges'
+``resource_metadata`` among them — lands under the prefix the client called.
+Opt-in: with ``SERVER_ROOT_PATHS`` unset the middleware is not added at all.
+"""
+
+import os
+from collections.abc import Sequence
+from contextvars import ContextVar
+from typing import Final
+
+from starlette.types import ASGIApp, Receive, Scope, Send
+
+from litellm._logging import verbose_proxy_logger
+
+SERVER_ROOT_PATHS_ENV: Final = "SERVER_ROOT_PATHS"
+
+# The effective ``root_path`` for the currently-handled request. Populated by
+# ``PerRequestRootPathMiddleware`` from the (possibly-mutated) scope so code
+# that emits URLs off the request path — the 401 challenges' resource_metadata
+# and ``get_custom_url``'s SSO callbacks among them — can pick up the prefix
+# the client actually called without threading scope through every call site.
+# ``None`` means "middleware did not run" (the ``SERVER_ROOT_PATHS`` env is
+# unset, so no per-request prefix exists); readers fall back to the scalar
+# ``SERVER_ROOT_PATH`` in that case, which matches the pre-middleware behavior.
+_request_root_path_var: Final[ContextVar[str | None]] = ContextVar("_request_root_path_var", default=None)
+
+
+def get_request_root_path() -> str:
+    """Return the effective ``root_path`` for the current request.
+
+    Reads the value ``PerRequestRootPathMiddleware`` stashed for this request;
+    falls back through :func:`~litellm.proxy.utils.get_server_root_path` (i.e.
+    the ``SERVER_ROOT_PATH`` env) when the middleware did not run — the
+    scalar-only deployment. Delegating to the existing helper keeps every
+    existing ``monkeypatch.setattr("litellm.proxy.utils.get_server_root_path"``
+    test override working, and keeps a single source of truth for the scalar.
+    """
+    value: Final = _request_root_path_var.get()
+    if value is not None:
+        return value
+    # Lazy import: utils.py imports this module (via the lazy import inside
+    # get_custom_url), so a top-level import would build a cycle at load time.
+    from litellm.proxy.utils import get_server_root_path  # noqa: PLC0415  # lazy import breaks a two-way dep
+
+    return get_server_root_path()
+
+
+def normalize_root_paths(raw_paths: Sequence[str]) -> tuple[str, ...]:
+    """Strip whitespace and trailing slashes, dedupe, order longest-first;
+    warn and drop entries missing a leading ``/`` and the bare root."""
+    kept: Final[list[str]] = []  # mutable-ok: local accumulator; escapes only as a tuple
+    for entry in raw_paths:
+        candidate = entry.strip()
+        if not candidate:
+            continue
+        if not candidate.startswith("/"):
+            verbose_proxy_logger.warning(
+                "%s entry %r does not start with '/' and will be ignored.",
+                SERVER_ROOT_PATHS_ENV,
+                entry,
+            )
+            continue
+        candidate = candidate.rstrip("/")
+        if not candidate:
+            verbose_proxy_logger.warning(
+                "%s entry %r is the bare root and will be ignored; a root-mounted deployment needs no entry.",
+                SERVER_ROOT_PATHS_ENV,
+                entry,
+            )
+            continue
+        if candidate not in kept:
+            kept.append(candidate)
+    return tuple(sorted(kept, key=len, reverse=True))
+
+
+def get_server_root_paths() -> tuple[str, ...]:
+    """The normalized ``SERVER_ROOT_PATHS`` prefixes, empty when unset."""
+    configured: Final = os.getenv(SERVER_ROOT_PATHS_ENV, "")
+    if not configured.strip():
+        return ()
+    return normalize_root_paths(configured.split(","))
+
+
+class PerRequestRootPathMiddleware:
+    """Sets ``scope["root_path"]`` to the configured prefix matching the
+    request path on a whole-segment boundary. ``scope["path"]`` is left
+    untouched (Starlette strips ``root_path`` at route-match time). Must be
+    the outermost middleware so inner middlewares and the router see the
+    resolved value; a matched prefix overrides a scalar ``SERVER_ROOT_PATH``
+    for that request.
+    """
+
+    def __init__(self, app: ASGIApp, root_paths: Sequence[str]) -> None:
+        self.app = app
+        self.root_paths: Final = normalize_root_paths(root_paths)
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if scope["type"] in ("http", "websocket"):
+            path: Final = scope.get("path", "")
+            for prefix in self.root_paths:
+                if path == prefix or path.startswith(prefix + "/"):
+                    scope["root_path"] = prefix  # rebind-ok: ASGI middleware contract; Router and base_url read it
+                    break
+            # Stash the effective root_path (matched prefix, or the scope's
+            # existing value when nothing matched — i.e. FastAPI's scalar
+            # SERVER_ROOT_PATH) so code that emits URLs off the request path
+            # picks the same prefix the router will resolve the request under.
+            token: Final = _request_root_path_var.set(str(scope.get("root_path", "")))
+            try:
+                await self.app(scope, receive, send)
+            finally:
+                _request_root_path_var.reset(token)
+            return
+        await self.app(scope, receive, send)
diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index ba5714fe950..0915b8dd1b9 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -598,6 +598,10 @@ from litellm.proxy.middleware.admission_control_middleware import (
 from litellm.proxy.middleware.in_flight_requests_middleware import (
     InFlightRequestsMiddleware,
 )
+from litellm.proxy.middleware.per_request_root_path_middleware import (
+    PerRequestRootPathMiddleware,
+    get_server_root_paths,
+)
 from litellm.proxy.middleware.prometheus_auth_middleware import PrometheusAuthMiddleware
 from litellm.proxy.middleware.request_size_limit_middleware import (
     RequestSizeLimitMiddleware,
@@ -18384,6 +18388,22 @@ app.add_middleware(
     get_settings=lambda: get_admission_control_settings(general_settings),
     state=admission_control_state,
 )
+# Added last on purpose - last-added is outermost, and the client-visible URL
+# prefix must be resolved into scope["root_path"] before any inner middleware
+# or the router inspects the path. Only added when SERVER_ROOT_PATHS is
+# configured, so the default deployment's middleware stack is unchanged.
+_server_root_paths: Final = get_server_root_paths()
+if _server_root_paths:
+    if server_root_path and server_root_path != "/":
+        verbose_proxy_logger.warning(
+            "Both SERVER_ROOT_PATH=%r and SERVER_ROOT_PATHS=%r are set. A request "
+            "matching a SERVER_ROOT_PATHS prefix overrides the scalar root_path for "
+            "that request; unmatched requests keep SERVER_ROOT_PATH. Configure one "
+            "mechanism or the other.",
+            server_root_path,
+            _server_root_paths,
+        )
+    app.add_middleware(PerRequestRootPathMiddleware, root_paths=_server_root_paths)
 
 
 async def _stream_mcp_asgi_response(handle_fn, scope: dict, receive) -> "StreamingResponse":
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index f1103ce6a29..ddf31cb1d8a 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -7212,7 +7212,19 @@ def get_custom_url(request_base_url: str, route: str | None = None) -> str:
     else:
         base_url = request_base_url
 
-    server_root_path: Final = get_server_root_path()
+    # get_request_root_path() returns the prefix the router is actually
+    # resolving this request under: the matched SERVER_ROOT_PATHS entry when
+    # PerRequestRootPathMiddleware ran, otherwise the SERVER_ROOT_PATH scalar.
+    # This keeps the emitted URL under one prefix — the one the client called —
+    # instead of stacking the scalar onto a request already living under a
+    # dynamic prefix (which would produce /tenant-a/legacy/... — a path that
+    # doesn't exist). join_paths()'s tail-dedup then collapses the append when
+    # base_url (i.e. request.base_url) already ends in the same prefix.
+    from litellm.proxy.middleware.per_request_root_path_middleware import (  # noqa: PLC0415  # lazy: middleware imports utils
+        get_request_root_path,
+    )
+
+    server_root_path: Final = get_request_root_path()
     if route is not None:
         if server_root_path != "":
             # First join base_url with server_root_path, then with route
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py
index d67d0df4d0e..1b003e11993 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_adapter.py
@@ -476,16 +476,47 @@ def test_raise_public_plain_unauthorized_has_no_challenge():
 
 
 @pytest.mark.parametrize(
-    "root_path, expected_prefix",
+    "root_path",
     [
-        ("/", ""),  # "/" means no prefix
-        ("", ""),  # empty means no prefix
-        ("/api/v1", "/api/v1"),  # a real root path is prepended verbatim
+        "/",  # "/" means no prefix
+        "",  # empty means no prefix
     ],
 )
-def test_oauth_protected_resource_path_honors_root_path(root_path, expected_prefix):
+def test_oauth_protected_resource_path_no_prefix(root_path, monkeypatch):
+    monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
     path = oauth_protected_resource_path(root_path, _server(alias="my-srv"))
-    assert path == f"/.well-known/oauth-protected-resource{expected_prefix}/mcp/my-srv"
+    assert path == "/.well-known/oauth-protected-resource/mcp/my-srv"
+
+
+def test_oauth_protected_resource_path_scalar_prefix_uses_rfc8414_insertion(monkeypatch):
+    # A scalar SERVER_ROOT_PATH deployment registers the well-known routes with
+    # the prefix inserted (via well_known_root_suffix at import time). The URL
+    # must match that insertion or a client fetching it 404s.
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/api/v1")
+    path = oauth_protected_resource_path("/api/v1", _server(alias="my-srv"))
+    assert path == "/.well-known/oauth-protected-resource/api/v1/mcp/my-srv"
+
+
+def test_oauth_protected_resource_path_per_request_prefix_goes_before_wellknown(monkeypatch):
+    # Per-request deployment: SERVER_ROOT_PATHS matched /tenant-a for this
+    # request but the scalar SERVER_ROOT_PATH is unset. Routes were registered
+    # without the well-known insertion, so the URL must place the prefix
+    # *before* .well-known — PerRequestRootPathMiddleware strips it and the
+    # router matches the un-inserted route.
+    monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+    path = oauth_protected_resource_path("/tenant-a", _server(alias="my-srv"))
+    assert path == "/tenant-a/.well-known/oauth-protected-resource/mcp/my-srv"
+
+
+def test_oauth_protected_resource_path_dynamic_prefix_wins_over_scalar(monkeypatch):
+    # Both env vars configured: the middleware matched a SERVER_ROOT_PATHS
+    # prefix (/tenant-a) that differs from the scalar (/legacy). The URL must
+    # advertise /tenant-a — the prefix the client called — with no /legacy
+    # segment stacked onto it. Same review-fix invariant get_custom_url pins.
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+    path = oauth_protected_resource_path("/tenant-a", _server(alias="my-srv"))
+    assert path == "/tenant-a/.well-known/oauth-protected-resource/mcp/my-srv"
+    assert "/legacy" not in path
 
 
 @pytest.mark.parametrize(
@@ -510,7 +541,11 @@ def test_raise_user_oauth_challenge_points_at_per_server_prm():
     )
 
 
-def test_raise_user_oauth_challenge_includes_server_root_path():
+def test_raise_user_oauth_challenge_includes_server_root_path(monkeypatch):
+    # The scalar deployment: routes are registered with the prefix inserted
+    # (via well_known_root_suffix at import time), so the challenge URL uses
+    # the RFC 8414 §3 insertion form.
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/api/v1")
     with pytest.raises(HTTPException) as exc_info:
         raise_user_oauth_challenge(_server(alias="my-srv"), root_path="/api/v1")
     assert (
@@ -519,6 +554,20 @@ def test_raise_user_oauth_challenge_includes_server_root_path():
     )
 
 
+def test_raise_user_oauth_challenge_per_request_prefix_is_routable(monkeypatch):
+    # Per-request deployment (SERVER_ROOT_PATHS matched /tenant-a): the
+    # challenge URL must place /tenant-a before .well-known so the client's
+    # discovery fetch routes through the same middleware strip the original
+    # request went through. The scalar-inserted form would 404 here.
+    monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+    with pytest.raises(HTTPException) as exc_info:
+        raise_user_oauth_challenge(_server(alias="my-srv"), root_path="/tenant-a")
+    assert (
+        exc_info.value.headers["WWW-Authenticate"]
+        == 'Bearer resource_metadata="/tenant-a/.well-known/oauth-protected-resource/mcp/my-srv"'
+    )
+
+
 def test_raise_token_exchange_challenge_is_rfc9728_invalid_token():
     from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (
         raise_token_exchange_challenge,
@@ -535,17 +584,30 @@ def test_raise_token_exchange_challenge_is_rfc9728_invalid_token():
     assert "error_description=" in www
 
 
-def test_raise_token_exchange_challenge_includes_server_root_path():
+def test_raise_token_exchange_challenge_includes_server_root_path(monkeypatch):
     from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (
         raise_token_exchange_challenge,
     )
 
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/api/v1")
     with pytest.raises(HTTPException) as exc_info:
         raise_token_exchange_challenge(_server(alias="obo-srv"), root_path="/api/v1")
     www = exc_info.value.headers["WWW-Authenticate"]
     assert 'resource_metadata="/.well-known/oauth-protected-resource/api/v1/mcp/obo-srv"' in www
 
 
+def test_raise_token_exchange_challenge_per_request_prefix_is_routable(monkeypatch):
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (
+        raise_token_exchange_challenge,
+    )
+
+    monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+    with pytest.raises(HTTPException) as exc_info:
+        raise_token_exchange_challenge(_server(alias="obo-srv"), root_path="/tenant-a")
+    www = exc_info.value.headers["WWW-Authenticate"]
+    assert 'resource_metadata="/tenant-a/.well-known/oauth-protected-resource/mcp/obo-srv"' in www
+
+
 def test_raise_token_exchange_challenge_static_form_is_unchanged_without_step_up():
     from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (
         raise_token_exchange_challenge,
@@ -589,9 +651,7 @@ def test_id_jag_client_secret_maps_to_config():
     # ID-JAG asserts the user's id_token; the access_token default maps to id_token.
     assert spec.config.subject_token_type == "urn:ietf:params:oauth:token-type:id_token"
     assert isinstance(spec.config.client_auth, ClientSecretAuth)
-    assert spec.config.client_auth.client_secret.get_secret_value() == (
-        "litellm-client-secret"
-    )
+    assert spec.config.client_auth.client_secret.get_secret_value() == ("litellm-client-secret")
 
 
 def test_id_jag_private_key_maps_to_private_key_jwt_auth():
@@ -617,9 +677,7 @@ def test_id_jag_private_key_wins_over_client_secret():
 
 
 def test_id_jag_honors_explicit_subject_token_type():
-    spec = to_server_spec(
-        _id_jag_server(subject_token_type="urn:ietf:params:oauth:token-type:saml2")
-    )
+    spec = to_server_spec(_id_jag_server(subject_token_type="urn:ietf:params:oauth:token-type:saml2"))
     assert spec is not None and isinstance(spec.config, IdJagConfig)
     assert spec.config.subject_token_type == "urn:ietf:params:oauth:token-type:saml2"
 
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
index a9bb24bbef9..be4206a1faf 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
@@ -10342,6 +10342,230 @@ async def test_upstream_resource_sent_on_dcr_bridge_relay_authorize():
     assert query["client_id"] == ["caller-client"]
 
 
+# ---------------------------------------------------------------------------
+# Per-request root_path (SERVER_ROOT_PATHS / PerRequestRootPathMiddleware):
+# one app fronting several client-visible URL path prefixes, each prefix's
+# discovery documents emitting URLs under the prefix the client called
+# (RFC 9728 §3 exact-match). A scalar PROXY_BASE_URL / SERVER_ROOT_PATH can
+# encode at most one prefix per pod; these tests pin the N-prefix case.
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def _no_proxy_base_url(monkeypatch):
+    """Discovery must derive URLs from the request in these tests, so the
+    scalar env overrides are cleared."""
+    monkeypatch.delenv("PROXY_BASE_URL", raising=False)
+    monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+
+
+@pytest.fixture
+def _isolated_mcp_registry():
+    """Fixture-owned registry state: snapshot the shared registry, hand the
+    test an empty one, restore afterwards so nothing leaks between cases."""
+    from litellm.proxy._experimental.mcp_server.mcp_server_manager import (
+        global_mcp_server_manager,
+    )
+
+    saved = dict(global_mcp_server_manager.registry)
+    global_mcp_server_manager.registry.clear()
+    try:
+        yield global_mcp_server_manager.registry
+    finally:
+        global_mcp_server_manager.registry.clear()
+        global_mcp_server_manager.registry.update(saved)
+
+
+def _prefixed_discovery_client(prefixes):
+    from fastapi import FastAPI
+    from fastapi.testclient import TestClient
+
+    from litellm.proxy._experimental.mcp_server.discoverable_endpoints import router
+    from litellm.proxy.middleware.per_request_root_path_middleware import (
+        PerRequestRootPathMiddleware,
+    )
+
+    app = FastAPI()
+    app.include_router(router)
+    app.add_middleware(PerRequestRootPathMiddleware, root_paths=prefixes)
+    return TestClient(app)
+
+
+class TestPerRequestRootPathDiscovery:
+    def test_prefixed_wellknown_not_routable_without_middleware(self, _isolated_mcp_registry):
+        """Control: on a plain app (the only shape a scalar root_path can
+        express), a prefixed well-known request 404s before any discovery
+        builder runs — the routing gap this feature exists to close."""
+        from fastapi import FastAPI
+        from fastapi.testclient import TestClient
+
+        from litellm.proxy._experimental.mcp_server.discoverable_endpoints import router
+
+        server = _create_oauth2_server(server_id="srv_a", name="server_a", server_name="server_a", alias="server_a")
+        _isolated_mcp_registry[server.server_id] = server
+        app = FastAPI()
+        app.include_router(router)
+        client = TestClient(app)
+        resp = client.get("/tenant-a/.well-known/oauth-protected-resource/mcp/server_a")
+        assert resp.status_code == 404
+
+    def test_two_prefixes_one_app_each_resource_matches_the_called_url(
+        self, _no_proxy_base_url, _isolated_mcp_registry
+    ):
+        """The multi-origin case itself: two prefixes served by the same app,
+        each per-server document's ``resource`` equal to the URL its client
+        called — including the prefix."""
+        for sid, name in (("srv_a", "server_a"), ("srv_b", "server_b")):
+            _isolated_mcp_registry[sid] = _create_oauth2_server(server_id=sid, name=name, server_name=name, alias=name)
+        client = _prefixed_discovery_client(["/tenant-a", "/tenant-b"])
+
+        resp_a = client.get("/tenant-a/.well-known/oauth-protected-resource/mcp/server_a")
+        resp_b = client.get("/tenant-b/.well-known/oauth-protected-resource/mcp/server_b")
+
+        assert resp_a.status_code == 200
+        assert resp_a.json()["resource"] == "http://testserver/tenant-a/mcp/server_a"
+        assert resp_b.status_code == 200
+        assert resp_b.json()["resource"] == "http://testserver/tenant-b/mcp/server_b"
+
+        # Every URL the document advertises stays under the request's
+        # prefix, so it resolves on this same app.
+        for auth_server in resp_a.json()["authorization_servers"]:
+            assert auth_server.startswith("http://testserver/tenant-a/")
+
+    def test_unprefixed_requests_unchanged_on_the_same_app(self, _no_proxy_base_url, _isolated_mcp_registry):
+        """Backward compat on the very same app: a root request emits the
+        document byte-identical to a deployment without the middleware."""
+        server = _create_oauth2_server(server_id="srv_a", name="server_a", server_name="server_a", alias="server_a")
+        _isolated_mcp_registry[server.server_id] = server
+        client = _prefixed_discovery_client(["/tenant-a"])
+        resp = client.get("/.well-known/oauth-protected-resource/mcp/server_a")
+        assert resp.status_code == 200
+        assert resp.json()["resource"] == "http://testserver/mcp/server_a"
+
+    def test_unlisted_prefix_404s(self, _no_proxy_base_url):
+        client = _prefixed_discovery_client(["/tenant-a"])
+        assert client.get("/tenant-c/.well-known/oauth-protected-resource/mcp/server_a").status_code == 404
+
+    def test_aggregate_documents_and_as_endpoints_under_prefix(self, _no_proxy_base_url, _isolated_mcp_registry):
+        """Aggregate PRM/AS documents carry the prefix, and the advertised
+        authorize endpoint actually resolves under it — the 404 trap that
+        invalidated prefixing discovery URLs without per-request routing
+        (#35226 review round 1)."""
+        client = _prefixed_discovery_client(["/tenant-a"])
+
+        prm = client.get("/tenant-a/.well-known/oauth-protected-resource/mcp")
+        asm = client.get("/tenant-a/.well-known/oauth-authorization-server/mcp")
+
+        assert prm.status_code == 200
+        assert prm.json()["resource"] == "http://testserver/tenant-a/mcp"
+        assert prm.json()["authorization_servers"] == ["http://testserver/tenant-a/mcp"]
+
+        assert asm.status_code == 200
+        assert asm.json()["issuer"] == "http://testserver/tenant-a/mcp"
+        assert asm.json()["authorization_endpoint"] == "http://testserver/tenant-a/authorize"
+
+        # The prefixed authorize URL routes to the real handler (not 404):
+        # under per-request root_path the whole app is reachable per-prefix,
+        # so discovery may advertise prefixed AS endpoints safely.
+        assert client.get("/tenant-a/authorize").status_code != 404
+
+    def test_passthrough_challenge_metadata_url_carries_prefix(self, _no_proxy_base_url):
+        """The WWW-Authenticate resource_metadata URL a 401 advertises must
+        land under the request's prefix, or the client is bounced to a
+        document whose ``resource`` cannot match the URL it called."""
+        from litellm.proxy._experimental.mcp_server.oauth_utils import (
+            get_passthrough_resource_metadata_url,
+        )
+
+        def _scope(path, root_path=None):
+            scope = {
+                "type": "http",
+                "method": "POST",
+                "path": path,
+                "headers": [],
+                "scheme": "http",
+                "server": ("testserver", 80),
+                "client": ("1.2.3.4", 4444),
+            }
+            if root_path is not None:
+                scope["root_path"] = root_path
+            return scope
+
+        prefixed = get_passthrough_resource_metadata_url(
+            scope=_scope("/tenant-a/mcp/github", root_path="/tenant-a"),
+            server_name="github",
+        )
+        assert prefixed == "http://testserver/tenant-a/.well-known/oauth-protected-resource/mcp/github"
+
+        # Regression guard: no root_path → today's URL, unchanged.
+        bare = get_passthrough_resource_metadata_url(
+            scope=_scope("/mcp/github"),
+            server_name="github",
+        )
+        assert bare == "http://testserver/.well-known/oauth-protected-resource/mcp/github"
+
+    def test_user_oauth_challenge_url_routes_and_resource_matches_client_url(
+        self, _no_proxy_base_url, _isolated_mcp_registry
+    ):
+        """The reviewer's expected end-state, pinned end-to-end: an MCP endpoint
+        raising ``raise_user_oauth_challenge`` under a per-request prefix must
+        emit a resource_metadata URL the client can actually fetch, and the
+        document it returns must carry the same prefix the client originally
+        called. If either half breaks the client's discovery is dead."""
+        import re
+
+        from fastapi import FastAPI, HTTPException, Request
+        from fastapi.testclient import TestClient
+
+        from litellm.proxy._experimental.mcp_server.discoverable_endpoints import router
+        from litellm.proxy._experimental.mcp_server.outbound_credentials.adapter import (
+            raise_user_oauth_challenge,
+        )
+        from litellm.proxy.middleware.per_request_root_path_middleware import (
+            PerRequestRootPathMiddleware,
+            get_request_root_path,
+        )
+
+        server = _create_oauth2_server(
+            server_id="srv_a", name="server_a", server_name="server_a", alias="server_a"
+        )
+        _isolated_mcp_registry[server.server_id] = server
+
+        app = FastAPI()
+        app.include_router(router)
+
+        @app.post("/mcp/{name}")
+        def _mcp(name: str, request: Request):
+            try:
+                raise_user_oauth_challenge(server, root_path=get_request_root_path())
+            except HTTPException as exc:
+                return {"www_authenticate": exc.headers["WWW-Authenticate"]}
+
+        app.add_middleware(PerRequestRootPathMiddleware, root_paths=["/tenant-a", "/tenant-b"])
+        client = TestClient(app)
+
+        for prefix, mcp_url in (
+            ("/tenant-a", "http://testserver/tenant-a/mcp/server_a"),
+            ("/tenant-b", "http://testserver/tenant-b/mcp/server_a"),
+            ("", "http://testserver/mcp/server_a"),
+        ):
+            call = client.post(f"{prefix}/mcp/server_a")
+            assert call.status_code == 200, call.text
+            www = call.json()["www_authenticate"]
+            match = re.search(r'resource_metadata="([^"]+)"', www)
+            assert match, www
+            discovery = client.get(match.group(1))
+            # The challenge URL must route (a client that can't fetch it has
+            # no way to reach the resource metadata).
+            assert discovery.status_code == 200, (
+                f"challenge URL {match.group(1)} for prefix {prefix!r} 404s; "
+                "the client can't reach the resource metadata."
+            )
+            # And the doc's `resource` must equal the URL the client called
+            # (RFC 9728 §3 exact match): a mismatch bounces a strict client.
+            assert discovery.json()["resource"] == mcp_url, discovery.json()
+
+
 def _s256(verifier: str) -> str:
     return urlsafe_b64encode(hashlib.sha256(verifier.encode("ascii")).digest()).rstrip(b"=").decode("ascii")
 
diff --git a/tests/test_litellm/proxy/middleware/test_per_request_root_path_middleware.py b/tests/test_litellm/proxy/middleware/test_per_request_root_path_middleware.py
new file mode 100644
index 00000000000..61eea00bf7d
--- /dev/null
+++ b/tests/test_litellm/proxy/middleware/test_per_request_root_path_middleware.py
@@ -0,0 +1,302 @@
+"""Tests for PerRequestRootPathMiddleware (``SERVER_ROOT_PATHS``).
+
+One deployment fronting several client-visible URL path prefixes: the matched
+prefix becomes that request's ``root_path``, so Starlette route matching and
+``request.base_url`` — and therefore every URL the proxy emits, the MCP OAuth
+discovery documents among them — resolve under the prefix the client called.
+"""
+
+import pytest
+from fastapi import FastAPI, Request
+from fastapi.testclient import TestClient
+
+from litellm.proxy.middleware.per_request_root_path_middleware import (
+    PerRequestRootPathMiddleware,
+    get_request_root_path,
+    get_server_root_paths,
+    normalize_root_paths,
+)
+
+
+class TestNormalizeRootPaths:
+    def test_strips_whitespace_and_trailing_slash(self):
+        assert normalize_root_paths([" /tenant-a/ ", "/tenant-b"]) == (
+            "/tenant-a",
+            "/tenant-b",
+        )
+
+    def test_drops_empty_entries(self):
+        assert normalize_root_paths(["", "  ", "/tenant-a"]) == ("/tenant-a",)
+
+    def test_drops_entries_without_leading_slash(self):
+        # A typo'd entry must not silently match nothing at request time.
+        assert normalize_root_paths(["tenant-a", "/tenant-b"]) == ("/tenant-b",)
+
+    def test_drops_bare_root(self):
+        # "/" would turn every request into a root_path rewrite; a
+        # root-mounted deployment needs no entry at all.
+        assert normalize_root_paths(["/", "/tenant-a"]) == ("/tenant-a",)
+
+    def test_dedupes(self):
+        assert normalize_root_paths(["/t", "/t/", " /t "]) == ("/t",)
+
+    def test_longest_first_for_nested_prefixes(self):
+        # Longest-first ordering is what makes the most-specific nested
+        # prefix win at match time.
+        assert normalize_root_paths(["/t", "/t/deep"]) == ("/t/deep", "/t")
+
+
+class TestGetServerRootPaths:
+    def test_unset_env_is_empty(self, monkeypatch):
+        monkeypatch.delenv("SERVER_ROOT_PATHS", raising=False)
+        assert get_server_root_paths() == ()
+
+    def test_empty_env_is_empty(self, monkeypatch):
+        monkeypatch.setenv("SERVER_ROOT_PATHS", "")
+        assert get_server_root_paths() == ()
+
+    def test_comma_separated_entries(self, monkeypatch):
+        monkeypatch.setenv("SERVER_ROOT_PATHS", "/tenant-a, /tenant-b/")
+        assert get_server_root_paths() == ("/tenant-a", "/tenant-b")
+
+
+def _capture_scope_middleware(root_paths):
+    """Middleware wired to a downstream that records the scope it received."""
+    captured = {}
+
+    async def downstream(scope, receive, send):
+        captured.update(scope)
+        await send({"type": "http.response.start", "status": 200, "headers": []})
+        await send({"type": "http.response.body", "body": b""})
+
+    return PerRequestRootPathMiddleware(downstream, root_paths=root_paths), captured
+
+
+async def _run(mw, scope):
+    async def receive():
+        return {"type": "http.request", "body": b"", "more_body": False}
+
+    async def send(message):
+        pass
+
+    await mw(scope, receive, send)
+
+
+class TestPerRequestRootPathMiddleware:
+    @pytest.mark.asyncio
+    async def test_matched_prefix_becomes_root_path_path_untouched(self):
+        # Starlette's router strips root_path from the (unmodified) path at
+        # match time, so the middleware must NOT rewrite scope["path"].
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(mw, {"type": "http", "path": "/tenant-a/mcp/x", "method": "GET", "headers": []})
+        assert captured["root_path"] == "/tenant-a"
+        assert captured["path"] == "/tenant-a/mcp/x"
+
+    @pytest.mark.asyncio
+    async def test_exact_prefix_matches(self):
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(mw, {"type": "http", "path": "/tenant-a", "method": "GET", "headers": []})
+        assert captured["root_path"] == "/tenant-a"
+
+    @pytest.mark.asyncio
+    async def test_segment_boundary_prevents_sibling_match(self):
+        # /tenant-ab must not match the /tenant-a prefix.
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(mw, {"type": "http", "path": "/tenant-ab/mcp", "method": "GET", "headers": []})
+        assert "root_path" not in captured
+
+    @pytest.mark.asyncio
+    async def test_unmatched_path_untouched(self):
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(mw, {"type": "http", "path": "/chat/completions", "method": "GET", "headers": []})
+        assert "root_path" not in captured
+        assert captured["path"] == "/chat/completions"
+
+    @pytest.mark.asyncio
+    async def test_longest_nested_prefix_wins(self):
+        mw, captured = _capture_scope_middleware(["/t", "/t/deep"])
+        await _run(mw, {"type": "http", "path": "/t/deep/mcp", "method": "GET", "headers": []})
+        assert captured["root_path"] == "/t/deep"
+
+    @pytest.mark.asyncio
+    async def test_matched_prefix_overrides_scalar_root_path(self):
+        # FastAPI(root_path=SERVER_ROOT_PATH) stamps the scalar before the
+        # middleware stack runs; a matched dynamic prefix wins for that
+        # request (combining both mechanisms is warned about at startup).
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(
+            mw,
+            {"type": "http", "path": "/tenant-a/mcp", "root_path": "/legacy", "method": "GET", "headers": []},
+        )
+        assert captured["root_path"] == "/tenant-a"
+
+    @pytest.mark.asyncio
+    async def test_unmatched_request_keeps_scalar_root_path(self):
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(
+            mw,
+            {"type": "http", "path": "/legacy/mcp", "root_path": "/legacy", "method": "GET", "headers": []},
+        )
+        assert captured["root_path"] == "/legacy"
+
+    @pytest.mark.asyncio
+    async def test_websocket_scope_matched(self):
+        mw, captured = _capture_scope_middleware(["/tenant-a"])
+        await _run(mw, {"type": "websocket", "path": "/tenant-a/ws", "headers": []})
+        assert captured["root_path"] == "/tenant-a"
+
+    @pytest.mark.asyncio
+    async def test_lifespan_scope_passes_through(self):
+        called = {}
+
+        async def downstream(scope, receive, send):
+            called["scope"] = scope
+
+        mw = PerRequestRootPathMiddleware(downstream, root_paths=["/tenant-a"])
+        await _run(mw, {"type": "lifespan"})
+        assert called["scope"] == {"type": "lifespan"}
+
+
+def _routed_client(prefixes):
+    app = FastAPI()
+
+    @app.get("/where")
+    def where(request: Request):
+        return {
+            "base_url": str(request.base_url),
+            "root_path": request.scope.get("root_path", ""),
+        }
+
+    app.add_middleware(PerRequestRootPathMiddleware, root_paths=prefixes)
+    return TestClient(app)
+
+
+class TestEndToEndRouting:
+    def test_two_prefixes_route_on_one_app(self):
+        # The property a scalar SERVER_ROOT_PATH cannot provide: two
+        # client-visible prefixes served by the same app, each request
+        # reconstructing its own base URL.
+        client = _routed_client(["/tenant-a", "/tenant-b"])
+
+        resp_a = client.get("/tenant-a/where")
+        resp_b = client.get("/tenant-b/where")
+
+        assert resp_a.status_code == 200
+        assert resp_a.json() == {
+            "base_url": "http://testserver/tenant-a/",
+            "root_path": "/tenant-a",
+        }
+        assert resp_b.status_code == 200
+        assert resp_b.json() == {
+            "base_url": "http://testserver/tenant-b/",
+            "root_path": "/tenant-b",
+        }
+
+    def test_unprefixed_route_still_served(self):
+        client = _routed_client(["/tenant-a"])
+        resp = client.get("/where")
+        assert resp.status_code == 200
+        assert resp.json()["base_url"] == "http://testserver/"
+
+    def test_unlisted_prefix_404s(self):
+        client = _routed_client(["/tenant-a"])
+        assert client.get("/tenant-c/where").status_code == 404
+
+
+class TestGetRequestRootPath:
+    """``get_request_root_path`` is the accessor that plumbs the middleware's
+    resolved prefix to code that doesn't have scope in hand — the 401 challenge
+    builders in ``mcp_server_manager`` / ``server`` and ``get_custom_url`` on the
+    SSO callback path. Reading the SERVER_ROOT_PATH scalar there would emit URLs
+    under a prefix the client didn't call, and stack a second prefix onto ones it
+    did (the two review points this fixture pins)."""
+
+    def test_falls_back_to_server_root_path_env_outside_a_request(self, monkeypatch):
+        # Outside a request the middleware's ContextVar is unset. The scalar
+        # env still owns the answer, so pre-middleware call sites (module-load
+        # UI URL builders, background tasks) behave exactly as they did.
+        monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+        assert get_request_root_path() == "/legacy"
+
+    def test_returns_empty_string_when_no_env_and_no_request(self, monkeypatch):
+        monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+        assert get_request_root_path() == ""
+
+    def test_returns_matched_prefix_inside_a_request(self, monkeypatch):
+        # With both env vars set, a request matching a SERVER_ROOT_PATHS prefix
+        # must see that prefix — not the SERVER_ROOT_PATH scalar — so the URL
+        # it emits stays under the prefix the router will resolve it against.
+        monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+
+        seen: list[str] = []
+        app = FastAPI()
+
+        @app.get("/where")
+        def where():
+            seen.append(get_request_root_path())
+            return {}
+
+        app.add_middleware(PerRequestRootPathMiddleware, root_paths=["/tenant-a"])
+        client = TestClient(app)
+
+        assert client.get("/tenant-a/where").status_code == 200
+        assert seen == ["/tenant-a"]
+
+    def test_unmatched_request_falls_through_to_scope_scalar(self, monkeypatch):
+        # A request the middleware saw but did not match keeps whatever
+        # scope["root_path"] the app was mounted under (the scalar). The
+        # ContextVar still reflects the effective per-request answer, so
+        # emitted URLs and the router agree even on the fallback path.
+        monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+        seen: list[str] = []
+
+        app = FastAPI(root_path="/legacy")
+
+        @app.get("/where")
+        def where():
+            seen.append(get_request_root_path())
+            return {}
+
+        app.add_middleware(PerRequestRootPathMiddleware, root_paths=["/tenant-a"])
+        client = TestClient(app)
+
+        assert client.get("/legacy/where").status_code == 200
+        assert seen == ["/legacy"]
+
+    def test_each_request_sees_its_own_prefix(self, monkeypatch):
+        # Two sequential requests through the same app must each see the
+        # prefix they arrived under, so one tenant's client is never sent
+        # the URL of another tenant's origin.
+        monkeypatch.delenv("SERVER_ROOT_PATH", raising=False)
+        seen: list[tuple[str, str]] = []
+        app = FastAPI()
+
+        @app.get("/where")
+        def where(tag: str):
+            seen.append((tag, get_request_root_path()))
+            return {}
+
+        app.add_middleware(PerRequestRootPathMiddleware, root_paths=["/tenant-a", "/tenant-b"])
+        client = TestClient(app)
+        assert client.get("/tenant-a/where?tag=a").status_code == 200
+        assert client.get("/tenant-b/where?tag=b").status_code == 200
+        assert seen == [("a", "/tenant-a"), ("b", "/tenant-b")]
+
+    def test_context_var_reset_after_request(self, monkeypatch):
+        # A ContextVar left set after the request finishes would poison the
+        # module-load-time callers that read it lazily (they'd think they were
+        # inside a request under the last-seen prefix).
+        monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+
+        app = FastAPI()
+
+        @app.get("/where")
+        def where():
+            return {"prefix": get_request_root_path()}
+
+        app.add_middleware(PerRequestRootPathMiddleware, root_paths=["/tenant-a"])
+        client = TestClient(app)
+        assert client.get("/tenant-a/where").json() == {"prefix": "/tenant-a"}
+        # After the request finishes, the scalar-env fallback owns the answer
+        # again — nothing was left stashed from the last request's scope.
+        assert get_request_root_path() == "/legacy"
diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py
index 9393ec0f8e6..b7bb58378d4 100644
--- a/tests/test_litellm/proxy/test_proxy_server.py
+++ b/tests/test_litellm/proxy/test_proxy_server.py
@@ -9224,6 +9224,82 @@ class TestLazyFeatureMiddleware:
         else:
             assert loads == [], f"{case}: feature must not load"
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        "env_root_path,scope_root_path,request_path,should_load,case",
+        [
+            # Per-request root_path (PerRequestRootPathMiddleware under
+            # SERVER_ROOT_PATHS) with no scalar env: strip and match.
+            ("", "/tenant-a", "/tenant-a/dummy/x", True, "per-request root_path strip"),
+            # scope root_path is authoritative over the cached env scalar.
+            ("/api/v1", "/tenant-a", "/tenant-a/dummy/x", True, "scope wins over env scalar"),
+            # Boundary check still applies to the per-request value.
+            ("", "/tenant-a", "/tenant-ab/dummy/x", False, "boundary check on scope root_path"),
+            # Empty scope root_path falls back to the env scalar.
+            ("/api/v1", "", "/api/v1/dummy/x", True, "empty scope falls back to env"),
+        ],
+    )
+    async def test_per_request_root_path_handling(
+        self, monkeypatch, env_root_path, scope_root_path, request_path, should_load, case
+    ):
+        """
+        ``scope["root_path"]`` must be stripped before prefix matching when
+        set — the scalar SERVER_ROOT_PATH lands there via
+        ``FastAPI(root_path=...)``, and PerRequestRootPathMiddleware
+        (SERVER_ROOT_PATHS) resolves a per-request prefix there. Otherwise
+        lazily-registered features — the MCP OAuth discovery router among
+        them — stay unloaded under a client-visible prefix and 404.
+        """
+        from fastapi import FastAPI
+
+        from litellm.proxy._lazy_features import (
+            LazyFeature,
+            LazyFeatureMiddleware,
+        )
+
+        monkeypatch.setenv("SERVER_ROOT_PATH", env_root_path)
+
+        loads = []
+
+        def fake_register(app, module):
+            loads.append(getattr(module, "__name__", "?"))
+
+        feat = LazyFeature(
+            name=f"dummy_prr_{case}",
+            module_path="json",
+            path_prefixes=("/dummy",),
+            register_fn=fake_register,
+        )
+
+        async def downstream(scope, receive, send):
+            await send({"type": "http.response.start", "status": 200, "headers": []})
+            await send({"type": "http.response.body", "body": b""})
+
+        target_app = FastAPI()
+        mw = LazyFeatureMiddleware(downstream, fastapi_app=target_app, features=(feat,))
+
+        async def receive():
+            return {"type": "http.request", "body": b"", "more_body": False}
+
+        async def send(message):
+            pass
+
+        await mw(
+            {
+                "type": "http",
+                "path": request_path,
+                "root_path": scope_root_path,
+                "method": "GET",
+                "headers": [],
+            },
+            receive,
+            send,
+        )
+        if should_load:
+            assert loads == ["json"], f"{case}: expected feature to load"
+        else:
+            assert loads == [], f"{case}: feature must not load"
+
     @pytest.mark.asyncio
     async def test_concurrent_first_requests_only_register_once(self):
         """
diff --git a/tests/test_litellm/proxy/utils/helpers/test_url_helpers.py b/tests/test_litellm/proxy/utils/helpers/test_url_helpers.py
index 31ea1bdce74..5f23c5fc20e 100644
--- a/tests/test_litellm/proxy/utils/helpers/test_url_helpers.py
+++ b/tests/test_litellm/proxy/utils/helpers/test_url_helpers.py
@@ -314,3 +314,76 @@ def test_normalize_route_for_root_path_error_path_when_route_not_under_root(
     _clear_url_env(monkeypatch)
     monkeypatch.setenv("SERVER_ROOT_PATH", "/proxy")
     assert normalize_route_for_root_path("/other/v1/chat") is None
+
+
+# ---------------------------------------------------------------------------
+# get_custom_url under a per-request prefix (SERVER_ROOT_PATHS):
+# when a request lives under a dynamic prefix, request.base_url already
+# contains it. Appending the SERVER_ROOT_PATH scalar on top produced
+# double-prefixed SSO callback / login URLs (a path that doesn't exist on
+# the deployment). These pin that the emitted URL now stays under one
+# prefix — the one the request actually arrived on.
+# ---------------------------------------------------------------------------
+
+
+def test_get_custom_url_uses_per_request_prefix_when_middleware_ran(monkeypatch):
+    """The middleware stashes the effective per-request prefix in a ContextVar.
+    ``get_custom_url`` reads that in preference to the SERVER_ROOT_PATH scalar,
+    and ``join_paths``'s tail-dedup collapses the append so a request whose
+    ``base_url`` already ends in ``/tenant-a`` does not become
+    ``/tenant-a/legacy/route``."""
+    _clear_url_env(monkeypatch)
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+
+    from litellm.proxy.middleware.per_request_root_path_middleware import (
+        _request_root_path_var,
+    )
+
+    token = _request_root_path_var.set("/tenant-a")
+    try:
+        # request.base_url already carries the tenant prefix; the scalar
+        # SERVER_ROOT_PATH must not be re-appended on top.
+        result = get_custom_url(
+            request_base_url="https://request.example.com/tenant-a/",
+            route="/v1/chat",
+        )
+    finally:
+        _request_root_path_var.reset(token)
+
+    assert result == "https://request.example.com/tenant-a/v1/chat"
+
+
+def test_get_custom_url_no_double_prefix_when_both_env_vars_configured(monkeypatch):
+    """Regression guard for the review point: with SERVER_ROOT_PATH also set
+    (scalar-legacy) and the request matched by SERVER_ROOT_PATHS (per-request),
+    the emitted URL is under one prefix — the per-request one — never both."""
+    _clear_url_env(monkeypatch)
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+    monkeypatch.setenv("SERVER_ROOT_PATHS", "/tenant-a")
+
+    from litellm.proxy.middleware.per_request_root_path_middleware import (
+        _request_root_path_var,
+    )
+
+    token = _request_root_path_var.set("/tenant-a")
+    try:
+        # No "/legacy" ever appears — the fix pins the reviewer's expected
+        # behavior: only one prefix should apply per request.
+        result = get_custom_url("https://api.example.com/tenant-a", "/sso/callback")
+    finally:
+        _request_root_path_var.reset(token)
+
+    assert result == "https://api.example.com/tenant-a/sso/callback"
+    assert "/legacy" not in result
+
+
+def test_get_custom_url_scalar_only_still_stamps_root_path(monkeypatch):
+    """Pre-middleware deployments have not opted into SERVER_ROOT_PATHS at all;
+    the ContextVar stays unset and the SERVER_ROOT_PATH scalar owns the answer
+    — the behavior every existing scalar-only deployment relies on."""
+    _clear_url_env(monkeypatch)
+    monkeypatch.setenv("SERVER_ROOT_PATH", "/legacy")
+
+    result = get_custom_url("https://request.example.com", "/v1/chat")
+
+    assert result == "https://request.example.com/legacy/v1/chat"

From 1ae3216120a49fd2ac1ea6a2f43e42e8a33c276b Mon Sep 17 00:00:00 2001
From: tin-berri <tin@berri.ai>
Date: Mon, 7 Sep 2026 12:05:43 -0700
Subject: [PATCH 75/79] fix(router): preserve default heuristic updates
 (#40007)

---
 .../auto_router_tuning_baseline.py            | 14 ++++--
 .../test_model_management_endpoints.py        | 10 ++--
 .../proxy/proxy_server/test_lifecycle.py      | 21 ++++++++
 .../test_auto_router_tuning_baseline.py       | 48 +++++++++++++++++--
 4 files changed, 80 insertions(+), 13 deletions(-)

diff --git a/litellm/router_utils/auto_router_tuning_baseline.py b/litellm/router_utils/auto_router_tuning_baseline.py
index 1e6412856a6..b82269d5824 100644
--- a/litellm/router_utils/auto_router_tuning_baseline.py
+++ b/litellm/router_utils/auto_router_tuning_baseline.py
@@ -10,7 +10,7 @@ from pydantic import ValidationError
 
 from litellm.router_strategy.complexity_router.config import ComplexityRouterConfig
 
-TUNING_BASELINE_PARAM_NAME: Final = "auto_router_tuning_baseline"
+TUNING_BASELINE_PARAM_NAME: Final = "auto_router_tuning_baseline_v2"
 
 HEURISTIC_V1_TUNING_FIELDS: Final = (
     "tiers",
@@ -29,6 +29,8 @@ HEURISTIC_V1_TUNING_FIELDS: Final = (
     "keyword_tier_rules",
 )
 
+_TUNING_FIELD_SET: Final = frozenset(HEURISTIC_V1_TUNING_FIELDS)
+
 _V1_SCORING_CLASSIFIER_TYPES: Final = frozenset({"heuristic", "heuristic_first", "hybrid"})
 _AUTO_ROUTER_COMPLEXITY_PREFIX: Final = "auto_router/complexity_router"
 _EMPTY: Final[Mapping[str, object]] = MappingProxyType({})
@@ -40,12 +42,16 @@ def _mapping(value: object) -> Mapping[str, object]:
 
 
 def tuning_fingerprint(complexity_router_config: object) -> str | None:
-    """Digest of normalized heuristic-v1 tuning fields, or None when the config is invalid."""
+    """Digest of explicitly supplied heuristic-v1 tuning fields, normalized, or None when invalid."""
+    raw: Final = _mapping(complexity_router_config)
     try:
-        validated: Final = ComplexityRouterConfig.model_validate(_mapping(complexity_router_config))
+        validated: Final = ComplexityRouterConfig.model_validate(raw)
     except ValidationError:
         return None
-    payload: Final = validated.model_dump(mode="json", include=frozenset(HEURISTIC_V1_TUNING_FIELDS))
+    supplied: Final = ((_TUNING_FIELD_SET - frozenset(("tier_model_configs",))) & frozenset(raw)) | (
+        frozenset(("tier_model_configs",)) if validated.tier_model_configs else frozenset()
+    )
+    payload: Final = validated.model_dump(mode="json", include=supplied)
     return hashlib.sha256(json.dumps(payload, sort_keys=True, separators=(",", ":")).encode()).hexdigest()
 
 
diff --git a/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py
index 63c08529b4d..d90338f8480 100644
--- a/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py
+++ b/tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py
@@ -4447,13 +4447,13 @@ class TestStrategyRouterWriteValidation:
             self.litellm_proxymodeltable = MagicMock(
                 create=AsyncMock(),
                 update=AsyncMock(),
-                find_many=AsyncMock(
-                    return_value=tuple(
-                        LiteLLM_ProxyModelTable.model_validate(row) for row in self.tuning_rows
-                    )
-                ),
+                find_many=AsyncMock(side_effect=self._find_many),
             )
 
+        async def _find_many(self, where: object = None) -> tuple[LiteLLM_ProxyModelTable, ...]:
+            json.dumps(where)  # prisma serializes the filter with json.dumps and rejects a mappingproxy
+            return tuple(LiteLLM_ProxyModelTable.model_validate(row) for row in self.tuning_rows)
+
         @property
         def db(self) -> "TestStrategyRouterWriteValidation._FakeTx":
             return self
diff --git a/tests/test_litellm/proxy/proxy_server/test_lifecycle.py b/tests/test_litellm/proxy/proxy_server/test_lifecycle.py
index a518c9289ae..7db85dc6943 100644
--- a/tests/test_litellm/proxy/proxy_server/test_lifecycle.py
+++ b/tests/test_litellm/proxy/proxy_server/test_lifecycle.py
@@ -821,6 +821,27 @@ def test_proxy_startup_event_warns_for_global_budget_without_database():
     )
 
 
+@pytest.mark.asyncio
+async def test_tuning_baseline_v2_is_created_alongside_the_legacy_row():
+    from litellm.router_utils.auto_router_tuning_baseline import DEFAULT_TUNING_FINGERPRINT
+
+    prisma_client = MagicMock()
+    prisma_client.db.litellm_config.find_unique = AsyncMock(return_value=None)
+    prisma_client.db.litellm_config.create = AsyncMock()
+    deployment = {
+        "model_name": "a",
+        "litellm_params": {"model": "auto_router/complexity_router", "complexity_router_config": {}},
+    }
+
+    result = await ProxyStartupEvent._load_heuristic_v1_tuning_baselines(prisma_client, [deployment])
+
+    assert result == {'yaml:["a",[]]': DEFAULT_TUNING_FINGERPRINT}
+    assert prisma_client.db.litellm_config.create.await_args.kwargs["data"] == {
+        "param_name": "auto_router_tuning_baseline_v2",
+        "param_value": json.dumps(dict(result)),
+    }
+
+
 @pytest.mark.asyncio
 async def test_tuning_baseline_waits_for_a_complete_db_model_census(monkeypatch):
     prisma_client = MagicMock()
diff --git a/tests/test_litellm/router_utils/test_auto_router_tuning_baseline.py b/tests/test_litellm/router_utils/test_auto_router_tuning_baseline.py
index c3395ab641f..fa7a96adb20 100644
--- a/tests/test_litellm/router_utils/test_auto_router_tuning_baseline.py
+++ b/tests/test_litellm/router_utils/test_auto_router_tuning_baseline.py
@@ -72,6 +72,9 @@ class TestTuningFingerprint:
             config["classifier_llm_config"] = {"model": "judge"}
         assert tuning_fingerprint(config) != DEFAULT_TUNING_FINGERPRINT
 
+    def test_explicit_empty_tier_model_configs_follow_omission(self) -> None:
+        assert tuning_fingerprint({"tier_model_configs": {}}) == DEFAULT_TUNING_FINGERPRINT
+
     def test_tier_model_overrides_change_the_fingerprint(self) -> None:
         plain = tuning_fingerprint({"tiers": {"SIMPLE": "x"}})
         with_override = tuning_fingerprint(
@@ -80,11 +83,39 @@ class TestTuningFingerprint:
         assert plain != with_override
 
     def test_non_tuning_fields_do_not_change_the_fingerprint(self) -> None:
-        assert tuning_fingerprint({"return_raw_model_name": True, "session_affinity": True}) == DEFAULT_TUNING_FINGERPRINT
+        assert (
+            tuning_fingerprint({"return_raw_model_name": True, "session_affinity": True}) == DEFAULT_TUNING_FINGERPRINT
+        )
 
     def test_invalid_config_has_no_fingerprint(self) -> None:
         assert tuning_fingerprint({"tier_boundaries": "not-a-mapping"}) is None
 
+    def test_a_release_changing_a_shipped_default_is_not_an_operator_edit(self, monkeypatch) -> None:
+        """An omitted setting follows the shipped default and stays off the quota when that default moves;
+        only what the operator wrote is fingerprinted, so an explicit value equal to the old default still counts."""
+        import litellm.router_strategy.complexity_router.config as config_module
+
+        untouched = _router("a", {})
+        tiers_only = _router("b", {"tiers": _TIERS})
+        pinned = _router("c", {"dimension_weights": dict(config_module.DEFAULT_DIMENSION_WEIGHTS)})
+        baselines = snapshot_tuning_baselines([untouched, tiers_only, pinned])
+        assert tuning_fingerprint(pinned["litellm_params"]["complexity_router_config"]) != DEFAULT_TUNING_FINGERPRINT
+
+        monkeypatch.setattr(
+            config_module,
+            "DEFAULT_DIMENSION_WEIGHTS",
+            {**config_module.DEFAULT_DIMENSION_WEIGHTS, "codePresence": 0.99},
+        )
+        monkeypatch.setattr(
+            config_module,
+            "DEFAULT_TIER_BOUNDARIES",
+            {**config_module.DEFAULT_TIER_BOUNDARIES, "simple_medium": 0.42},
+        )
+
+        assert tuning_fingerprint({}) == DEFAULT_TUNING_FINGERPRINT
+        assert mutable_tuned_identities([untouched, tiers_only, pinned], baselines) == frozenset()
+        assert mutable_tuned_identities([untouched], snapshot_tuning_baselines([])) == frozenset()
+
 
 class TestRouterIdentity:
     def test_db_rows_key_on_model_id_and_yaml_rows_on_name_and_tags(self) -> None:
@@ -161,12 +192,18 @@ class TestQuota:
         new_c = _router("c", {"tiers": _TIERS})
 
         assert tuning_quota_violation(candidate=edited_a, others=[legacy_b], baselines=baselines, limit=1) is None
-        assert tuning_quota_violation(candidate=edited_a, others=[edited_a, legacy_b], baselines=baselines, limit=1) is None
+        assert (
+            tuning_quota_violation(candidate=edited_a, others=[edited_a, legacy_b], baselines=baselines, limit=1)
+            is None
+        )
         assert tuning_quota_violation(candidate=legacy_a, others=[edited_b], baselines=baselines, limit=1) is None
         assert tuning_quota_violation(candidate=edited_b, others=[edited_a], baselines=baselines, limit=1) is not None
         assert tuning_quota_violation(candidate=new_c, others=[edited_a], baselines=baselines, limit=1) is not None
         assert tuning_quota_violation(candidate=new_c, others=[edited_a], baselines=baselines, limit=None) is None
-        assert tuning_quota_violation(candidate=legacy_a, others=[edited_a, edited_b], baselines=baselines, limit=1) is None
+        assert (
+            tuning_quota_violation(candidate=legacy_a, others=[edited_a, edited_b], baselines=baselines, limit=1)
+            is None
+        )
 
     def test_reverting_to_baseline_frees_the_quota(self) -> None:
         legacy_a = _router("a", {"tiers": _TIERS})
@@ -174,7 +211,10 @@ class TestQuota:
         baselines = snapshot_tuning_baselines([legacy_a, legacy_b])
         edited_b = _router("b", {"tiers": _TIERS})
         assert tuning_quota_violation(candidate=edited_b, others=[legacy_a], baselines=baselines, limit=1) is None
-        assert tuning_quota_violation(candidate=edited_b, others=[legacy_a, edited_b], baselines=baselines, limit=1) is None
+        assert (
+            tuning_quota_violation(candidate=edited_b, others=[legacy_a, edited_b], baselines=baselines, limit=1)
+            is None
+        )
 
     def test_violation_message_names_the_limit_and_remedy(self) -> None:
         message = tuning_limit_violation(held=2, limit=1)

From eea4c860f76ab5eadbe082ff36d081c00ec7346b Mon Sep 17 00:00:00 2001
From: tin-berri <tin@berri.ai>
Date: Mon, 7 Sep 2026 12:12:22 -0700
Subject: [PATCH 76/79] feat(ui): add key-scoped auto-router usage tab (#39999)

* feat(ui): add key-scoped auto-router usage tab

GET /auto_router/benchmarks takes an optional api_key filter, applied in the
rollup aggregate on the primary key's leading column. Proxy admins get a
separate Auto-router usage tab on key detail pages with spend, baseline,
savings, tier routing, cache metrics and the existing router selector

* fix(ui): share key analytics date range
---
 litellm/proxy/db/autorouter_session_rollup.py |  4 +-
 .../auto_router_endpoints.py                  |  2 +
 .../spend/test_autorouter_session_rollup.py   | 32 ++++++
 .../test_auto_router_endpoints.py             |  5 +-
 .../AutoRouterBenchmarksTab.test.tsx          | 28 +++++-
 .../_components/AutoRouterBenchmarksTab.tsx   |  9 +-
 .../useAutoRouterBenchmarks.test.ts           | 31 +++++-
 .../_components/useAutoRouterBenchmarks.ts    |  4 +-
 ...useDailyActivityRange.integration.test.tsx |  8 +-
 .../useDailyActivityRange.test.tsx            | 10 +-
 .../_components/useDailyActivityRange.ts      | 24 +++--
 ...KeyAutoRouterUsageTab.integration.test.tsx | 98 +++++++++++++++++++
 .../templates/KeyAutoRouterUsageTab.tsx       | 18 ++++
 .../KeySavingsTab.integration.test.tsx        | 12 ++-
 .../components/templates/KeySavingsTab.tsx    | 19 ++--
 .../templates/key_info_view.test.tsx          | 56 +++++++++++
 .../components/templates/key_info_view.tsx    | 26 ++++-
 ui/litellm-dashboard/src/lib/http/schema.d.ts |  2 +
 18 files changed, 354 insertions(+), 34 deletions(-)
 create mode 100644 ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.integration.test.tsx
 create mode 100644 ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.tsx

diff --git a/litellm/proxy/db/autorouter_session_rollup.py b/litellm/proxy/db/autorouter_session_rollup.py
index 9c637a62dc1..b33bffbaaa1 100644
--- a/litellm/proxy/db/autorouter_session_rollup.py
+++ b/litellm/proxy/db/autorouter_session_rollup.py
@@ -37,7 +37,9 @@ CACHE_TTL_1H_SECONDS: Final = 3600
 AUTOROUTER_BENCHMARKS_SQL: Final = """
 WITH windowed AS (
     SELECT * FROM "LiteLLM_AutoRouterSession"
-    WHERE last_turn_at >= $1::timestamp AND first_turn_at < $2::timestamp
+    WHERE last_turn_at >= $1::timestamp
+      AND first_turn_at < $2::timestamp
+      AND ($3::text IS NULL OR api_key = $3::text)
 ),
 tier_maps AS (
     SELECT router_name, router_type, jsonb_object_agg(tier, tier_turns) AS tier_turns
diff --git a/litellm/proxy/management_endpoints/auto_router_endpoints.py b/litellm/proxy/management_endpoints/auto_router_endpoints.py
index 2a7813bc140..0f0323b45f8 100644
--- a/litellm/proxy/management_endpoints/auto_router_endpoints.py
+++ b/litellm/proxy/management_endpoints/auto_router_endpoints.py
@@ -645,6 +645,7 @@ async def get_auto_router_benchmarks(
         str | None, Query(description="YYYY-MM-DD UTC, inclusive (defaults to 30 days before end_date)")
     ] = None,
     end_date: Annotated[str | None, Query(description="YYYY-MM-DD UTC, inclusive (defaults to today)")] = None,
+    api_key: Annotated[str | None, Query(description="Filter to one virtual key token hash")] = None,
 ) -> AutoRouterBenchmarksResponse:
     """
     Benchmarks for the auto-router dashboard: session shape, savings against the configured
@@ -681,6 +682,7 @@ async def get_auto_router_benchmarks(
         AUTOROUTER_BENCHMARKS_SQL,
         start_day.isoformat(),
         (end_day + timedelta(days=1)).isoformat(),
+        api_key,
     )
     rows: Final = _SESSION_AGG_ROWS.validate_python(raw_rows or ())
     groups: Final = (
diff --git a/tests/proxy_behavior/spend/test_autorouter_session_rollup.py b/tests/proxy_behavior/spend/test_autorouter_session_rollup.py
index 7bc61c40ea0..c2272f3d20d 100644
--- a/tests/proxy_behavior/spend/test_autorouter_session_rollup.py
+++ b/tests/proxy_behavior/spend/test_autorouter_session_rollup.py
@@ -169,6 +169,7 @@ async def test_the_benchmarks_aggregate_reads_only_overlapping_sessions(db):
         AUTOROUTER_BENCHMARKS_SQL,
         (T0 - timedelta(days=1)).isoformat(),
         (T0 + timedelta(days=1)).isoformat(),
+        None,
     )
     matching = [row for row in rows if row["router_name"] == router]
     assert len(matching) == 1
@@ -181,6 +182,33 @@ async def test_the_benchmarks_aggregate_reads_only_overlapping_sessions(db):
     assert grouped["session_seconds"] == pytest.approx(60.0)
 
 
+async def test_the_benchmarks_aggregate_can_filter_to_one_key(db):
+    router = f"r-{uuid.uuid4()}"
+    first_key = f"k-{uuid.uuid4()}"
+    second_key = f"k-{uuid.uuid4()}"
+    await _turn(db, first_key, "A", T0, session_id=f"s-{uuid.uuid4()}", router=router, saved=0.5)
+    await _turn(db, second_key, "A", T0, session_id=f"s-{uuid.uuid4()}", router=router, saved=9.0)
+
+    rows = await db.query_raw(
+        AUTOROUTER_BENCHMARKS_SQL,
+        (T0 - timedelta(days=1)).isoformat(),
+        (T0 + timedelta(days=1)).isoformat(),
+        first_key,
+    )
+    matching = [row for row in rows if row["router_name"] == router]
+    assert len(matching) == 1
+    assert matching[0]["sessions"] == 1
+    assert matching[0]["saved_spend"] == pytest.approx(0.5)
+
+    unknown_key_rows = await db.query_raw(
+        AUTOROUTER_BENCHMARKS_SQL,
+        (T0 - timedelta(days=1)).isoformat(),
+        (T0 + timedelta(days=1)).isoformat(),
+        f"k-{uuid.uuid4()}",
+    )
+    assert [row for row in unknown_key_rows if row["router_name"] == router] == []
+
+
 async def test_a_reconfigured_alias_reports_each_router_type_as_its_own_group(db):
     key = f"k-{uuid.uuid4()}"
     router = f"r-{uuid.uuid4()}"
@@ -191,6 +219,7 @@ async def test_a_reconfigured_alias_reports_each_router_type_as_its_own_group(db
         AUTOROUTER_BENCHMARKS_SQL,
         (T0 - timedelta(days=1)).isoformat(),
         (T0 + timedelta(days=1)).isoformat(),
+        None,
     )
     matching = sorted(
         (row for row in rows if row["router_name"] == router),
@@ -253,6 +282,7 @@ async def test_the_benchmarks_aggregate_sums_tier_turns_across_sessions(db):
         AUTOROUTER_BENCHMARKS_SQL,
         (T0 - timedelta(days=1)).isoformat(),
         (T0 + timedelta(days=1)).isoformat(),
+        None,
     )
     grouped = next(row for row in rows if row["router_name"] == router)
     assert grouped["tier_turns"] == {"simple": 2, "complex": 1}
@@ -280,6 +310,7 @@ async def test_tier_maps_stay_separate_per_router_type_on_a_reconfigured_alias(d
         AUTOROUTER_BENCHMARKS_SQL,
         (T0 - timedelta(days=1)).isoformat(),
         (T0 + timedelta(days=1)).isoformat(),
+        None,
     )
     by_type = {row["router_type"]: row["tier_turns"] for row in rows if row["router_name"] == router}
     assert by_type == {"complexity": {"medium": 1}, "quality": {"2": 1}}
@@ -294,6 +325,7 @@ async def test_a_window_with_no_tiered_turns_aggregates_to_an_empty_map(db):
         AUTOROUTER_BENCHMARKS_SQL,
         (T0 - timedelta(days=1)).isoformat(),
         (T0 + timedelta(days=1)).isoformat(),
+        None,
     )
     grouped = next(row for row in rows if row["router_name"] == router)
     assert grouped["tier_turns"] == {}
diff --git a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
index 21f8d985f22..35e76c96c14 100644
--- a/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
+++ b/tests/test_litellm/proxy/management_endpoints/test_auto_router_endpoints.py
@@ -489,6 +489,7 @@ class TestAutoRouterBenchmarks:
         monkeypatch: pytest.MonkeyPatch,
         rows: Sequence[Mapping[str, object]],
         model_list: Sequence[object],
+        api_key: str | None = None,
     ) -> AutoRouterBenchmarksResponse:
         from litellm.proxy import proxy_server
         from litellm.proxy.management_endpoints.auto_router_endpoints import get_auto_router_benchmarks
@@ -503,6 +504,7 @@ class TestAutoRouterBenchmarks:
             user_api_key_dict=ADMIN,
             start_date="2026-07-01",
             end_date="2026-08-01",
+            api_key=api_key,
         )
 
     ROW = _SessionAggRow(
@@ -652,8 +654,9 @@ class TestAutoRouterBenchmarks:
             user_api_key_dict=ADMIN,
             start_date="2026-07-01",
             end_date="2026-08-01",
+            api_key="key-hash",
         )
-        assert captured["params"] == ("2026-07-01T00:00:00", "2026-08-02T00:00:00")
+        assert captured["params"] == ("2026-07-01T00:00:00", "2026-08-02T00:00:00", "key-hash")
         assert response.routers_in_scope == 1
         assert response.groups[0].router_name == "live-auto"
         assert response.groups[0].saved_pct == response.totals.saved_pct == 75.0
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
index ce0cd75cd36..e7c6adc478c 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
@@ -22,7 +22,7 @@ vi.mock("@/components/shared/advanced_date_picker", () => ({
 
 import { useAutoRouters } from "@/app/(dashboard)/hooks/models/useModels";
 
-import AutoRouterBenchmarksTab from "./AutoRouterBenchmarksTab";
+import AutoRouterBenchmarksTab, { AutoRouterUsageView } from "./AutoRouterBenchmarksTab";
 import type {
   AutoRouterBenchmarkGroup,
   AutoRouterBenchmarksResponse,
@@ -359,13 +359,37 @@ describe("AutoRouterBenchmarksTab", () => {
     mockHook({ data: response([group()]) });
     const { dateValue, onDateChange } = renderTab();
 
-    expect(vi.mocked(useAutoRouterBenchmarks)).toHaveBeenCalledWith("sk-test", dateValue);
+    expect(vi.mocked(useAutoRouterBenchmarks)).toHaveBeenCalledWith("sk-test", dateValue, undefined);
     expect(screen.getByText("Jul 6 – Aug 5 (UTC)")).toBeInTheDocument();
 
     fireEvent.click(screen.getByTestId("date-picker"));
     expect(onDateChange).toHaveBeenCalledWith({ from: new Date(2026, 7, 1), to: new Date(2026, 7, 5) });
   });
 
+  it("scopes the query to one key when the usage view is mounted for a key", () => {
+    mockHook({ data: response([group()]) });
+    const dateValue = { from: new Date(2026, 6, 6), to: new Date(2026, 7, 5) };
+    const activity = {
+      dateValue,
+      onDateChange: vi.fn(),
+      results: [],
+      loading: false,
+      isFetchingMore: false,
+      progress: { currentPage: 1, totalPages: 1 },
+      cancelled: false,
+      cancel: vi.fn(),
+    };
+    render(
+      <QueryClientProvider client={new QueryClient()}>
+        <AutoRouterUsageView accessToken="sk-test" activity={activity} apiKey="key-hash-1" />
+      </QueryClientProvider>,
+    );
+
+    expect(vi.mocked(useAutoRouterBenchmarks)).toHaveBeenCalledWith("sk-test", dateValue, "key-hash-1");
+    expect(screen.getByText("Total estimated savings")).toBeInTheDocument();
+    expect(screen.queryByRole("tab", { name: "Shadow Evals" })).not.toBeInTheDocument();
+  });
+
   it("shows usage by default and mounts shadow evals only when its sub-tab is selected", () => {
     mockHook({ data: response([group()]) });
     renderTab();
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
index 09a0cf0242b..39e6b0fd390 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
@@ -273,12 +273,13 @@ const BenchmarksBody: React.FC<BenchmarksBodyProps> = ({ isPending, error, data,
 
 interface AutoRouterBenchmarksTabProps {
   accessToken: string | null;
-  activity: DailyActivityRange;
+  activity: Pick<DailyActivityRange, "dateValue" | "onDateChange">;
+  apiKey?: string;
 }
 
-const UsageView: React.FC<AutoRouterBenchmarksTabProps> = ({ accessToken, activity }) => {
+export const AutoRouterUsageView: React.FC<AutoRouterBenchmarksTabProps> = ({ accessToken, activity, apiKey }) => {
   const { dateValue, onDateChange } = activity;
-  const { data, isPending, error } = useAutoRouterBenchmarks(accessToken, dateValue);
+  const { data, isPending, error } = useAutoRouterBenchmarks(accessToken, dateValue, apiKey);
   const [selectedKey, setSelectedKey] = useState<string>(ALL_ROUTERS);
   const { data: autoRouters } = useAutoRouters();
 
@@ -347,7 +348,7 @@ const AutoRouterBenchmarksTab: React.FC<AutoRouterBenchmarksTabProps> = ({ acces
       </TabsList>
 
       <TabsContent value="usage" keepMounted={visitedTabs.includes("usage")}>
-        <UsageView accessToken={accessToken} activity={activity} />
+        <AutoRouterUsageView accessToken={accessToken} activity={activity} />
       </TabsContent>
       <TabsContent value="shadow-evals" keepMounted={visitedTabs.includes("shadow-evals")}>
         <ShadowEvalSection />
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.test.ts
index e8a858ef3a5..1258c967ec9 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.test.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.test.ts
@@ -3,7 +3,9 @@ import { describe, expect, it, vi } from "vitest";
 vi.mock("@/components/networking", () => ({ formatDate: vi.fn() }));
 vi.mock("@/lib/http/api", () => ({ $api: { useQuery: vi.fn() } }));
 
-import { benchmarksWindow } from "./useAutoRouterBenchmarks";
+import { $api } from "@/lib/http/api";
+
+import { benchmarksWindow, useAutoRouterBenchmarks } from "./useAutoRouterBenchmarks";
 
 const localDay =
   (offsetHours: number) =>
@@ -44,3 +46,30 @@ describe("benchmarksWindow", () => {
     expect(benchmarksWindow({ to: now }, now, pacific)).toEqual({});
   });
 });
+
+describe("useAutoRouterBenchmarks", () => {
+  const range = { from: new Date("2026-07-06T19:00:00Z"), to: new Date("2026-08-05T19:00:00Z") };
+
+  it("forwards the key hash as the endpoint's api_key filter", () => {
+    useAutoRouterBenchmarks("sk-test", range, "key-hash-1");
+
+    const [, path, init] = vi.mocked($api.useQuery).mock.calls.at(-1) as unknown as [
+      string,
+      string,
+      { params: { query: Record<string, string | undefined> } },
+    ];
+    expect(path).toBe("/auto_router/benchmarks");
+    expect(init.params.query.api_key).toBe("key-hash-1");
+  });
+
+  it("leaves the read deployment-wide when no key is given", () => {
+    useAutoRouterBenchmarks("sk-test", range);
+
+    const [, , init] = vi.mocked($api.useQuery).mock.calls.at(-1) as unknown as [
+      string,
+      string,
+      { params: { query: Record<string, string | undefined> } },
+    ];
+    expect(init.params.query.api_key).toBeUndefined();
+  });
+});
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.ts
index 155c9d0c696..284342f0aec 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useAutoRouterBenchmarks.ts
@@ -23,10 +23,10 @@ export const benchmarksWindow = (
   };
 };
 
-export const useAutoRouterBenchmarks = (accessToken: string | null, range: DateRange) =>
+export const useAutoRouterBenchmarks = (accessToken: string | null, range: DateRange, apiKey?: string) =>
   $api.useQuery(
     "get",
     "/auto_router/benchmarks",
-    { params: { query: benchmarksWindow(range, new Date()) } },
+    { params: { query: { ...benchmarksWindow(range, new Date()), api_key: apiKey } } },
     { enabled: Boolean(accessToken && range.from && range.to), retry: false },
   );
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.integration.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.integration.test.tsx
index 78280e32eed..51dfe9fda90 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.integration.test.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.integration.test.tsx
@@ -24,12 +24,18 @@ describe("useScopedDailyActivityRange wiring", () => {
     );
     global.fetch = mockFetch;
 
-    renderHook(() => useScopedDailyActivityRange("sk-token", { userId: "u1", apiKey: "hash-abc" }));
+    const activity = {
+      dateValue: { from: new Date(2026, 7, 1), to: new Date(2026, 7, 10) },
+      onDateChange: vi.fn(),
+    };
+    renderHook(() => useScopedDailyActivityRange("sk-token", { userId: "u1", apiKey: "hash-abc" }, activity));
 
     await waitFor(() => expect(mockFetch).toHaveBeenCalled());
 
     const url = String(mockFetch.mock.calls[0][0]);
     expect(url).toContain("user_id=u1");
     expect(url).toContain("api_key=hash-abc");
+    expect(url).toContain("start_date=2026-08-01");
+    expect(url).toContain("end_date=2026-08-10");
   });
 });
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx
index 6c9281060a4..00902aa9fdd 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.test.tsx
@@ -25,11 +25,19 @@ vi.mock("@/components/networking", () => ({
 }));
 
 import { userDailyActivityAggregatedCall } from "@/components/networking";
-import { useDailyActivityRange } from "./useDailyActivityRange";
+import { useActivityDateRange, useDailyActivityRange } from "./useDailyActivityRange";
 
 const argsOfLastCall = () => mockUsePaginatedDailyActivity.mock.calls.at(-1)?.[0].args as unknown[];
 
 describe("useDailyActivityRange", () => {
+  it("offers date-range state without starting a daily-activity query", () => {
+    const { result } = renderHook(() => useActivityDateRange());
+
+    expect(result.current.dateValue.from).toBeInstanceOf(Date);
+    expect(result.current.dateValue.to).toBeInstanceOf(Date);
+    expect(mockUsePaginatedDailyActivity).not.toHaveBeenCalled();
+  });
+
   it("queries every user's activity for an admin", () => {
     renderHook(() => useDailyActivityRange("test-token", "u1", "proxy_admin"));
 
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts
index 8ed57e36a94..3435b57dbc8 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useDailyActivityRange.ts
@@ -37,14 +37,20 @@ export interface DailyActivityScope {
   apiKey?: string | null;
 }
 
-export const useScopedDailyActivityRange = (
-  accessToken: string | null,
-  scope: DailyActivityScope,
-): DailyActivityRange => {
+export type ActivityDateRange = Pick<DailyActivityRange, "dateValue" | "onDateChange">;
+
+export const useActivityDateRange = (): ActivityDateRange => {
   const initialFrom = useMemo(() => new Date(new Date().getTime() - THIRTY_DAYS_MS), []);
   const initialTo = useMemo(() => new Date(), []);
   const [dateValue, setDateValue] = useState<DateRange>({ from: initialFrom, to: initialTo });
+  return { dateValue, onDateChange: setDateValue };
+};
 
+export const useScopedDailyActivityRange = (
+  accessToken: string | null,
+  scope: DailyActivityScope,
+  { dateValue, onDateChange }: ActivityDateRange,
+): DailyActivityRange => {
   const startTime = dateValue.from ?? null;
   const endTime = dateValue.to ?? null;
   const { userId, apiKey = null } = scope;
@@ -63,7 +69,7 @@ export const useScopedDailyActivityRange = (
 
   return {
     dateValue,
-    onDateChange: setDateValue,
+    onDateChange,
     results: data.results as DailyData[],
     loading,
     isFetchingMore,
@@ -77,7 +83,7 @@ export const useDailyActivityRange = (
   accessToken: string | null,
   userId: string | null,
   userRole: string,
-): DailyActivityRange =>
-  useScopedDailyActivityRange(accessToken, {
-    userId: spendScopeUserId(userRole, userId),
-  });
+): DailyActivityRange => {
+  const dateRange = useActivityDateRange();
+  return useScopedDailyActivityRange(accessToken, { userId: spendScopeUserId(userRole, userId) }, dateRange);
+};
diff --git a/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.integration.test.tsx b/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.integration.test.tsx
new file mode 100644
index 00000000000..8cfd9d1941e
--- /dev/null
+++ b/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.integration.test.tsx
@@ -0,0 +1,98 @@
+import { screen } from "@testing-library/react";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { renderWithProviders, testQueryClient } from "../../../tests/test-utils";
+import KeyAutoRouterUsageTab from "./KeyAutoRouterUsageTab";
+
+vi.mock("@/app/(dashboard)/hooks/useAuthorized", () => ({
+  default: () => ({ accessToken: "test-token", userId: "admin-123", userRole: "Admin" }),
+}));
+
+const jsonResponse = (body: unknown) =>
+  new Response(JSON.stringify(body), { status: 200, headers: { "content-type": "application/json" } });
+
+const cache = {
+  coverage_pct: 100,
+  hit_rate_pct: 50,
+  same_model: { turns: 2, hits: 1, hit_rate_pct: 50 },
+  first_visit: { turns: 1, hits: 0, hit_rate_pct: 0 },
+  return_to_tier: { turns: 1, hits: 1, hit_rate_pct: 100 },
+  unordered_turns: 0,
+  return_misses_expired: 0,
+  return_misses_within_ttl: 0,
+  return_misses_unknown: 0,
+  ttl_5m_turns: 0,
+  ttl_1h_turns: 0,
+};
+
+const stats = {
+  sessions: 2,
+  turns: 4,
+  avg_turns_per_session: 2,
+  avg_session_seconds: 30,
+  avg_tokens_per_session: 100,
+  spend: 1.25,
+  saved_spend: 8.75,
+  baseline_spend: 10,
+  saved_pct: 87.5,
+  saved_per_session: 4.375,
+  cache,
+};
+
+const benchmarks = {
+  start_date: "2025-01-01",
+  end_date: "2025-01-31",
+  routers_in_scope: 2,
+  totals: stats,
+  groups: [
+    { router_name: "router-one", router_type: "complexity", tier_turns: { SIMPLE: 4 }, ...stats },
+    {
+      router_name: "router-two",
+      router_type: "complexity",
+      tier_turns: { SIMPLE: 1 },
+      ...stats,
+      spend: 0.25,
+      saved_spend: 0.75,
+      baseline_spend: 1,
+    },
+  ],
+};
+
+const noDeployments = { data: [], total_count: 0, current_page: 1, total_pages: 1, size: 1000 };
+const fetchMock = vi.fn(async (request: Request | string) => {
+  const url = typeof request === "string" ? request : request.url;
+  if (url.includes("/auto_router/benchmarks")) return jsonResponse(benchmarks);
+  return jsonResponse(noDeployments);
+});
+const requestedUrls = () =>
+  fetchMock.mock.calls.map(([request]) => (typeof request === "string" ? request : request.url));
+
+describe("KeyAutoRouterUsageTab", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    testQueryClient.clear();
+    vi.stubGlobal("fetch", fetchMock);
+  });
+
+  it("renders this key's spend, baseline, savings and per-router filter", async () => {
+    const activity = {
+      dateValue: { from: new Date(2025, 0, 1), to: new Date(2025, 0, 31) },
+      onDateChange: vi.fn(),
+    };
+    renderWithProviders(<KeyAutoRouterUsageTab accessToken="test-token" keyToken="key-hash-1" activity={activity} />);
+
+    expect(await screen.findByText("$8.75")).toBeInTheDocument();
+    expect(screen.getByText("Actual auto-router spend")).toBeInTheDocument();
+    expect(screen.getByText("$1.25")).toBeInTheDocument();
+    expect(screen.getByText("Estimated spend at highest-tier model")).toBeInTheDocument();
+    expect(screen.getByText("$10.00")).toBeInTheDocument();
+    expect(screen.getByText("Auto-router prompt caching")).toBeInTheDocument();
+    expect(screen.getAllByText("50.0%").length).toBeGreaterThan(0);
+    expect(screen.getByText("All auto-routers")).toBeInTheDocument();
+
+    const benchmarkUrl = new URL(requestedUrls().find((url) => url.includes("/auto_router/benchmarks")) ?? "");
+    expect(benchmarkUrl.searchParams.get("api_key")).toBe("key-hash-1");
+    expect(benchmarkUrl.searchParams.get("start_date")).toBe("2025-01-01");
+    expect(benchmarkUrl.searchParams.get("end_date")).toBe("2025-01-31");
+  });
+});
diff --git a/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.tsx b/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.tsx
new file mode 100644
index 00000000000..9e1cee1ec2e
--- /dev/null
+++ b/ui/litellm-dashboard/src/components/templates/KeyAutoRouterUsageTab.tsx
@@ -0,0 +1,18 @@
+"use client";
+
+import React from "react";
+
+import { AutoRouterUsageView } from "@/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab";
+import type { ActivityDateRange } from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange";
+
+interface KeyAutoRouterUsageTabProps {
+  accessToken: string | null;
+  keyToken: string;
+  activity: ActivityDateRange;
+}
+
+const KeyAutoRouterUsageTab: React.FC<KeyAutoRouterUsageTabProps> = ({ accessToken, keyToken, activity }) => (
+  <AutoRouterUsageView accessToken={accessToken} activity={activity} apiKey={keyToken} />
+);
+
+export default KeyAutoRouterUsageTab;
diff --git a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx
index 6c7d65ee566..385c3967d02 100644
--- a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx
+++ b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.integration.test.tsx
@@ -46,6 +46,11 @@ const mockActivity = (
 
 const scopedRange = () => vi.spyOn(useScopedDailyActivityRangeModule, "useScopedDailyActivityRange");
 
+const activity = {
+  dateValue: { from: new Date(2025, 0, 1), to: new Date(2025, 0, 31) },
+  onDateChange: vi.fn(),
+};
+
 const renderTab = (props: Partial<React.ComponentProps<typeof KeySavingsTab>> = {}) =>
   render(
     <KeySavingsTab
@@ -53,6 +58,7 @@ const renderTab = (props: Partial<React.ComponentProps<typeof KeySavingsTab>> =
       keyToken="key-abc123"
       userId="user-123"
       userRole="Internal User"
+      activity={activity}
       {...props}
     />,
   );
@@ -114,7 +120,7 @@ describe("KeySavingsTab", () => {
 
     renderTab({ userId: "user-456", userRole: "Internal User" });
 
-    expect(hook).toHaveBeenCalledWith("test-token", { userId: "user-456", apiKey: "key-abc123" });
+    expect(hook).toHaveBeenCalledWith("test-token", { userId: "user-456", apiKey: "key-abc123" }, activity);
     expect(screen.getByTestId("key-savings-scope-note")).toHaveTextContent("Showing your own requests");
   });
 
@@ -123,7 +129,7 @@ describe("KeySavingsTab", () => {
 
     renderTab({ userId: "admin-123", userRole: "Admin" });
 
-    expect(hook).toHaveBeenCalledWith("test-token", { userId: null, apiKey: "key-abc123" });
+    expect(hook).toHaveBeenCalledWith("test-token", { userId: null, apiKey: "key-abc123" }, activity);
     expect(screen.queryByTestId("key-savings-scope-note")).not.toBeInTheDocument();
   });
 
@@ -132,7 +138,7 @@ describe("KeySavingsTab", () => {
 
     renderTab({ userId: "org-admin-1", userRole: "Org Admin" });
 
-    expect(hook).toHaveBeenCalledWith("test-token", { userId: "org-admin-1", apiKey: "key-abc123" });
+    expect(hook).toHaveBeenCalledWith("test-token", { userId: "org-admin-1", apiKey: "key-abc123" }, activity);
     expect(screen.getByTestId("key-savings-scope-note")).toHaveTextContent("Showing your own requests");
   });
 });
diff --git a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx
index 967c3c7eff0..c33529eb042 100644
--- a/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx
+++ b/ui/litellm-dashboard/src/components/templates/KeySavingsTab.tsx
@@ -22,7 +22,10 @@ import {
   usd,
   withStartAnchor,
 } from "@/app/(dashboard)/cost-optimization/_components/costOptimizationUtils";
-import { useScopedDailyActivityRange } from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange";
+import {
+  useScopedDailyActivityRange,
+  type ActivityDateRange,
+} from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange";
 
 interface KeySavingsTabProps {
   accessToken: string | null;
@@ -30,19 +33,19 @@ interface KeySavingsTabProps {
   keyToken: string;
   userId: string | null;
   userRole: string;
+  activity: ActivityDateRange;
 }
 
-const KeySavingsTab: React.FC<KeySavingsTabProps> = ({ accessToken, keyToken, userId, userRole }) => {
+const KeySavingsTab: React.FC<KeySavingsTabProps> = ({ accessToken, keyToken, userId, userRole, activity }) => {
   // Proxy admins read the whole key. For anyone else the endpoint applies the caller's own user_id
   // alongside the key filter, so the figures cover only that viewer's requests on this key -- said
   // plainly in the scope note below rather than left to be misread as the key's total.
   const readsWholeKey = hasProxyWideSpendView(userRole);
-  const activity = useScopedDailyActivityRange(accessToken, {
-    userId: spendScopeUserId(userRole, userId),
-    apiKey: keyToken,
-  });
-
-  const { dateValue, onDateChange, results, loading, isFetchingMore } = activity;
+  const { dateValue, onDateChange, results, loading, isFetchingMore } = useScopedDailyActivityRange(
+    accessToken,
+    { userId: spendScopeUserId(userRole, userId), apiKey: keyToken },
+    activity,
+  );
   const startTime = dateValue.from ?? null;
   const endTime = dateValue.to ?? null;
 
diff --git a/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx b/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
index 7a27cc41e5e..b403255b329 100644
--- a/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
+++ b/ui/litellm-dashboard/src/components/templates/key_info_view.test.tsx
@@ -26,6 +26,30 @@ vi.mock("./key_edit_view", () => ({
   },
 }));
 
+import type { ActivityDateRange } from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange";
+
+const AnalyticsDateControl = ({ activity, keyToken }: { activity: ActivityDateRange; keyToken: string }) => (
+  <div>
+    <span data-testid="key-auto-router-usage">{keyToken}</span>
+    <output aria-label="Selected dates">{activity.dateValue.from?.toISOString()}</output>
+    {[1, 10].map((day) => (
+      <button
+        key={day}
+        onClick={() => activity.onDateChange({ from: new Date(Date.UTC(2026, 7, day)), to: new Date(2026, 7, 20) })}
+      >
+        Select August {day}
+      </button>
+    ))}
+  </div>
+);
+
+vi.mock("./KeyAutoRouterUsageTab", () => ({
+  default: (props: React.ComponentProps<typeof AnalyticsDateControl>) => <AnalyticsDateControl {...props} />,
+}));
+vi.mock("./KeySavingsTab", () => ({
+  default: (props: React.ComponentProps<typeof AnalyticsDateControl>) => <AnalyticsDateControl {...props} />,
+}));
+
 vi.mock("@/app/(dashboard)/hooks/useTeams", () => ({
   default: vi.fn(),
 }));
@@ -176,6 +200,38 @@ describe("KeyInfoView", () => {
     await userEvent.click(await screen.findByRole("button", { name: /more key actions/i }));
   };
 
+  it("shows key-scoped auto-router usage as its own admin tab", async () => {
+    vi.mocked(useAuthorized).mockReturnValue({ ...baseUseAuthorizedMock, userRole: "Admin" });
+    renderWithProviders(<KeyInfoView keyData={MOCK_KEY_DATA} onClose={() => {}} keyId="test-key-id" teams={[]} />);
+
+    await userEvent.click(screen.getByRole("tab", { name: "Auto-router usage" }));
+
+    expect(screen.getByTestId("key-auto-router-usage")).toHaveTextContent("test-token-123");
+  });
+
+  it("preserves dates in both directions across unmounted analytics panels", async () => {
+    vi.mocked(useAuthorized).mockReturnValue({ ...baseUseAuthorizedMock, userRole: "Admin" });
+    renderWithProviders(<KeyInfoView keyData={MOCK_KEY_DATA} onClose={() => {}} keyId="test-key-id" teams={[]} />);
+
+    expect(screen.queryByLabelText("Selected dates")).not.toBeInTheDocument();
+    await userEvent.click(screen.getByRole("tab", { name: "Savings" }));
+    await userEvent.click(screen.getByRole("button", { name: "Select August 1" }));
+    await userEvent.click(screen.getByRole("tab", { name: "Auto-router usage" }));
+    expect(screen.getByLabelText("Selected dates")).toHaveTextContent("2026-08-01T00:00:00.000Z");
+    await userEvent.click(screen.getByRole("button", { name: "Select August 10" }));
+    await userEvent.click(screen.getByRole("tab", { name: "Settings" }));
+    expect(screen.queryByLabelText("Selected dates")).not.toBeInTheDocument();
+    await userEvent.click(screen.getByRole("tab", { name: "Savings" }));
+    expect(screen.getByLabelText("Selected dates")).toHaveTextContent("2026-08-10T00:00:00.000Z");
+  });
+
+  it("does not offer the admin-only auto-router usage tab to an internal user", () => {
+    vi.mocked(useAuthorized).mockReturnValue({ ...baseUseAuthorizedMock, userRole: "Internal User" });
+    renderWithProviders(<KeyInfoView keyData={MOCK_KEY_DATA} onClose={() => {}} keyId="test-key-id" teams={[]} />);
+
+    expect(screen.queryByRole("tab", { name: "Auto-router usage" })).not.toBeInTheDocument();
+  });
+
   describe("last updated", () => {
     const renderWithTimestamps = (overrides: Partial<KeyResponse>) => {
       vi.mocked(useAuthorized).mockReturnValue(baseUseAuthorizedMock);
diff --git a/ui/litellm-dashboard/src/components/templates/key_info_view.tsx b/ui/litellm-dashboard/src/components/templates/key_info_view.tsx
index 0dd0dd6d6af..f5c682a2ee0 100644
--- a/ui/litellm-dashboard/src/components/templates/key_info_view.tsx
+++ b/ui/litellm-dashboard/src/components/templates/key_info_view.tsx
@@ -16,8 +16,15 @@ import { modelGroupHref, teamDetailHref } from "@/utils/entityLinks";
 import { BadgeLink } from "@/components/shared/BadgeLink";
 import { KeyInfoHeader } from "./KeyInfoHeader";
 import KeySavingsTab from "./KeySavingsTab";
+import KeyAutoRouterUsageTab from "./KeyAutoRouterUsageTab";
+import { useActivityDateRange } from "@/app/(dashboard)/cost-optimization/_components/useDailyActivityRange";
 import { useEffect, useState } from "react";
-import { isProxyAdminRole, isUserTeamAdminForSingleTeam, rolesWithWriteAccess } from "../../utils/roles";
+import {
+  hasProxyWideSpendView,
+  isProxyAdminRole,
+  isUserTeamAdminForSingleTeam,
+  rolesWithWriteAccess,
+} from "../../utils/roles";
 import { mapDisplayToInternalNames, mapInternalToDisplayNames } from "../callback_info_helpers";
 import AutoRotationView from "../common_components/AutoRotationView";
 import DeleteResourceModal from "../common_components/DeleteResourceModal";
@@ -81,6 +88,7 @@ export default function KeyInfoView({
   backButtonText = "Back to Keys",
 }: KeyInfoViewProps) {
   const { accessToken, userId: userID, userRole, premiumUser } = useAuthorized();
+  const activityDateRange = useActivityDateRange();
   const queryClient = useQueryClient();
   const canEditGuardrails = premiumUser || (userRole != null && rolesWithWriteAccess.includes(userRole));
   const { teams: teamsData } = useTeams();
@@ -618,6 +626,11 @@ export default function KeyInfoView({
           <TabsTrigger value="savings" className="flex-none rounded-none px-4 py-2">
             Savings
           </TabsTrigger>
+          {hasProxyWideSpendView(userRole) && (
+            <TabsTrigger value="auto-router-usage" className="flex-none rounded-none px-4 py-2">
+              Auto-router usage
+            </TabsTrigger>
+          )}
           <TabsTrigger value="settings" className="flex-none rounded-none px-4 py-2">
             Settings
           </TabsTrigger>
@@ -761,9 +774,20 @@ export default function KeyInfoView({
               keyToken={currentKeyData.token}
               userId={userID}
               userRole={userRole}
+              activity={activityDateRange}
             />
           </TabsContent>
 
+          {hasProxyWideSpendView(userRole) && (
+            <TabsContent value="auto-router-usage">
+              <KeyAutoRouterUsageTab
+                accessToken={accessToken}
+                keyToken={currentKeyData.token}
+                activity={activityDateRange}
+              />
+            </TabsContent>
+          )}
+
           {/* Settings Panel */}
           <TabsContent value="settings" keepMounted>
             <Card className="block p-6">
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 36fd744efc0..842a3da4122 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -41393,6 +41393,8 @@ export interface operations {
                 start_date?: string | null;
                 /** @description YYYY-MM-DD UTC, inclusive (defaults to today) */
                 end_date?: string | null;
+                /** @description Filter to one virtual key token hash */
+                api_key?: string | null;
             };
             header?: never;
             path?: never;

From cc287a7d8f484dc3eb063aff33999068f733b00c Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 7 Sep 2026 12:23:47 -0700
Subject: [PATCH 77/79] fix(ui): hide the Create Vector Store flow from non
 proxy admins (#40148)

* fix(ui): hide the Create Vector Store flow from non proxy admins

The vector stores page rendered the Create Vector Store tab, the
+ Add Vector Store button and a GET /credentials call for every role,
while the proxy only lets proxy admins call POST /vector_store/new and
GET /credentials. Internal users landed on the create form and got an
Only proxy admin error toast. Gate all three on isProxyAdminRole and
default everyone else to the Manage tab, matching the Indexes tab and
the Add Model gating.

Resolves LIT-7131

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(ui): exclude view-only admin sessions from the vector store create flow

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../vector-stores/_components/index.test.tsx  | 51 ++++++++++++++++---
 .../vector-stores/_components/index.tsx       | 35 ++++++++-----
 .../app/(dashboard)/vector-stores/page.tsx    |  6 ++-
 3 files changed, 70 insertions(+), 22 deletions(-)

diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.test.tsx
index 7137da201d8..3224496b13a 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.test.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.test.tsx
@@ -45,7 +45,7 @@ describe("VectorStoreManagement loading state", () => {
 
   it("should resolve the loading state when accessToken is null instead of showing the skeleton forever", async () => {
     const user = userEvent.setup();
-    render(<VectorStoreManagement accessToken={null} userID={null} userRole={null} />);
+    render(<VectorStoreManagement accessToken={null} userID={null} userRole={null} isViewOnly={false} />);
     await openManageTab(user);
     expect(await screen.findByText("table-loaded")).toBeInTheDocument();
     expect(mockVectorStoreListCall).not.toHaveBeenCalled();
@@ -59,7 +59,7 @@ describe("VectorStoreManagement loading state", () => {
         resolveFetch = resolve;
       }),
     );
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
     await openManageTab(user);
     expect(screen.getByText("table-loading")).toBeInTheDocument();
 
@@ -69,6 +69,43 @@ describe("VectorStoreManagement loading state", () => {
   });
 });
 
+describe("VectorStoreManagement create flow visibility", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    mockVectorStoreListCall.mockResolvedValue({ data: [] });
+    mockCredentialListCall.mockResolvedValue({ credentials: [] });
+  });
+
+  it.each([
+    { label: "Internal User", userRole: "Internal User", isViewOnly: false },
+    { label: "Internal Viewer", userRole: "Internal Viewer", isViewOnly: true },
+    { label: "proxy_admin_viewer session (userRole Admin, isViewOnly)", userRole: "Admin", isViewOnly: true },
+    { label: "Org Admin", userRole: "Org Admin", isViewOnly: false },
+  ])(
+    "should hide the Create Vector Store tab and button and skip /credentials for $label",
+    async ({ userRole, isViewOnly }) => {
+      render(
+        <VectorStoreManagement accessToken="sk-test" userID="user-1" userRole={userRole} isViewOnly={isViewOnly} />,
+      );
+      await waitFor(() => expect(mockVectorStoreListCall).toHaveBeenCalledWith("sk-test"));
+      expect(screen.queryByRole("tab", { name: "Create Vector Store" })).not.toBeInTheDocument();
+      expect(screen.getByRole("tab", { name: "Manage Vector Stores" })).toHaveAttribute("aria-selected", "true");
+      expect(await screen.findByText("table-loaded")).toBeInTheDocument();
+      expect(screen.queryByRole("button", { name: "+ Add Vector Store" })).not.toBeInTheDocument();
+      expect(mockCredentialListCall).not.toHaveBeenCalled();
+    },
+  );
+
+  it("should keep the Create Vector Store tab and button and fetch /credentials for a proxy admin", async () => {
+    const user = userEvent.setup();
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
+    await waitFor(() => expect(mockCredentialListCall).toHaveBeenCalledWith("sk-test"));
+    expect(screen.getByRole("tab", { name: "Create Vector Store" })).toHaveAttribute("aria-selected", "true");
+    await openManageTab(user);
+    expect(screen.getByRole("button", { name: "+ Add Vector Store" })).toBeInTheDocument();
+  });
+});
+
 describe("VectorStoreManagement Indexes tab", () => {
   beforeEach(() => {
     vi.clearAllMocks();
@@ -88,7 +125,7 @@ describe("VectorStoreManagement Indexes tab", () => {
         },
       ],
     });
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
     await user.click(screen.getByRole("tab", { name: "Indexes" }));
     expect(await screen.findByText("support-docs-index")).toBeInTheDocument();
     expect(screen.getByText("support-docs-store")).toBeInTheDocument();
@@ -96,7 +133,7 @@ describe("VectorStoreManagement Indexes tab", () => {
   });
 
   it("should not render the Indexes tab for an Admin Viewer", async () => {
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin Viewer" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin Viewer" isViewOnly={true} />);
     await waitFor(() => expect(mockVectorStoreListCall).toHaveBeenCalledWith("sk-test"));
     expect(screen.getByRole("tab", { name: "Manage Vector Stores" })).toBeInTheDocument();
     expect(screen.queryByRole("tab", { name: "Indexes" })).not.toBeInTheDocument();
@@ -125,7 +162,7 @@ describe("VectorStoreManagement Indexes tab", () => {
         },
       ],
     });
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
     await user.click(screen.getByRole("tab", { name: "Indexes" }));
     await user.click(await screen.findByRole("button", { name: "support-docs-store" }));
     expect(await screen.findByTestId("vector-store-info-view")).toHaveTextContent("vs-1");
@@ -135,7 +172,7 @@ describe("VectorStoreManagement Indexes tab", () => {
   it("should link to the feature docs and a GitHub issue for unsupported providers on the Indexes tab", async () => {
     const user = userEvent.setup();
     mockIndexesListCall.mockResolvedValue({ object: "list", data: [] });
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
     await user.click(screen.getByRole("tab", { name: "Indexes" }));
     expect(screen.getByRole("link", { name: "vector store index docs" })).toHaveAttribute(
       "href",
@@ -149,7 +186,7 @@ describe("VectorStoreManagement Indexes tab", () => {
   });
 
   it("should not call indexesListCall until the Indexes tab is clicked", async () => {
-    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" />);
+    render(<VectorStoreManagement accessToken="sk-test" userID="user-1" userRole="Admin" isViewOnly={false} />);
     await waitFor(() => expect(mockVectorStoreListCall).toHaveBeenCalledWith("sk-test"));
     expect(screen.getByRole("tab", { name: "Indexes" })).toBeInTheDocument();
     expect(mockIndexesListCall).not.toHaveBeenCalled();
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.tsx
index 1745c710c51..285a96520bd 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/index.tsx
@@ -24,9 +24,10 @@ interface VectorStoreProps {
   accessToken: string | null;
   userID: string | null;
   userRole: string | null;
+  isViewOnly: boolean;
 }
 
-const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID, userRole }) => {
+const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID, userRole, isViewOnly }) => {
   const [vectorStores, setVectorStores] = useState<VectorStore[]>([]);
   const [isLoadingVectorStores, setIsLoadingVectorStores] = useState(true);
   const [isCreateModalVisible, setIsCreateModalVisible] = useState(false);
@@ -37,7 +38,9 @@ const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID
   const [selectedVectorStoreId, setSelectedVectorStoreId] = useState<string | null>(null);
   const [editVectorStore, setEditVectorStore] = useState(false);
   const [isDeleting, setIsDeleting] = useState(false);
-  const { onTabChange, hasVisited } = useVisitedTabs("create");
+  const canCreateVectorStores = isProxyAdminRole(userRole || "") && !isViewOnly;
+  const defaultTab = canCreateVectorStores ? "create" : "manage";
+  const { onTabChange, hasVisited } = useVisitedTabs(defaultTab);
 
   const fetchVectorStores = async () => {
     if (!accessToken) {
@@ -56,7 +59,7 @@ const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID
   };
 
   const fetchCredentials = async () => {
-    if (!accessToken) return;
+    if (!accessToken || !canCreateVectorStores) return;
     try {
       const response = await credentialListCall(accessToken);
       setCredentials(response.credentials || []);
@@ -153,11 +156,13 @@ const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID
           You can use vector stores to store and retrieve LLM embeddings.
         </p>
 
-        <Tabs defaultValue="create" onValueChange={onTabChange}>
+        <Tabs defaultValue={defaultTab} onValueChange={onTabChange}>
           <TabsList variant="line" className="mb-6 h-auto w-full justify-start rounded-none p-0">
-            <TabsTrigger value="create" className="flex-none rounded-none px-4 py-2">
-              Create Vector Store
-            </TabsTrigger>
+            {canCreateVectorStores && (
+              <TabsTrigger value="create" className="flex-none rounded-none px-4 py-2">
+                Create Vector Store
+              </TabsTrigger>
+            )}
             <TabsTrigger value="manage" className="flex-none rounded-none px-4 py-2">
               Manage Vector Stores
             </TabsTrigger>
@@ -171,14 +176,18 @@ const VectorStoreManagement: React.FC<VectorStoreProps> = ({ accessToken, userID
             )}
           </TabsList>
 
-          <TabsContent keepMounted={hasVisited("create")} value="create">
-            <CreateVectorStore accessToken={accessToken} onSuccess={handleVectorStoreCreated} />
-          </TabsContent>
+          {canCreateVectorStores && (
+            <TabsContent keepMounted={hasVisited("create")} value="create">
+              <CreateVectorStore accessToken={accessToken} onSuccess={handleVectorStoreCreated} />
+            </TabsContent>
+          )}
 
           <TabsContent keepMounted={hasVisited("manage")} value="manage">
-            <Button className="mb-4" onClick={() => setIsCreateModalVisible(true)}>
-              + Add Vector Store
-            </Button>
+            {canCreateVectorStores && (
+              <Button className="mb-4" onClick={() => setIsCreateModalVisible(true)}>
+                + Add Vector Store
+              </Button>
+            )}
 
             <div className="grid grid-cols-1 gap-2 pt-2 pb-2 w-full mt-2">
               <VectorStoreTable
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/page.tsx b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/page.tsx
index 34759152d11..94fedb557a7 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/page.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/vector-stores/page.tsx
@@ -4,6 +4,8 @@ import VectorStoreManagement from "./_components";
 import useAuthorized from "@/app/(dashboard)/hooks/useAuthorized";
 
 export default function VectorStores() {
-  const { accessToken, userRole, userId } = useAuthorized();
-  return <VectorStoreManagement accessToken={accessToken} userRole={userRole} userID={userId} />;
+  const { accessToken, userRole, userId, isViewOnly } = useAuthorized();
+  return (
+    <VectorStoreManagement accessToken={accessToken} userRole={userRole} userID={userId} isViewOnly={isViewOnly} />
+  );
 }

From c949843157f890ea253b51af4fa825e78543feb8 Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 7 Sep 2026 12:24:04 -0700
Subject: [PATCH 78/79] fix(ui): send empty vector_stores when the last team
 vector store is removed (#40144)

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../src/components/team/TeamInfo.test.tsx           | 13 +++++++++++++
 .../src/components/team/TeamInfo.tsx                |  2 +-
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/ui/litellm-dashboard/src/components/team/TeamInfo.test.tsx b/ui/litellm-dashboard/src/components/team/TeamInfo.test.tsx
index 041ba4b548d..b286c8c1303 100644
--- a/ui/litellm-dashboard/src/components/team/TeamInfo.test.tsx
+++ b/ui/litellm-dashboard/src/components/team/TeamInfo.test.tsx
@@ -2052,6 +2052,19 @@ describe("TeamInfoView - the exact bytes the update call sends", () => {
     expect(objectPermission.agent_access_groups).toStrictEqual([]);
   });
 
+  it("sends an empty vector_stores array after the last vector store chip is removed", async () => {
+    const user = userEvent.setup({ delay: null });
+    await openEditor(user);
+
+    await user.click(within(screen.getByLabelText("vs-1")).getByRole("button"));
+    expect(screen.queryByLabelText("vs-1")).not.toBeInTheDocument();
+
+    const payload = await save(user);
+
+    const objectPermission = wireBody(payload).object_permission as Record<string, unknown>;
+    expect(objectPermission.vector_stores).toStrictEqual([]);
+  });
+
   it("resends every stored value once both sections are opened", async () => {
     const user = userEvent.setup({ delay: null });
     await openEditor(user);
diff --git a/ui/litellm-dashboard/src/components/team/TeamInfo.tsx b/ui/litellm-dashboard/src/components/team/TeamInfo.tsx
index efd69a79fc0..4947cec1441 100644
--- a/ui/litellm-dashboard/src/components/team/TeamInfo.tsx
+++ b/ui/litellm-dashboard/src/components/team/TeamInfo.tsx
@@ -1039,7 +1039,7 @@ const TeamInfoView: React.FC<TeamInfoProps> = ({
       delete values.agents_and_groups;
 
       // Handle vector stores permissions
-      if (values.vector_stores && values.vector_stores.length > 0) {
+      if (values.vector_stores) {
         updateData.object_permission.vector_stores = values.vector_stores;
       }
 

From 192e38fa7ba2f529cfaad3bcfc28d2613aca28d5 Mon Sep 17 00:00:00 2001
From: Yassin Kortam <yassin@berri.ai>
Date: Mon, 7 Sep 2026 12:28:38 -0700
Subject: [PATCH 79/79] feat(skills): semantic search over the LiteLLM-hosted
 skill registry (#39401)

* feat(skills): semantic search over the LiteLLM-hosted skill registry

Adds GET /v1/skills?query= (custom_llm_provider=litellm_proxy) and a
skill_search MCP virtual tool, ranking the caller's accessible skills by
semantic similarity, mirroring the A2A agent registry search (LIT-6309).

Also fixes a pre-existing bug where create_skill() dropped description and
instructions for the litellm_proxy provider, which left every LiteLLM-hosted
skill with no searchable text.

* fix(mcp): coerce skill_search top_k instead of raising 500 on malformed input

The MCP-REST skill_search dispatch validated raw tool arguments through a
pydantic model directly, so a non-numeric top_k raised a ValidationError
that the endpoint's catch-all turned into an HTTP 500. Mirrors the
agent_search branch's tolerant coerce_top_k handling instead.

* fix(skills): enforce key limits on search embeddings and bound the semantic index

Semantic search embeddings now run the same pre_call_hook the /embeddings
route runs, so key rate limits, budgets and guardrails apply before the
embedding model is called. The shared SemanticTextIndex caps cached vectors
and evicts the least recently searched entries, and each skill's embedded
text is capped so one skill cannot inflate the embedding batch

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(skills): surface proxy 429s from search embeddings instead of a 503

ProxyRateLimitError is also an OpenAIError, so the search engine was folding
a key rate limit into skill_search_unavailable. Proxy HTTPExceptions now
propagate so the caller gets the same 429 the /embeddings route returns

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(skills): import assert_never from typing_extensions for Python 3.10

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(skills): embed the request as the pre-call hooks returned it, not the original text

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* refactor(skills): keep the litellm_proxy provider check for GET /v1/skills?query= inside llms/

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* test(skills): move the GET /v1/skills?query= endpoint tests under tests/test_litellm/proxy

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .github/workflows/test-unit.yml               |   1 +
 litellm/__init__.py                           |   1 +
 .../llms/litellm_proxy/skills/constants.py    |   5 +
 litellm/llms/litellm_proxy/skills/handler.py  |  19 +-
 .../llms/litellm_proxy/skills/skill_search.py | 161 +++++++
 .../litellm_proxy/skills/transformation.py    |  11 +-
 .../mcp_server/rest_endpoints.py              |  13 +
 .../proxy/_experimental/mcp_server/server.py  |   9 +
 .../_experimental/mcp_server/tool_search.py   |  68 ++-
 litellm/proxy/_lazy_openapi_snapshot.json     |  57 ++-
 litellm/proxy/agent_endpoints/agent_search.py |   7 +-
 litellm/proxy/agent_endpoints/endpoints.py    |   3 +-
 .../anthropic_endpoints/skills_endpoints.py   |  97 +++-
 .../proxy/common_utils/semantic_text_index.py |  79 +++-
 litellm/skills/main.py                        |   6 +-
 litellm/types/llms/anthropic_skills.py        |   9 +
 .../litellm_proxy/skills/test_skill_search.py | 436 ++++++++++++++++++
 .../mcp_server/test_mcp_tool_search.py        |  50 +-
 .../agent_endpoints/test_agent_search.py      |  27 +-
 .../test_skills_endpoints.py                  | 175 +++++++
 tests/test_litellm/skills/test_skills_main.py |  57 +++
 ui/litellm-dashboard/src/lib/http/schema.d.ts |  14 +
 22 files changed, 1259 insertions(+), 46 deletions(-)
 create mode 100644 litellm/llms/litellm_proxy/skills/skill_search.py
 create mode 100644 tests/test_litellm/llms/litellm_proxy/skills/test_skill_search.py
 create mode 100644 tests/test_litellm/proxy/anthropic_endpoints/test_skills_endpoints.py
 create mode 100644 tests/test_litellm/skills/test_skills_main.py

diff --git a/.github/workflows/test-unit.yml b/.github/workflows/test-unit.yml
index 33245ec5b5f..cc606339a20 100644
--- a/.github/workflows/test-unit.yml
+++ b/.github/workflows/test-unit.yml
@@ -116,6 +116,7 @@ jobs:
               tests/test_litellm/rerank_api
               tests/test_litellm/rust_bridge
               tests/test_litellm/sandbox
+              tests/test_litellm/skills
               tests/test_litellm/test_router
               tests/test_litellm/vector_stores
               tests/test_litellm/videos
diff --git a/litellm/__init__.py b/litellm/__init__.py
index 42c0ea881fd..dc2f40af46e 100644
--- a/litellm/__init__.py
+++ b/litellm/__init__.py
@@ -495,6 +495,7 @@ public_model_groups: Optional[List[str]] = None
 public_agent_groups: Optional[List[str]] = None
 agent_search_embedding_model: Optional[str] = None
 mcp_tool_search: Optional[Mapping[str, object]] = None
+skill_search_embedding_model: Optional[str] = None
 # Supports both old format (Dict[str, str]) and new format (Dict[str, Dict[str, Any]])
 # New format: { "displayName": { "url": "...", "index": 0 } }
 # Old format: { "displayName": "url" } (for backward compatibility)
diff --git a/litellm/llms/litellm_proxy/skills/constants.py b/litellm/llms/litellm_proxy/skills/constants.py
index a6c88718f11..04a3a7dbc91 100644
--- a/litellm/llms/litellm_proxy/skills/constants.py
+++ b/litellm/llms/litellm_proxy/skills/constants.py
@@ -16,3 +16,8 @@ DEFAULT_MAX_ITERATIONS: Final[int] = 10
 
 DEFAULT_SANDBOX_TIMEOUT: Final[int] = 120
 """Default timeout in seconds for sandbox code execution."""
+
+MAX_SKILLS_PER_SEARCH: Final[int] = 5000
+"""Upper bound on how many of the caller's accessible skills a single semantic
+search embeds. Ranking runs in memory over this candidate set (no tsvector/DB-side
+filtering yet), so this caps worst-case embedding cost per search request."""
diff --git a/litellm/llms/litellm_proxy/skills/handler.py b/litellm/llms/litellm_proxy/skills/handler.py
index 73f6ed23092..9b625cb0571 100644
--- a/litellm/llms/litellm_proxy/skills/handler.py
+++ b/litellm/llms/litellm_proxy/skills/handler.py
@@ -6,11 +6,15 @@ Used by the transformation layer and skills injection hook.
 """
 
 import uuid
+from collections.abc import Sequence
 from typing import Final
 
 from litellm._logging import verbose_logger
 from litellm.caching.in_memory_cache import InMemoryCache
-from litellm.llms.litellm_proxy.skills.constants import LITELLM_SKILL_ID_PREFIX
+from litellm.llms.litellm_proxy.skills.constants import (
+    LITELLM_SKILL_ID_PREFIX,
+    MAX_SKILLS_PER_SEARCH,
+)
 from litellm.proxy._types import LiteLLM_SkillsTable, NewSkillRequest, UserAPIKeyAuth
 from litellm.proxy.common_utils.resource_ownership import (
     get_primary_resource_owner_scope,
@@ -131,6 +135,19 @@ class LiteLLMSkillsHandler:
         )
         return [_prisma_skill_to_litellm(s) for s in skills]
 
+    @staticmethod
+    async def list_skills_for_search(
+        user_api_key_dict: UserAPIKeyAuth | None = None,
+    ) -> Sequence[LiteLLM_SkillsTable]:
+        """Every skill the caller can access, for ranking. Same owner-scope filter as
+        ``list_skills``, but unpaginated (up to ``MAX_SKILLS_PER_SEARCH``) since a query
+        must be scored against the whole accessible set, not one page of it."""
+        return await LiteLLMSkillsHandler.list_skills(
+            limit=MAX_SKILLS_PER_SEARCH,
+            offset=0,
+            user_api_key_dict=user_api_key_dict,
+        )
+
     @staticmethod
     async def _load_skill(skill_id: str) -> object | None:
         """Cache-first read of the Prisma skill row. Owner-scope filtering
diff --git a/litellm/llms/litellm_proxy/skills/skill_search.py b/litellm/llms/litellm_proxy/skills/skill_search.py
new file mode 100644
index 00000000000..f975c6c4cab
--- /dev/null
+++ b/litellm/llms/litellm_proxy/skills/skill_search.py
@@ -0,0 +1,161 @@
+"""Semantic ranking over the LiteLLM-hosted skill registry, shared by GET /v1/skills?query= and the skill_search MCP tool."""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Final, TypeAlias
+
+from pydantic import BaseModel, ConfigDict
+
+from litellm.llms.litellm_proxy.skills.constants import MAX_SKILLS_PER_SEARCH
+from litellm.llms.litellm_proxy.skills.handler import LiteLLMSkillsHandler
+from litellm.proxy.common_utils.semantic_text_index import (
+    Embedder,
+    EmbeddingFailed,
+    SemanticTextIndex,
+    router_embedder,
+)
+from litellm.types.utils import LlmProviders
+
+if TYPE_CHECKING:
+    from litellm.proxy._types import LiteLLM_SkillsTable, UserAPIKeyAuth
+    from litellm.proxy.utils import ProxyLogging
+    from litellm.router import Router
+
+DEFAULT_SKILL_SEARCH_TOP_K: Final = 5
+MAX_SKILL_SEARCH_TOP_K: Final = 100
+"""Matches the ``le=100`` bound GET /v1/skills?query= enforces via FastAPI's Query
+validation, so the MCP tool can't return a larger payload than the REST endpoint allows."""
+MAX_SKILL_SEARCH_TEXT_CHARS: Final = 4000
+"""Per-skill cap on the title + description + instructions text that gets embedded, so one
+search embeds at most ``MAX_SKILLS_PER_SEARCH * MAX_SKILL_SEARCH_TEXT_CHARS`` characters no
+matter how long the stored instructions are."""
+
+
+@dataclass(frozen=True, slots=True)
+class SkillSearchHit:
+    skill: LiteLLM_SkillsTable
+    score: float
+
+
+@dataclass(frozen=True, slots=True)
+class SkillSearchHits:
+    hits: tuple[SkillSearchHit, ...]
+
+
+@dataclass(frozen=True, slots=True)
+class SkillSearchNotConfigured:
+    reason: str
+
+
+@dataclass(frozen=True, slots=True)
+class SkillSearchEmbeddingFailed:
+    reason: str
+
+
+@dataclass(frozen=True, slots=True)
+class SkillSearchUnsupportedProvider:
+    reason: str
+
+
+SkillSearchOutcome: TypeAlias = SkillSearchHits | SkillSearchNotConfigured | SkillSearchEmbeddingFailed
+HostedSkillSearchOutcome: TypeAlias = SkillSearchOutcome | SkillSearchUnsupportedProvider
+
+
+class SkillSearchResult(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    skill_id: str
+    display_title: str | None
+    description: str | None
+    score: float
+
+
+def skill_search_text(skill: LiteLLM_SkillsTable) -> str:
+    joined: Final = "\n".join(part for part in (skill.display_title, skill.description, skill.instructions) if part)
+    return joined[:MAX_SKILL_SEARCH_TEXT_CHARS]
+
+
+def skill_search_result(hit: SkillSearchHit) -> SkillSearchResult:
+    return SkillSearchResult(
+        skill_id=hit.skill.skill_id,
+        display_title=hit.skill.display_title,
+        description=hit.skill.description,
+        score=hit.score,
+    )
+
+
+class SkillSearchIndex:
+    """Caches one vector per distinct skill text per embedding model, so repeat searches only embed the query."""
+
+    def __init__(self, max_entries: int = MAX_SKILLS_PER_SEARCH) -> None:
+        self._index: Final = SemanticTextIndex(max_entries=max_entries)
+
+    async def search(
+        self,
+        query: str,
+        skills: Sequence[LiteLLM_SkillsTable],
+        top_k: int,
+        embed: Embedder,
+        embedding_model: str,
+    ) -> SkillSearchHits | SkillSearchEmbeddingFailed:
+        texts: Final = tuple(skill_search_text(skill) for skill in skills)
+        scores: Final = await self._index.scores(query, texts, embed, embedding_model)
+        if isinstance(scores, EmbeddingFailed):
+            return SkillSearchEmbeddingFailed(reason=scores.reason)
+        ranked: Final = sorted(
+            (SkillSearchHit(skill=skill, score=score) for skill, score in zip(skills, scores, strict=True)),
+            key=lambda hit: hit.score,
+            reverse=True,
+        )
+        return SkillSearchHits(hits=tuple(ranked[:top_k]))
+
+
+global_skill_search_index: Final = SkillSearchIndex()
+
+
+async def search_skills(
+    query: str,
+    skills: Sequence[LiteLLM_SkillsTable],
+    top_k: int,
+    router: Router | None,
+    embedding_model: str | None,
+    index: SkillSearchIndex,
+    user_api_key_dict: UserAPIKeyAuth,
+    proxy_logging_obj: ProxyLogging,
+) -> SkillSearchOutcome:
+    if embedding_model is None:
+        return SkillSearchNotConfigured(
+            reason="skill search needs litellm_settings.skill_search_embedding_model set to an embedding model from model_list"
+        )
+    if router is None:
+        return SkillSearchNotConfigured(reason="skill search needs a model_list so the embedding model can be called")
+    embed: Final = router_embedder(router, embedding_model, user_api_key_dict, proxy_logging_obj)
+    return await index.search(query, skills, top_k, embed, embedding_model)
+
+
+async def search_hosted_skills(
+    custom_llm_provider: str | None,
+    query: str,
+    top_k: int,
+    router: Router | None,
+    embedding_model: str | None,
+    index: SkillSearchIndex,
+    user_api_key_dict: UserAPIKeyAuth,
+    proxy_logging_obj: ProxyLogging,
+) -> HostedSkillSearchOutcome:
+    """GET /v1/skills?query= for the skills LiteLLM hosts itself: only ``litellm_proxy`` has a registry to rank."""
+    if custom_llm_provider != LlmProviders.LITELLM_PROXY.value:
+        return SkillSearchUnsupportedProvider(reason="query is only supported for custom_llm_provider=litellm_proxy")
+    skills: Final = await LiteLLMSkillsHandler.list_skills_for_search(user_api_key_dict=user_api_key_dict)
+    return await search_skills(
+        query=query,
+        skills=skills,
+        top_k=top_k,
+        router=router,
+        embedding_model=embedding_model,
+        index=index,
+        user_api_key_dict=user_api_key_dict,
+        proxy_logging_obj=proxy_logging_obj,
+    )
diff --git a/litellm/llms/litellm_proxy/skills/transformation.py b/litellm/llms/litellm_proxy/skills/transformation.py
index c972dc349c9..9fc2d2cbb45 100644
--- a/litellm/llms/litellm_proxy/skills/transformation.py
+++ b/litellm/llms/litellm_proxy/skills/transformation.py
@@ -154,7 +154,7 @@ class LiteLLMSkillsTransformationHandler:
             user_api_key_dict=user_api_key_dict,
         )
 
-        return self._db_skill_to_response(db_skill)
+        return self.db_skill_to_response(db_skill)
 
     def list_skills_handler(
         self,
@@ -222,7 +222,9 @@ class LiteLLMSkillsTransformationHandler:
             user_api_key_dict=user_api_key_dict,
         )
 
-        skills: Final = [self._db_skill_to_response(s) for s in db_skills]
+        skills: Final = [  # mutable-ok: ListSkillsResponse.data needs list[Skill]; never mutated after
+            self.db_skill_to_response(s) for s in db_skills
+        ]
         return ListSkillsResponse(
             data=skills,
             has_more=len(skills) >= limit,
@@ -288,7 +290,7 @@ class LiteLLMSkillsTransformationHandler:
             skill_id=skill_id,
             user_api_key_dict=user_api_key_dict,
         )
-        return self._db_skill_to_response(db_skill)
+        return self.db_skill_to_response(db_skill)
 
     def delete_skill_handler(
         self,
@@ -354,7 +356,7 @@ class LiteLLMSkillsTransformationHandler:
             type=result.get("type", "skill_deleted"),
         )
 
-    def _db_skill_to_response(self, db_skill: "LiteLLM_SkillsTable") -> Skill:
+    def db_skill_to_response(self, db_skill: "LiteLLM_SkillsTable") -> Skill:
         """
         Convert a database skill record to Anthropic-compatible Skill response.
 
@@ -375,4 +377,5 @@ class LiteLLMSkillsTransformationHandler:
             latest_version=db_skill.latest_version,
             source=db_skill.source or "custom",
             type="skill",
+            description=db_skill.description,
         )
diff --git a/litellm/proxy/_experimental/mcp_server/rest_endpoints.py b/litellm/proxy/_experimental/mcp_server/rest_endpoints.py
index 5fbfad54a39..102129ffdd0 100644
--- a/litellm/proxy/_experimental/mcp_server/rest_endpoints.py
+++ b/litellm/proxy/_experimental/mcp_server/rest_endpoints.py
@@ -104,6 +104,9 @@ if MCP_AVAILABLE:
     from mcp.types import Tool as MCPTool
 
     from litellm.experimental_mcp_client.client import MCPClient
+    from litellm.llms.litellm_proxy.skills.skill_search import (
+        DEFAULT_SKILL_SEARCH_TOP_K,
+    )
     from litellm.proxy._experimental.mcp_server.mcp_server_manager import (
         _UPSTREAM_OAUTH_DISCOVERY_AUTH_TYPES,
         global_mcp_server_manager,
@@ -188,10 +191,12 @@ if MCP_AVAILABLE:
             AGENT_SEARCH_TOOL_NAME,
             DEFAULT_AGENT_SEARCH_TOP_K,
             MCP_TOOL_SEARCH_TOOL_NAME,
+            SKILL_SEARCH_TOOL_NAME,
             coerce_top_k,
             handle_agent_search,
             handle_mcp_tool_call,
             handle_mcp_tool_search,
+            handle_skill_search,
         )
         from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
         from litellm.proxy.proxy_server import general_settings, proxy_config, proxy_logging_obj
@@ -210,6 +215,14 @@ if MCP_AVAILABLE:
                 ),
                 user_api_key_dict=user_api_key_dict,
             )
+        if tool_name == SKILL_SEARCH_TOOL_NAME:
+            return await handle_skill_search(
+                query=str(tool_arguments.get("query", "")),
+                top_k=coerce_top_k(
+                    tool_arguments.get("top_k", DEFAULT_SKILL_SEARCH_TOP_K), default=DEFAULT_SKILL_SEARCH_TOP_K
+                ),
+                user_api_key_dict=user_api_key_dict,
+            )
         rest_client_ip: Final = IPAddressUtils.get_mcp_client_ip(request)
         (
             virtual_mcp_auth_header,
diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py
index 60a9af89cc3..0b424c31c4b 100644
--- a/litellm/proxy/_experimental/mcp_server/server.py
+++ b/litellm/proxy/_experimental/mcp_server/server.py
@@ -911,15 +911,18 @@ if MCP_AVAILABLE:
         Returns a CallToolResult when ``name`` is a virtual tool, else ``None`` so
         the caller falls through to normal tool routing.
         """
+        from litellm.llms.litellm_proxy.skills.skill_search import DEFAULT_SKILL_SEARCH_TOP_K
         from litellm.proxy._experimental.mcp_server.tool_search import (
             AGENT_SEARCH_TOOL_NAME,
             DEFAULT_AGENT_SEARCH_TOP_K,
             MCP_TOOL_SEARCH_TOOL_NAME,
+            SKILL_SEARCH_TOOL_NAME,
             VIRTUAL_TOOL_NAMES,
             coerce_top_k,
             handle_agent_search,
             handle_mcp_tool_call,
             handle_mcp_tool_search,
+            handle_skill_search,
         )
 
         if name not in VIRTUAL_TOOL_NAMES:
@@ -961,6 +964,12 @@ if MCP_AVAILABLE:
                 top_k=coerce_top_k(args.get("top_k", DEFAULT_AGENT_SEARCH_TOP_K), default=DEFAULT_AGENT_SEARCH_TOP_K),
                 user_api_key_dict=user_api_key_auth,
             )
+        if name == SKILL_SEARCH_TOOL_NAME:
+            return await handle_skill_search(
+                query=str(args.get("query", "")),
+                top_k=coerce_top_k(args.get("top_k", DEFAULT_SKILL_SEARCH_TOP_K), default=DEFAULT_SKILL_SEARCH_TOP_K),
+                user_api_key_dict=user_api_key_auth,
+            )
         virtual_logging_obj: Final = await _build_virtual_call_logging_obj(
             name=name,
             arguments=args,
diff --git a/litellm/proxy/_experimental/mcp_server/tool_search.py b/litellm/proxy/_experimental/mcp_server/tool_search.py
index af02c11ad86..f19340d30cb 100644
--- a/litellm/proxy/_experimental/mcp_server/tool_search.py
+++ b/litellm/proxy/_experimental/mcp_server/tool_search.py
@@ -11,6 +11,7 @@ from pydantic import ValidationError
 from typing_extensions import ReadOnly, Required, assert_never
 
 import litellm
+from litellm.llms.litellm_proxy.skills.skill_search import DEFAULT_SKILL_SEARCH_TOP_K
 from litellm.proxy.agent_endpoints.agent_search import DEFAULT_AGENT_SEARCH_TOP_K
 from litellm.proxy.common_utils.semantic_text_index import (
     Embedder,
@@ -30,7 +31,10 @@ MCP_TOOL_SEARCH_SETTINGS_KEY: Final[str] = "mcp_tool_search"
 MCP_TOOL_SEARCH_TOOL_NAME: Final[str] = "mcp_tool_search"
 MCP_TOOL_CALL_TOOL_NAME: Final[str] = "mcp_tool_call"
 AGENT_SEARCH_TOOL_NAME: Final[str] = "agent_search"
-VIRTUAL_TOOL_NAMES: Final = frozenset((MCP_TOOL_SEARCH_TOOL_NAME, MCP_TOOL_CALL_TOOL_NAME, AGENT_SEARCH_TOOL_NAME))
+SKILL_SEARCH_TOOL_NAME: Final[str] = "skill_search"
+VIRTUAL_TOOL_NAMES: Final = frozenset(
+    (MCP_TOOL_SEARCH_TOOL_NAME, MCP_TOOL_CALL_TOOL_NAME, AGENT_SEARCH_TOOL_NAME, SKILL_SEARCH_TOOL_NAME)
+)
 
 
 def coerce_top_k(value: Any, default: int = 5) -> int:
@@ -199,8 +203,28 @@ _AGENT_SEARCH_DEFINITION: Final[VirtualToolDefinition] = {
 }
 
 
+_SKILL_SEARCH_DEFINITION: Final[VirtualToolDefinition] = {
+    "name": SKILL_SEARCH_TOOL_NAME,
+    "description": "Find registered skills by describing what you need in natural language. Returns the best "
+    "matching skills you can access, ranked by semantic similarity, each with its skill_id, display_title, "
+    "description, and score.",
+    "inputSchema": {
+        "type": "object",
+        "properties": {
+            "query": {"type": "string", "description": "What you need the skill to do, in natural language."},
+            "top_k": {
+                "type": "integer",
+                "description": "Maximum number of skills to return.",
+                "default": DEFAULT_SKILL_SEARCH_TOP_K,
+            },
+        },
+        "required": _json_array("query"),
+    },
+}
+
+
 def get_virtual_tool_definitions() -> tuple[VirtualToolDefinition, ...]:
-    return (_MCP_TOOL_SEARCH_DEFINITION, _MCP_TOOL_CALL_DEFINITION, _AGENT_SEARCH_DEFINITION)
+    return (_MCP_TOOL_SEARCH_DEFINITION, _MCP_TOOL_CALL_DEFINITION, _AGENT_SEARCH_DEFINITION, _SKILL_SEARCH_DEFINITION)
 
 
 def _text_tool_result(text: str, is_error: bool) -> CallToolResult:
@@ -223,7 +247,7 @@ async def handle_agent_search(query: str, top_k: int, user_api_key_dict: UserAPI
     )
     from litellm.proxy.agent_endpoints.auth.agent_permission_handler import accessible_agents
     from litellm.proxy.common_utils.rbac_utils import check_feature_access_for_user
-    from litellm.proxy.proxy_server import llm_router
+    from litellm.proxy.proxy_server import llm_router, proxy_logging_obj
 
     await check_feature_access_for_user(user_api_key_dict, "agents")
     outcome: Final = await search_agents(
@@ -234,6 +258,7 @@ async def handle_agent_search(query: str, top_k: int, user_api_key_dict: UserAPI
         embedding_model=litellm.agent_search_embedding_model,
         index=global_agent_search_index,
         user_api_key_dict=user_api_key_dict,
+        proxy_logging_obj=proxy_logging_obj,
     )
     match outcome:
         case AgentSearchHits(hits):
@@ -245,6 +270,39 @@ async def handle_agent_search(query: str, top_k: int, user_api_key_dict: UserAPI
             assert_never(outcome)
 
 
+async def handle_skill_search(query: str, top_k: int, user_api_key_dict: UserAPIKeyAuth) -> CallToolResult:
+    from litellm.llms.litellm_proxy.skills.handler import LiteLLMSkillsHandler
+    from litellm.llms.litellm_proxy.skills.skill_search import (
+        MAX_SKILL_SEARCH_TOP_K,
+        SkillSearchEmbeddingFailed,
+        SkillSearchHits,
+        SkillSearchNotConfigured,
+        global_skill_search_index,
+        search_skills,
+        skill_search_result,
+    )
+    from litellm.proxy.proxy_server import llm_router, proxy_logging_obj
+
+    outcome: Final = await search_skills(
+        query=query,
+        skills=await LiteLLMSkillsHandler.list_skills_for_search(user_api_key_dict),
+        top_k=min(max(top_k, 1), MAX_SKILL_SEARCH_TOP_K),
+        router=llm_router,
+        embedding_model=litellm.skill_search_embedding_model,
+        index=global_skill_search_index,
+        user_api_key_dict=user_api_key_dict,
+        proxy_logging_obj=proxy_logging_obj,
+    )
+    match outcome:
+        case SkillSearchHits(hits):
+            results: Final = tuple(skill_search_result(hit).model_dump() for hit in hits)
+            return _text_tool_result(json.dumps(results), is_error=False)
+        case SkillSearchNotConfigured(reason) | SkillSearchEmbeddingFailed(reason):
+            return _text_tool_result(reason, is_error=True)
+        case _:
+            assert_never(outcome)
+
+
 async def handle_mcp_tool_search(
     query: str,
     top_k: int,
@@ -257,7 +315,7 @@ async def handle_mcp_tool_search(
     raw_headers: dict[str, str] | None = None,
 ) -> CallToolResult:
     from litellm.proxy._experimental.mcp_server.server import _list_mcp_tools
-    from litellm.proxy.proxy_server import llm_router
+    from litellm.proxy.proxy_server import llm_router, proxy_logging_obj
 
     settings: Final = mcp_tool_search_settings()
     if isinstance(settings, ValidationError):
@@ -271,7 +329,7 @@ async def handle_mcp_tool_search(
         )
     ranker: Final = (
         SemanticToolRanker(
-            embed=router_embedder(llm_router, settings.embedding_model, user_api_key_dict),
+            embed=router_embedder(llm_router, settings.embedding_model, user_api_key_dict, proxy_logging_obj),
             embedding_model=settings.embedding_model,
             index=global_mcp_tool_search_index,
         )
diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json
index 91a97ad6544..4093f2c5248 100644
--- a/litellm/proxy/_lazy_openapi_snapshot.json
+++ b/litellm/proxy/_lazy_openapi_snapshot.json
@@ -4687,6 +4687,17 @@
               "title": "Created At",
               "type": "string"
             },
+            "description": {
+              "anyOf": [
+                {
+                  "type": "string"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Description"
+            },
             "display_title": {
               "anyOf": [
                 {
@@ -4713,6 +4724,17 @@
               ],
               "title": "Latest Version"
             },
+            "search_score": {
+              "anyOf": [
+                {
+                  "type": "number"
+                },
+                {
+                  "type": "null"
+                }
+              ],
+              "title": "Search Score"
+            },
             "source": {
               "title": "Source",
               "type": "string"
@@ -4781,7 +4803,7 @@
     "paths": {
       "/v1/skills": {
         "get": {
-          "description": "List skills on Anthropic.\n\nRequires `?beta=true` query parameter.\n\nModel-based routing (for multi-account support):\n- Pass model via header: `x-litellm-model: claude-account-1`\n- Pass model via query: `?model=claude-account-1`\n- Pass model via body: `{\"model\": \"claude-account-1\"}`\n\nExample usage:\n```bash\n# Basic usage\ncurl \"http://localhost:4000/v1/skills?beta=true&limit=10\"       -H \"Authorization: Bearer your-key\"\n\n# With model-based routing\ncurl \"http://localhost:4000/v1/skills?beta=true&limit=10\"       -H \"Authorization: Bearer your-key\"       -H \"x-litellm-model: claude-account-1\"\n```\n\nReturns: ListSkillsResponse with list of skills",
+          "description": "List skills on Anthropic.\n\nRequires `?beta=true` query parameter.\n\nModel-based routing (for multi-account support):\n- Pass model via header: `x-litellm-model: claude-account-1`\n- Pass model via query: `?model=claude-account-1`\n- Pass model via body: `{\"model\": \"claude-account-1\"}`\n\nExample usage:\n```bash\n# Basic usage\ncurl \"http://localhost:4000/v1/skills?beta=true&limit=10\"       -H \"Authorization: Bearer your-key\"\n\n# With model-based routing\ncurl \"http://localhost:4000/v1/skills?beta=true&limit=10\"       -H \"Authorization: Bearer your-key\"       -H \"x-litellm-model: claude-account-1\"\n```\n\nPass `?custom_llm_provider=litellm_proxy&query=<task>` to rank the LiteLLM-hosted skills you can\naccess by semantic similarity instead of paging through the whole registry:\n```bash\ncurl \"http://localhost:4000/v1/skills?custom_llm_provider=litellm_proxy&query=summarize+a+pdf&top_k=5\"       -H \"Authorization: Bearer your-key\"\n```\n\nReturns: ListSkillsResponse with list of skills",
           "operationId": "list_skills_v1_skills_get",
           "parameters": [
             {
@@ -4849,6 +4871,39 @@
                 "default": "anthropic",
                 "title": "Custom Llm Provider"
               }
+            },
+            {
+              "description": "Describe what you need in natural language to rank the skills you can access by semantic similarity over their title and description. Each result carries a search_score. Only supported for custom_llm_provider=litellm_proxy. Requires litellm_settings.skill_search_embedding_model.",
+              "in": "query",
+              "name": "query",
+              "required": false,
+              "schema": {
+                "anyOf": [
+                  {
+                    "minLength": 1,
+                    "type": "string"
+                  },
+                  {
+                    "type": "null"
+                  }
+                ],
+                "description": "Describe what you need in natural language to rank the skills you can access by semantic similarity over their title and description. Each result carries a search_score. Only supported for custom_llm_provider=litellm_proxy. Requires litellm_settings.skill_search_embedding_model.",
+                "title": "Query"
+              }
+            },
+            {
+              "description": "With query: the maximum number of ranked skills to return.",
+              "in": "query",
+              "name": "top_k",
+              "required": false,
+              "schema": {
+                "default": 5,
+                "description": "With query: the maximum number of ranked skills to return.",
+                "maximum": 100,
+                "minimum": 1,
+                "title": "Top K",
+                "type": "integer"
+              }
             }
           ],
           "responses": {
diff --git a/litellm/proxy/agent_endpoints/agent_search.py b/litellm/proxy/agent_endpoints/agent_search.py
index 76e3fe6c5ad..65a89bb2c7a 100644
--- a/litellm/proxy/agent_endpoints/agent_search.py
+++ b/litellm/proxy/agent_endpoints/agent_search.py
@@ -18,6 +18,7 @@ from litellm.types.agents import AgentResponse
 
 if TYPE_CHECKING:
     from litellm.proxy._types import UserAPIKeyAuth
+    from litellm.proxy.utils import ProxyLogging
     from litellm.router import Router
 
 DEFAULT_AGENT_SEARCH_TOP_K: Final = 5
@@ -132,6 +133,7 @@ async def search_agents(
     embedding_model: str | None,
     index: AgentSearchIndex,
     user_api_key_dict: UserAPIKeyAuth,
+    proxy_logging_obj: ProxyLogging,
 ) -> AgentSearchOutcome:
     if embedding_model is None:
         return AgentSearchNotConfigured(
@@ -139,6 +141,5 @@ async def search_agents(
         )
     if router is None:
         return AgentSearchNotConfigured(reason="agent search needs a model_list so the embedding model can be called")
-    return await index.search(
-        query, agents, top_k, router_embedder(router, embedding_model, user_api_key_dict), embedding_model
-    )
+    embed: Final = router_embedder(router, embedding_model, user_api_key_dict, proxy_logging_obj)
+    return await index.search(query, agents, top_k, embed, embedding_model)
diff --git a/litellm/proxy/agent_endpoints/endpoints.py b/litellm/proxy/agent_endpoints/endpoints.py
index cc17672553b..aa8979a73c6 100644
--- a/litellm/proxy/agent_endpoints/endpoints.py
+++ b/litellm/proxy/agent_endpoints/endpoints.py
@@ -249,7 +249,7 @@ def _agent_search_error(status_code: int, error: str, message: str) -> HTTPExcep
 async def _rank_agents_by_query(
     query: str, agents: Sequence[AgentResponse], top_k: int, user_api_key_dict: UserAPIKeyAuth
 ) -> tuple[AgentResponse, ...]:
-    from litellm.proxy.proxy_server import llm_router
+    from litellm.proxy.proxy_server import llm_router, proxy_logging_obj
 
     outcome: Final = await search_agents(
         query=query,
@@ -259,6 +259,7 @@ async def _rank_agents_by_query(
         embedding_model=litellm.agent_search_embedding_model,
         index=global_agent_search_index,
         user_api_key_dict=user_api_key_dict,
+        proxy_logging_obj=proxy_logging_obj,
     )
     match outcome:
         case AgentSearchHits(hits):
diff --git a/litellm/proxy/anthropic_endpoints/skills_endpoints.py b/litellm/proxy/anthropic_endpoints/skills_endpoints.py
index 9390bf4c537..4426c0b547a 100644
--- a/litellm/proxy/anthropic_endpoints/skills_endpoints.py
+++ b/litellm/proxy/anthropic_endpoints/skills_endpoints.py
@@ -2,11 +2,23 @@
 Anthropic Skills API endpoints - /v1/skills
 """
 
-from typing import Final
+from types import MappingProxyType
+from typing import Annotated, Final
 
 import orjson
-from fastapi import APIRouter, Depends, Request, Response
+from fastapi import APIRouter, Depends, HTTPException, Query, Request, Response
+from typing_extensions import ReadOnly, TypedDict, assert_never
 
+import litellm
+from litellm.llms.litellm_proxy.skills.skill_search import (
+    DEFAULT_SKILL_SEARCH_TOP_K,
+    SkillSearchEmbeddingFailed,
+    SkillSearchHits,
+    SkillSearchNotConfigured,
+    SkillSearchUnsupportedProvider,
+    global_skill_search_index,
+    search_hosted_skills,
+)
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing
@@ -23,6 +35,51 @@ from litellm.types.llms.anthropic_skills import (
 router: Final = APIRouter()
 
 
+class _SkillSearchErrorDetail(TypedDict):
+    error: ReadOnly[str]
+    message: ReadOnly[str]
+
+
+def _skill_search_error(status_code: int, error: str, message: str) -> HTTPException:
+    detail: Final[_SkillSearchErrorDetail] = {"error": error, "message": message}
+    return HTTPException(status_code=status_code, detail=detail)
+
+
+async def _search_skills(
+    custom_llm_provider: str | None, query: str, top_k: int, user_api_key_dict: UserAPIKeyAuth
+) -> ListSkillsResponse:
+    from litellm.llms.litellm_proxy.skills.transformation import (
+        LiteLLMSkillsTransformationHandler,
+    )
+    from litellm.proxy.proxy_server import llm_router, proxy_logging_obj
+
+    outcome: Final = await search_hosted_skills(
+        custom_llm_provider=custom_llm_provider,
+        query=query,
+        top_k=top_k,
+        router=llm_router,
+        embedding_model=litellm.skill_search_embedding_model,
+        index=global_skill_search_index,
+        user_api_key_dict=user_api_key_dict,
+        proxy_logging_obj=proxy_logging_obj,
+    )
+    to_response: Final = LiteLLMSkillsTransformationHandler().db_skill_to_response
+    match outcome:
+        case SkillSearchHits(hits):
+            skills: Final = [  # mutable-ok: ListSkillsResponse.data requires list[Skill]; never mutated after
+                to_response(hit.skill).model_copy(update=MappingProxyType({"search_score": hit.score})) for hit in hits
+            ]
+            return ListSkillsResponse(data=skills, has_more=False, next_page=None)
+        case SkillSearchUnsupportedProvider(reason):
+            raise _skill_search_error(400, "skill_search_unsupported_provider", reason)
+        case SkillSearchNotConfigured(reason):
+            raise _skill_search_error(400, "skill_search_not_configured", reason)
+        case SkillSearchEmbeddingFailed(reason):
+            raise _skill_search_error(503, "skill_search_unavailable", reason)
+        case _:
+            assert_never(outcome)
+
+
 @router.post(
     "/v1/skills",
     tags=["[beta] Anthropic Skills API"],
@@ -134,32 +191,58 @@ async def list_skills(
     after_id: str | None = None,
     before_id: str | None = None,
     custom_llm_provider: str | None = "anthropic",
+    query: Annotated[
+        str | None,
+        Query(
+            min_length=1,
+            description="Describe what you need in natural language to rank the skills you can access by "
+            "semantic similarity over their title and description. Each result carries a search_score. "
+            "Only supported for custom_llm_provider=litellm_proxy. Requires "
+            "litellm_settings.skill_search_embedding_model.",
+        ),
+    ] = None,
+    top_k: Annotated[
+        int,
+        Query(ge=1, le=100, description="With query: the maximum number of ranked skills to return."),
+    ] = DEFAULT_SKILL_SEARCH_TOP_K,
     user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth),
 ):
     """
     List skills on Anthropic.
-    
+
     Requires `?beta=true` query parameter.
-    
+
     Model-based routing (for multi-account support):
     - Pass model via header: `x-litellm-model: claude-account-1`
     - Pass model via query: `?model=claude-account-1`
     - Pass model via body: `{"model": "claude-account-1"}`
-    
+
     Example usage:
     ```bash
     # Basic usage
     curl "http://localhost:4000/v1/skills?beta=true&limit=10" \
       -H "Authorization: Bearer your-key"
-    
+
     # With model-based routing
     curl "http://localhost:4000/v1/skills?beta=true&limit=10" \
       -H "Authorization: Bearer your-key" \
       -H "x-litellm-model: claude-account-1"
     ```
-    
+
+    Pass `?custom_llm_provider=litellm_proxy&query=<task>` to rank the LiteLLM-hosted skills you can
+    access by semantic similarity instead of paging through the whole registry:
+    ```bash
+    curl "http://localhost:4000/v1/skills?custom_llm_provider=litellm_proxy&query=summarize+a+pdf&top_k=5" \
+      -H "Authorization: Bearer your-key"
+    ```
+
     Returns: ListSkillsResponse with list of skills
     """
+    if query is not None:
+        return await _search_skills(
+            custom_llm_provider=custom_llm_provider, query=query, top_k=top_k, user_api_key_dict=user_api_key_dict
+        )
+
     from litellm.proxy.proxy_server import (
         general_settings,
         llm_router,
diff --git a/litellm/proxy/common_utils/semantic_text_index.py b/litellm/proxy/common_utils/semantic_text_index.py
index 0820459af49..b8d3595163e 100644
--- a/litellm/proxy/common_utils/semantic_text_index.py
+++ b/litellm/proxy/common_utils/semantic_text_index.py
@@ -5,10 +5,11 @@ from __future__ import annotations
 import math
 from collections.abc import Awaitable, Mapping, Sequence
 from dataclasses import dataclass
-from itertools import chain
+from itertools import chain, islice
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Final, Protocol, TypeAlias
 
+from fastapi import HTTPException
 from openai import OpenAIError
 from pydantic import BaseModel, ConfigDict
 
@@ -16,10 +17,14 @@ from litellm.exceptions import BudgetExceededError
 
 if TYPE_CHECKING:
     from litellm.proxy._types import UserAPIKeyAuth
+    from litellm.proxy.utils import ProxyLogging
     from litellm.router import Router
 
 Vector: TypeAlias = tuple[float, ...]
 
+DEFAULT_MAX_CACHED_VECTORS: Final = 5000
+"""Ceiling on how many (embedding model, text) vectors one index keeps; the least recently searched are evicted first."""
+
 
 class Embedder(Protocol):
     def __call__(self, texts: Sequence[str]) -> Awaitable[Sequence[Vector]]: ...
@@ -42,6 +47,16 @@ class _EmbeddingData(BaseModel):
     data: tuple[_EmbeddingItem, ...]
 
 
+class _EmbeddingRequest(BaseModel):
+    """The /embeddings-shaped request as the pre-call hooks (rate limits, budgets, guardrails) hand it back."""
+
+    model_config = ConfigDict(frozen=True, extra="ignore")
+
+    model: str
+    input: tuple[str, ...]
+    metadata: dict[str, object]  # mutable-ok: the router mutates the metadata dict it is handed
+
+
 def cosine_similarity(left: Vector, right: Vector) -> float:
     dot: Final = sum(a * b for a, b in zip(left, right, strict=True))
     norms: Final = math.sqrt(sum(a * a for a in left)) * math.sqrt(sum(b * b for b in right))
@@ -57,23 +72,40 @@ def embedding_spend_metadata(user_api_key_dict: UserAPIKeyAuth) -> dict[str, obj
     }
 
 
-def router_embedder(router: Router, embedding_model: str, user_api_key_dict: UserAPIKeyAuth) -> Embedder:
+def router_embedder(
+    router: Router, embedding_model: str, user_api_key_dict: UserAPIKeyAuth, proxy_logging_obj: ProxyLogging
+) -> Embedder:
+    """Embeds through the router after the same key rate-limit, budget and guardrail pre-call hooks /embeddings runs."""
+
     async def embed(texts: Sequence[str]) -> Sequence[Vector]:
-        batch: Final = list(texts)  # mutable-ok: Router.aembedding accepts only str | list input
+        request: Final = {  # mutable-ok: pre_call_hook mutates the request dict in place
+            "model": embedding_model,
+            "input": list(texts),  # mutable-ok: Router.aembedding accepts only str | list input
+            "metadata": embedding_spend_metadata(user_api_key_dict),
+        }
+        processed: Final = _EmbeddingRequest.model_validate(
+            await proxy_logging_obj.pre_call_hook(
+                user_api_key_dict=user_api_key_dict, data=request, call_type="aembedding"
+            )
+        )
         response: Final = await router.aembedding(
-            model=embedding_model, input=batch, metadata=embedding_spend_metadata(user_api_key_dict)
+            model=processed.model,
+            input=list(processed.input),  # mutable-ok: Router.aembedding accepts only str | list input
+            metadata=processed.metadata,
         )
         return tuple(item.embedding for item in _EmbeddingData.model_validate(response.model_dump()).data)
 
     return embed
 
 
-_NO_VECTORS: Final[Mapping[str, Vector]] = MappingProxyType({})
+_CacheKey: TypeAlias = tuple[str, str]
 
 
 async def _embed_all(embed: Embedder, texts: Sequence[str]) -> tuple[Vector, ...] | EmbeddingFailed:
     try:
         vectors: Final = tuple(await embed(texts))
+    except HTTPException:
+        raise
     except (OpenAIError, ValueError, BudgetExceededError) as exc:
         return EmbeddingFailed(reason=f"embedding the search query failed: {exc}")
     if len(vectors) != len(texts):
@@ -111,20 +143,34 @@ async def _embed_query_and_texts(
 
 
 class SemanticTextIndex:
-    """Caches one vector per distinct text per embedding model, so repeat searches only embed the query."""
+    """Caches one vector per distinct text per embedding model, so repeat searches only embed the query.
 
-    def __init__(self) -> None:
-        self._vectors: Mapping[str, Mapping[str, Vector]] = MappingProxyType({})
+    Holds at most ``max_entries`` vectors across all models: once full, the texts no recent search touched go first."""
 
-    def _merged(self, embedding_model: str, embedded: _Embedded) -> Mapping[str, Vector]:
-        kept: Final = MappingProxyType(
+    def __init__(self, max_entries: int = DEFAULT_MAX_CACHED_VECTORS) -> None:
+        self._max_entries: Final = max_entries
+        self._vectors: Mapping[_CacheKey, Vector] = MappingProxyType({})
+
+    def _cached(self, embedding_model: str) -> Mapping[str, Vector]:
+        return MappingProxyType(
+            {text: vector for (model, text), vector in self._vectors.items() if model == embedding_model}
+        )
+
+    def _merged(self, embedding_model: str, embedded: _Embedded, texts: Sequence[str]) -> Mapping[_CacheKey, Vector]:
+        dimension: Final = len(embedded.query_vector)
+        touched: Final = MappingProxyType({(embedding_model, text): embedded.vectors[text] for text in texts})
+        untouched: Final = MappingProxyType(
             {
-                text: vector
-                for text, vector in self._vectors.get(embedding_model, _NO_VECTORS).items()
-                if len(vector) == len(embedded.query_vector)
+                key: vector
+                for key, vector in chain(
+                    self._vectors.items(),
+                    (((embedding_model, text), vector) for text, vector in embedded.vectors.items()),
+                )
+                if key not in touched and (key[0] != embedding_model or len(vector) == dimension)
             }
         )
-        return MappingProxyType({**kept, **embedded.vectors})
+        ordered: Final = MappingProxyType({**untouched, **touched})
+        return MappingProxyType(dict(islice(ordered.items(), max(len(ordered) - self._max_entries, 0), None)))
 
     async def scores(
         self, query: str, texts: Sequence[str], embed: Embedder, embedding_model: str
@@ -132,11 +178,10 @@ class SemanticTextIndex:
         """Cosine similarity of `query` to each entry of `texts`, in the same order."""
         if not texts:
             return ()
-        cached: Final = self._vectors.get(embedding_model, _NO_VECTORS)
-        embedded: Final = await _embed_query_and_texts(embed, query, texts, cached)
+        embedded: Final = await _embed_query_and_texts(embed, query, texts, self._cached(embedding_model))
         if isinstance(embedded, EmbeddingFailed):
             return embedded
         if not _same_dimension(embedded.query_vector, embedded.vectors, texts):
             return EmbeddingFailed(reason=f"embedding model {embedding_model} returned vectors of mixed dimensions")
-        self._vectors = MappingProxyType({**self._vectors, embedding_model: self._merged(embedding_model, embedded)})
+        self._vectors = self._merged(embedding_model, embedded, texts)
         return tuple(cosine_similarity(embedded.query_vector, embedded.vectors[text]) for text in texts)
diff --git a/litellm/skills/main.py b/litellm/skills/main.py
index 002419dbad4..71fd78f11a3 100644
--- a/litellm/skills/main.py
+++ b/litellm/skills/main.py
@@ -182,10 +182,14 @@ def create_skill(
         if extra_body:
             create_request.update(extra_body)
 
-        # Route to LiteLLM DB if custom_llm_provider="litellm_proxy"
+        # Route to LiteLLM DB if custom_llm_provider="litellm_proxy". description/instructions
+        # arrive as top-level kwargs from the REST form endpoint, or nested in extra_body from
+        # the SDK convention used by other providers' create_request above.
         if custom_llm_provider == LlmProviders.LITELLM_PROXY.value:
             return _get_litellm_skills_handler().create_skill_handler(
                 display_title=display_title,
+                description=kwargs.get("description") or (extra_body.get("description") if extra_body else None),
+                instructions=kwargs.get("instructions") or (extra_body.get("instructions") if extra_body else None),
                 files=files,
                 metadata=_get_skill_request_metadata(kwargs, extra_body),
                 user_id=kwargs.get("user_id"),
diff --git a/litellm/types/llms/anthropic_skills.py b/litellm/types/llms/anthropic_skills.py
index 51eefe7154f..4b27f9b17ef 100644
--- a/litellm/types/llms/anthropic_skills.py
+++ b/litellm/types/llms/anthropic_skills.py
@@ -57,6 +57,15 @@ class Skill(BaseModel):
     updated_at: str
     """ISO 8601 timestamp of when the skill was last updated"""
 
+    description: str | None = None
+    """Description of the skill. Populated for the LiteLLM-hosted registry
+    (custom_llm_provider="litellm_proxy"); Anthropic's list endpoint does not
+    return a description, so this is None there."""
+
+    search_score: float | None = None
+    """Semantic similarity to the ``query`` passed to ``GET /v1/skills``. None
+    unless a query was given."""
+
 
 class ListSkillsResponse(BaseModel):
     """Response from listing skills"""
diff --git a/tests/test_litellm/llms/litellm_proxy/skills/test_skill_search.py b/tests/test_litellm/llms/litellm_proxy/skills/test_skill_search.py
new file mode 100644
index 00000000000..a0f22a59f0c
--- /dev/null
+++ b/tests/test_litellm/llms/litellm_proxy/skills/test_skill_search.py
@@ -0,0 +1,436 @@
+import asyncio
+import json
+from collections.abc import Sequence
+from types import MappingProxyType
+from typing import Final
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from openai import APIConnectionError
+
+import litellm
+from litellm.llms.litellm_proxy.skills.skill_search import (
+    MAX_SKILL_SEARCH_TEXT_CHARS,
+    SkillSearchEmbeddingFailed,
+    SkillSearchHits,
+    SkillSearchIndex,
+    SkillSearchNotConfigured,
+    search_skills,
+    skill_search_text,
+)
+from litellm.proxy._types import LiteLLM_SkillsTable, UserAPIKeyAuth
+from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+from litellm.proxy.common_utils.semantic_text_index import Vector, cosine_similarity
+
+CALLER: Final = UserAPIKeyAuth(api_key="hashed-caller-key", team_id="team-1", user_id="user-1")
+
+TRANSLATOR: Final = LiteLLM_SkillsTable(
+    skill_id="translate-file",
+    display_title="Document Translator",
+    description="Converts files from one language into another",
+    instructions="Take an uploaded document and produce it in the target language",
+)
+SQL_ANALYST: Final = LiteLLM_SkillsTable(
+    skill_id="warehouse-sql-analyst",
+    display_title="Warehouse SQL Analyst",
+    description="Runs SQL against the inventory database",
+)
+TRIP_PLANNER: Final = LiteLLM_SkillsTable(
+    skill_id="trip-planner",
+    display_title="Trip Planner",
+    description="Books flights and hotels",
+)
+SKILLS: Final = (TRANSLATOR, SQL_ANALYST, TRIP_PLANNER)
+
+VECTORS: Final = MappingProxyType(
+    {
+        "language translation": (1.0, 0.0, 0.0),
+        skill_search_text(TRANSLATOR): (0.9, 0.1, 0.0),
+        skill_search_text(SQL_ANALYST): (0.0, 1.0, 0.0),
+        skill_search_text(TRIP_PLANNER): (0.3, 0.0, 1.0),
+    }
+)
+
+
+def _pass_through_key_limits() -> MagicMock:
+    limits = MagicMock()
+    limits.pre_call_hook = AsyncMock(side_effect=lambda user_api_key_dict, data, call_type: data)
+    return limits
+
+
+def _embedding_router() -> MagicMock:
+    router = MagicMock()
+    router.aembedding = AsyncMock(
+        side_effect=lambda model, input, metadata: litellm.EmbeddingResponse(
+            model=model,
+            data=[{"object": "embedding", "index": i, "embedding": list(VECTORS[t])} for i, t in enumerate(input)],
+        )
+    )
+    return router
+
+
+class FakeEmbedder:
+    def __init__(self) -> None:
+        self.calls: list[tuple[str, ...]] = []  # mutable-ok: test spy recording embed inputs
+
+    async def __call__(self, texts: Sequence[str]) -> Sequence[Vector]:
+        self.calls.append(tuple(texts))
+        return tuple(VECTORS[text] for text in texts)
+
+
+class FixedDimensionEmbedder:
+    def __init__(self, dimensions: int) -> None:
+        self.dimensions: Final = dimensions
+        self.calls: list[tuple[str, ...]] = []  # mutable-ok: test spy recording embed inputs
+
+    async def __call__(self, texts: Sequence[str]) -> Sequence[Vector]:
+        self.calls.append(tuple(texts))
+        await asyncio.sleep(0)
+        return tuple((1.0,) * self.dimensions for _ in texts)
+
+
+class TestSkillSearchText:
+    def test_joins_title_description_and_instructions(self) -> None:
+        assert skill_search_text(TRANSLATOR) == (
+            "Document Translator\n"
+            "Converts files from one language into another\n"
+            "Take an uploaded document and produce it in the target language"
+        )
+
+    def test_missing_fields_fall_back_to_whatever_is_present(self) -> None:
+        assert skill_search_text(LiteLLM_SkillsTable(skill_id="bare", display_title="bare")) == "bare"
+
+    def test_all_fields_absent_is_an_empty_string(self) -> None:
+        assert skill_search_text(LiteLLM_SkillsTable(skill_id="empty")) == ""
+
+    def test_oversized_instructions_are_cut_so_one_skill_cannot_blow_up_the_embedding_batch(self) -> None:
+        bloated = LiteLLM_SkillsTable(
+            skill_id="bloated", display_title="Bloated", instructions="x" * (MAX_SKILL_SEARCH_TEXT_CHARS * 3)
+        )
+        text = skill_search_text(bloated)
+        assert len(text) == MAX_SKILL_SEARCH_TEXT_CHARS
+        assert text.startswith("Bloated\n")
+
+
+class TestCosineSimilarity:
+    def test_identical_direction_scores_one(self) -> None:
+        assert cosine_similarity((2.0, 0.0), (1.0, 0.0)) == pytest.approx(1.0)
+
+    def test_orthogonal_scores_zero(self) -> None:
+        assert cosine_similarity((1.0, 0.0), (0.0, 1.0)) == pytest.approx(0.0)
+
+    def test_zero_vector_scores_zero_instead_of_dividing(self) -> None:
+        assert cosine_similarity((0.0, 0.0), (1.0, 0.0)) == 0.0
+
+
+class TestSkillSearchIndex:
+    @pytest.mark.asyncio
+    async def test_ranks_by_similarity_and_truncates_to_top_k(self) -> None:
+        outcome = await SkillSearchIndex().search(
+            "language translation", SKILLS, top_k=2, embed=FakeEmbedder(), embedding_model="m"
+        )
+        assert isinstance(outcome, SkillSearchHits)
+        assert [hit.skill.skill_id for hit in outcome.hits] == ["translate-file", "trip-planner"]
+        assert outcome.hits[0].score > outcome.hits[1].score
+
+    @pytest.mark.asyncio
+    async def test_second_search_only_embeds_the_query(self) -> None:
+        index = SkillSearchIndex()
+        embedder = FakeEmbedder()
+        await index.search("language translation", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        await index.search("language translation", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        assert len(embedder.calls[0]) == 1 + len(SKILLS)
+        assert embedder.calls[1] == ("language translation",)
+
+    @pytest.mark.asyncio
+    async def test_switching_embedding_models_does_not_reuse_cached_vectors(self) -> None:
+        index = SkillSearchIndex()
+        await index.search("language translation", SKILLS, top_k=5, embed=FakeEmbedder(), embedding_model="small")
+        wide = FixedDimensionEmbedder(2)
+        outcome = await index.search("language translation", SKILLS, top_k=5, embed=wide, embedding_model="wide")
+        assert isinstance(outcome, SkillSearchHits)
+        assert len(wide.calls[0]) == 1 + len(SKILLS)
+
+    @pytest.mark.asyncio
+    async def test_cached_vectors_of_another_dimension_are_re_embedded(self) -> None:
+        index = SkillSearchIndex()
+        await index.search("language translation", SKILLS, top_k=5, embed=FakeEmbedder(), embedding_model="m")
+        fallback = FixedDimensionEmbedder(2)
+        outcome = await index.search("language translation", SKILLS, top_k=5, embed=fallback, embedding_model="m")
+        assert isinstance(outcome, SkillSearchHits)
+        assert fallback.calls == [
+            ("language translation",),
+            ("language translation", *(skill_search_text(skill) for skill in SKILLS)),
+        ]
+
+    @pytest.mark.asyncio
+    async def test_re_embedding_a_subset_drops_the_other_skills_old_vectors(self) -> None:
+        index = SkillSearchIndex()
+        await index.search("language translation", SKILLS, top_k=5, embed=FakeEmbedder(), embedding_model="m")
+        wide = FixedDimensionEmbedder(2)
+        await index.search("language translation", SKILLS[:1], top_k=5, embed=wide, embedding_model="m")
+        await index.search("language translation", SKILLS, top_k=5, embed=wide, embedding_model="m")
+        assert wide.calls[-1] == ("language translation", *(skill_search_text(skill) for skill in SKILLS[1:]))
+
+    @pytest.mark.asyncio
+    async def test_concurrent_searches_keep_each_others_vectors(self) -> None:
+        index = SkillSearchIndex()
+        embedder = FixedDimensionEmbedder(3)
+        await asyncio.gather(
+            index.search("q", SKILLS[:1], top_k=5, embed=embedder, embedding_model="m"),
+            index.search("q", SKILLS[1:], top_k=5, embed=embedder, embedding_model="m"),
+        )
+        await index.search("q", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        assert embedder.calls[-1] == ("q",)
+
+    @pytest.mark.asyncio
+    async def test_least_recently_searched_skills_are_evicted_once_the_index_is_full(self) -> None:
+        index = SkillSearchIndex(max_entries=len(SKILLS))
+        embedder = FixedDimensionEmbedder(3)
+        newcomer = LiteLLM_SkillsTable(skill_id="newcomer", display_title="Newcomer")
+        await index.search("q", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        await index.search("q", SKILLS[:1], top_k=5, embed=embedder, embedding_model="m")
+        await index.search("q", (newcomer,), top_k=5, embed=embedder, embedding_model="m")
+        await index.search("q", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        assert embedder.calls[-1] == ("q", skill_search_text(SKILLS[1]))
+
+    @pytest.mark.asyncio
+    async def test_deleted_skills_stop_occupying_the_index_after_enough_new_ones(self) -> None:
+        index = SkillSearchIndex(max_entries=2)
+        embedder = FixedDimensionEmbedder(3)
+        for generation in range(50):
+            skill = LiteLLM_SkillsTable(skill_id=f"gen-{generation}", display_title=f"Generation {generation}")
+            await index.search("q", (skill,), top_k=5, embed=embedder, embedding_model="m")
+        await index.search("q", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        await index.search("q", SKILLS, top_k=5, embed=embedder, embedding_model="m")
+        assert embedder.calls[-1] == ("q", skill_search_text(SKILLS[0]))
+
+    @pytest.mark.asyncio
+    async def test_mixed_dimensions_in_one_batch_become_embedding_failed(self) -> None:
+        async def mixed(texts: Sequence[str]) -> Sequence[Vector]:
+            return ((1.0, 0.0), *((1.0, 0.0, 0.0) for _ in texts[1:]))
+
+        outcome = await SkillSearchIndex().search("q", SKILLS, top_k=5, embed=mixed, embedding_model="m")
+        assert isinstance(outcome, SkillSearchEmbeddingFailed)
+        assert "mixed dimensions" in outcome.reason
+
+    @pytest.mark.asyncio
+    async def test_no_accessible_skills_returns_no_hits_without_embedding(self) -> None:
+        embedder = FakeEmbedder()
+        outcome = await SkillSearchIndex().search("anything", (), top_k=5, embed=embedder, embedding_model="m")
+        assert outcome == SkillSearchHits(hits=())
+        assert embedder.calls == []
+
+    @pytest.mark.asyncio
+    async def test_provider_error_becomes_embedding_failed(self) -> None:
+        async def failing(texts: Sequence[str]) -> Sequence[Vector]:
+            raise APIConnectionError(request=MagicMock())
+
+        outcome = await SkillSearchIndex().search("q", SKILLS, top_k=5, embed=failing, embedding_model="m")
+        assert isinstance(outcome, SkillSearchEmbeddingFailed)
+        assert "embedding the search query failed" in outcome.reason
+
+    @pytest.mark.asyncio
+    async def test_wrong_vector_count_becomes_embedding_failed(self) -> None:
+        async def short(texts: Sequence[str]) -> Sequence[Vector]:
+            return ((1.0, 0.0, 0.0),)
+
+        outcome = await SkillSearchIndex().search("q", SKILLS, top_k=5, embed=short, embedding_model="m")
+        assert isinstance(outcome, SkillSearchEmbeddingFailed)
+
+
+class TestSearchSkills:
+    @pytest.mark.asyncio
+    async def test_no_embedding_model_is_not_configured(self) -> None:
+        outcome = await search_skills(
+            "q",
+            SKILLS,
+            5,
+            router=MagicMock(),
+            embedding_model=None,
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
+        )
+        assert isinstance(outcome, SkillSearchNotConfigured)
+        assert "skill_search_embedding_model" in outcome.reason
+
+    @pytest.mark.asyncio
+    async def test_no_router_is_not_configured(self) -> None:
+        outcome = await search_skills(
+            "q",
+            SKILLS,
+            5,
+            router=None,
+            embedding_model="m",
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
+        )
+        assert isinstance(outcome, SkillSearchNotConfigured)
+
+    @pytest.mark.asyncio
+    async def test_router_embeddings_are_read_from_the_response(self) -> None:
+        router = _embedding_router()
+        outcome = await search_skills(
+            "language translation",
+            SKILLS,
+            1,
+            router=router,
+            embedding_model="text-embedding-3-small",
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
+        )
+        assert isinstance(outcome, SkillSearchHits)
+        assert [hit.skill.skill_id for hit in outcome.hits] == ["translate-file"]
+        assert router.aembedding.await_args.kwargs["model"] == "text-embedding-3-small"
+
+    @pytest.mark.asyncio
+    async def test_embedding_spend_is_attributed_to_the_calling_key(self) -> None:
+        router = _embedding_router()
+        await search_skills(
+            "language translation",
+            SKILLS,
+            1,
+            router=router,
+            embedding_model="text-embedding-3-small",
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
+        )
+        metadata = router.aembedding.await_args.kwargs["metadata"]
+        assert metadata["user_api_key"] == "hashed-caller-key"
+        assert metadata["user_api_key_team_id"] == "team-1"
+        assert metadata["user_api_key_user_id"] == "user-1"
+
+    @pytest.mark.asyncio
+    async def test_key_limits_are_checked_against_the_real_embedding_call_before_it_runs(self) -> None:
+        router = _embedding_router()
+        key_limits = _pass_through_key_limits()
+        await search_skills(
+            "language translation",
+            SKILLS,
+            1,
+            router=router,
+            embedding_model="text-embedding-3-small",
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=key_limits,
+        )
+        checked = key_limits.pre_call_hook.await_args.kwargs
+        assert checked["user_api_key_dict"] is CALLER
+        assert checked["call_type"] == "aembedding"
+        assert checked["data"]["model"] == "text-embedding-3-small"
+        assert checked["data"]["input"] == router.aembedding.await_args.kwargs["input"]
+        assert checked["data"]["metadata"]["user_api_key"] == "hashed-caller-key"
+
+    @pytest.mark.asyncio
+    async def test_the_embedding_model_sees_the_request_as_the_guardrails_rewrote_it(self) -> None:
+        router = MagicMock()
+        router.aembedding = AsyncMock(
+            side_effect=lambda model, input, metadata: litellm.EmbeddingResponse(
+                model=model,
+                data=[{"object": "embedding", "index": i, "embedding": [1.0, 0.0, 0.0]} for i in range(len(input))],
+            )
+        )
+        key_limits = MagicMock()
+        key_limits.pre_call_hook = AsyncMock(
+            side_effect=lambda user_api_key_dict, data, call_type: {
+                **data,
+                "input": ["[MASKED]" for _ in data["input"]],
+                "metadata": {**data["metadata"], "guardrail": "masked"},
+            }
+        )
+        await search_skills(
+            "language translation",
+            SKILLS,
+            1,
+            router=router,
+            embedding_model="text-embedding-3-small",
+            index=SkillSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=key_limits,
+        )
+        sent = tuple(call.kwargs for call in router.aembedding.await_args_list)
+        assert sent
+        assert all(set(call["input"]) == {"[MASKED]"} for call in sent)
+        assert all(call["metadata"]["guardrail"] == "masked" for call in sent)
+
+    @pytest.mark.asyncio
+    async def test_a_key_over_its_limit_never_reaches_the_embedding_model(self) -> None:
+        router = _embedding_router()
+        key_limits = MagicMock()
+        key_limits.pre_call_hook = AsyncMock(side_effect=ProxyRateLimitError(detail="rpm exceeded"))
+        with pytest.raises(ProxyRateLimitError) as raised:
+            await search_skills(
+                "language translation",
+                SKILLS,
+                1,
+                router=router,
+                embedding_model="text-embedding-3-small",
+                index=SkillSearchIndex(),
+                user_api_key_dict=CALLER,
+                proxy_logging_obj=key_limits,
+            )
+        assert raised.value.status_code == 429
+        router.aembedding.assert_not_awaited()
+
+
+@pytest.fixture
+def accessible_skills(monkeypatch: pytest.MonkeyPatch) -> AsyncMock:
+    list_for_search = AsyncMock(return_value=list(SKILLS))
+    monkeypatch.setattr(
+        "litellm.llms.litellm_proxy.skills.handler.LiteLLMSkillsHandler.list_skills_for_search", list_for_search
+    )
+    return list_for_search
+
+
+@pytest.fixture
+def key_limits(monkeypatch: pytest.MonkeyPatch) -> MagicMock:
+    limits = _pass_through_key_limits()
+    monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", limits)
+    return limits
+
+
+@pytest.fixture
+def embedding_router(monkeypatch: pytest.MonkeyPatch, key_limits: MagicMock) -> MagicMock:
+    router = _embedding_router()
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router)
+    monkeypatch.setattr(litellm, "skill_search_embedding_model", "text-embedding-3-small")
+    return router
+
+
+class TestHandleSkillSearchMCP:
+    @pytest.mark.asyncio
+    async def test_top_k_is_clamped_to_the_same_ceiling_as_rest(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        from litellm.llms.litellm_proxy.skills.skill_search import MAX_SKILL_SEARCH_TOP_K
+        from litellm.proxy._experimental.mcp_server.tool_search import handle_skill_search
+
+        many_skills: Final = tuple(
+            LiteLLM_SkillsTable(
+                skill_id=f"skill-{i}", display_title=TRIP_PLANNER.display_title, description=TRIP_PLANNER.description
+            )
+            for i in range(MAX_SKILL_SEARCH_TOP_K + 50)
+        )
+        accessible_skills.return_value = list(many_skills)
+
+        result = await handle_skill_search(
+            query="language translation", top_k=10_000, user_api_key_dict=UserAPIKeyAuth(user_id="u")
+        )
+        assert result.isError is False
+        assert len(json.loads(result.content[0].text)) == MAX_SKILL_SEARCH_TOP_K
+
+    @pytest.mark.asyncio
+    async def test_top_k_below_one_is_raised_to_one(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        from litellm.proxy._experimental.mcp_server.tool_search import handle_skill_search
+
+        result = await handle_skill_search(
+            query="language translation", top_k=0, user_api_key_dict=UserAPIKeyAuth(user_id="u")
+        )
+        assert result.isError is False
+        assert len(json.loads(result.content[0].text)) == 1
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_tool_search.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_tool_search.py
index 239f89ebd90..798b0001af1 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_tool_search.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_tool_search.py
@@ -24,6 +24,7 @@ from litellm.proxy._experimental.mcp_server.tool_search import (
     AGENT_SEARCH_TOOL_NAME,
     MCP_TOOL_CALL_TOOL_NAME,
     MCP_TOOL_SEARCH_TOOL_NAME,
+    SKILL_SEARCH_TOOL_NAME,
     SemanticToolRanker,
     ToolSearchResult,
     coerce_top_k,
@@ -272,8 +273,8 @@ class TestSearchTools:
 
 
 class TestGetVirtualToolDefinitions:
-    def test_returns_three_tools(self) -> None:
-        assert len(get_virtual_tool_definitions()) == 3
+    def test_returns_four_tools(self) -> None:
+        assert len(get_virtual_tool_definitions()) == 4
 
     def test_agent_search_schema_requires_query(self) -> None:
         tools = get_virtual_tool_definitions()
@@ -330,6 +331,7 @@ class TestGetVirtualToolDefinitions:
             MCP_TOOL_SEARCH_TOOL_NAME,
             MCP_TOOL_CALL_TOOL_NAME,
             AGENT_SEARCH_TOOL_NAME,
+            SKILL_SEARCH_TOOL_NAME,
         }
 
 
@@ -364,7 +366,12 @@ class TestListToolRestApiWithToolSearch:
 
         assert result["error"] is None
         tool_names = [t["name"] for t in result["tools"]]
-        assert set(tool_names) == {MCP_TOOL_SEARCH_TOOL_NAME, MCP_TOOL_CALL_TOOL_NAME, AGENT_SEARCH_TOOL_NAME}
+        assert set(tool_names) == {
+            MCP_TOOL_SEARCH_TOOL_NAME,
+            MCP_TOOL_CALL_TOOL_NAME,
+            AGENT_SEARCH_TOOL_NAME,
+            SKILL_SEARCH_TOOL_NAME,
+        }
 
     @pytest.mark.asyncio
     async def test_returns_full_catalog_when_flag_disabled(self) -> None:
@@ -737,6 +744,31 @@ class TestCallToolRestApiVirtualTools:
         assert mock_search.await_args.kwargs["top_k"] == 1
         assert mock_search.await_args.kwargs["agents"] == (translator,)
 
+    @pytest.mark.asyncio
+    async def test_skill_search_call_tolerates_malformed_top_k(self) -> None:
+        """Regression: a caller-supplied non-numeric top_k must be coerced to the default,
+        the same as agent_search, instead of raising a pydantic ValidationError that the
+        endpoint's catch-all turns into an HTTP 500."""
+        from mcp.types import CallToolResult, TextContent
+
+        from litellm.llms.litellm_proxy.skills.skill_search import DEFAULT_SKILL_SEARCH_TOP_K
+
+        user_api_key_dict = UserAPIKeyAuth(api_key="k", object_permission=_make_perm(mcp_tool_search_enabled=True))
+        request = self._make_request(
+            {"name": SKILL_SEARCH_TOOL_NAME, "arguments": {"query": "translate a document", "top_k": "not-a-number"}}
+        )
+        fake_result = CallToolResult(content=[TextContent(type="text", text="[]")], isError=False)
+        with patch(  # test-quality-ok: the embedding router only resolves via proxy_server globals, no injection seam
+            "litellm.proxy._experimental.mcp_server.tool_search.handle_skill_search",
+            new_callable=AsyncMock,
+            return_value=fake_result,
+        ) as mock_search:
+            result = await self._get_call_fn()(request=request, user_api_key_dict=user_api_key_dict)
+
+        assert result.isError is False
+        assert mock_search.await_args.kwargs["top_k"] == DEFAULT_SKILL_SEARCH_TOP_K
+        assert mock_search.await_args.kwargs["query"] == "translate a document"
+
     @pytest.mark.asyncio
     async def test_agent_search_call_reports_missing_embedding_model_as_tool_error(self) -> None:
         from litellm.proxy.agent_endpoints.agent_search import AgentSearchNotConfigured
@@ -782,10 +814,15 @@ class TestCallToolRestApiVirtualTools:
 
         router = MagicMock()
         router.aembedding = AsyncMock(side_effect=fake_aembedding)
+        key_limits = MagicMock()
+        key_limits.pre_call_hook = AsyncMock(side_effect=lambda user_api_key_dict, data, call_type: data)
         with (
             patch(  # test-quality-ok: the proxy's router is a module global; the handler reaches it the way production does
                 "litellm.proxy.proxy_server.llm_router", router
             ),
+            patch(  # test-quality-ok: the proxy's key-limit hooks are a module global; the embedding call runs them like /embeddings does
+                "litellm.proxy.proxy_server.proxy_logging_obj", key_limits
+            ),
             patch(  # test-quality-ok: the authorized catalog is the seam every virtual tool shares; the ranking under test stays real
                 "litellm.proxy._experimental.mcp_server.server._list_mcp_tools",
                 new_callable=AsyncMock,
@@ -795,6 +832,8 @@ class TestCallToolRestApiVirtualTools:
             result = await self._get_call_fn()(request=self._semantic_request(), user_api_key_dict=user_api_key_dict)
 
         assert mock_list.await_args.kwargs["user_api_key_auth"] is user_api_key_dict
+        assert key_limits.pre_call_hook.await_args.kwargs["call_type"] == "aembedding"
+        assert key_limits.pre_call_hook.await_args.kwargs["data"]["model"] == "emb"
         assert result.isError is False
         assert [t["name"] for t in json.loads(result.content[0].text)] == [FX_TOOL.name]
 
@@ -810,7 +849,9 @@ class TestCallToolRestApiVirtualTools:
         assert "mcp_tool_search.embedding_model" in result.content[0].text
 
     @pytest.mark.asyncio
-    async def test_mcp_tool_search_reports_invalid_settings_as_tool_error(self, monkeypatch: pytest.MonkeyPatch) -> None:
+    async def test_mcp_tool_search_reports_invalid_settings_as_tool_error(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
         monkeypatch.setattr(litellm, "mcp_tool_search", {"top_k": 0})
         user_api_key_dict = UserAPIKeyAuth(api_key="k", object_permission=_make_perm(mcp_tool_search_enabled=True))
         result = await self._get_call_fn()(request=self._semantic_request(), user_api_key_dict=user_api_key_dict)
@@ -1196,6 +1237,7 @@ class TestHandleListToolsVirtual:
             MCP_TOOL_SEARCH_TOOL_NAME,
             MCP_TOOL_CALL_TOOL_NAME,
             AGENT_SEARCH_TOOL_NAME,
+            SKILL_SEARCH_TOOL_NAME,
         }
 
 
diff --git a/tests/test_litellm/proxy/agent_endpoints/test_agent_search.py b/tests/test_litellm/proxy/agent_endpoints/test_agent_search.py
index daca244c0a1..c02ed1f37e5 100644
--- a/tests/test_litellm/proxy/agent_endpoints/test_agent_search.py
+++ b/tests/test_litellm/proxy/agent_endpoints/test_agent_search.py
@@ -211,11 +211,24 @@ class TestAgentSearchIndex:
         assert isinstance(outcome, AgentSearchEmbeddingFailed)
 
 
+def _pass_through_key_limits() -> MagicMock:
+    limits = MagicMock()
+    limits.pre_call_hook = AsyncMock(side_effect=lambda user_api_key_dict, data, call_type: data)
+    return limits
+
+
 class TestSearchAgents:
     @pytest.mark.asyncio
     async def test_no_embedding_model_is_not_configured(self) -> None:
         outcome = await search_agents(
-            "q", AGENTS, 5, router=MagicMock(), embedding_model=None, index=AgentSearchIndex(), user_api_key_dict=CALLER
+            "q",
+            AGENTS,
+            5,
+            router=MagicMock(),
+            embedding_model=None,
+            index=AgentSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
         )
         assert isinstance(outcome, AgentSearchNotConfigured)
         assert "agent_search_embedding_model" in outcome.reason
@@ -223,7 +236,14 @@ class TestSearchAgents:
     @pytest.mark.asyncio
     async def test_no_router_is_not_configured(self) -> None:
         outcome = await search_agents(
-            "q", AGENTS, 5, router=None, embedding_model="m", index=AgentSearchIndex(), user_api_key_dict=CALLER
+            "q",
+            AGENTS,
+            5,
+            router=None,
+            embedding_model="m",
+            index=AgentSearchIndex(),
+            user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
         )
         assert isinstance(outcome, AgentSearchNotConfigured)
 
@@ -244,6 +264,7 @@ class TestSearchAgents:
             embedding_model="text-embedding-3-small",
             index=AgentSearchIndex(),
             user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
         )
         assert isinstance(outcome, AgentSearchHits)
         assert [hit.agent.agent_id for hit in outcome.hits] == ["translator"]
@@ -266,6 +287,7 @@ class TestSearchAgents:
             embedding_model="text-embedding-3-small",
             index=AgentSearchIndex(),
             user_api_key_dict=CALLER,
+            proxy_logging_obj=_pass_through_key_limits(),
         )
         metadata = router.aembedding.await_args.kwargs["metadata"]
         assert metadata["user_api_key"] == "hashed-caller-key"
@@ -302,6 +324,7 @@ def embedding_router(monkeypatch: pytest.MonkeyPatch) -> MagicMock:
         )
     )
     monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router)
+    monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", _pass_through_key_limits())
     monkeypatch.setattr(litellm, "agent_search_embedding_model", "text-embedding-3-small")
     return router
 
diff --git a/tests/test_litellm/proxy/anthropic_endpoints/test_skills_endpoints.py b/tests/test_litellm/proxy/anthropic_endpoints/test_skills_endpoints.py
new file mode 100644
index 00000000000..ae6b1471c93
--- /dev/null
+++ b/tests/test_litellm/proxy/anthropic_endpoints/test_skills_endpoints.py
@@ -0,0 +1,175 @@
+from types import MappingProxyType
+from typing import Final
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from openai import APIConnectionError
+
+import litellm
+from litellm.llms.litellm_proxy.skills.skill_search import skill_search_text
+from litellm.proxy._types import LiteLLM_SkillsTable, LitellmUserRoles, UserAPIKeyAuth
+from litellm.proxy.anthropic_endpoints.skills_endpoints import router
+from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
+from litellm.proxy.common_utils.proxy_rate_limit_error import ProxyRateLimitError
+
+TRANSLATOR: Final = LiteLLM_SkillsTable(
+    skill_id="translate-file",
+    display_title="Document Translator",
+    description="Converts files from one language into another",
+    instructions="Take an uploaded document and produce it in the target language",
+)
+SQL_ANALYST: Final = LiteLLM_SkillsTable(
+    skill_id="warehouse-sql-analyst",
+    display_title="Warehouse SQL Analyst",
+    description="Runs SQL against the inventory database",
+)
+TRIP_PLANNER: Final = LiteLLM_SkillsTable(
+    skill_id="trip-planner",
+    display_title="Trip Planner",
+    description="Books flights and hotels",
+)
+SKILLS: Final = (TRANSLATOR, SQL_ANALYST, TRIP_PLANNER)
+
+VECTORS: Final = MappingProxyType(
+    {
+        "language translation": (1.0, 0.0, 0.0),
+        skill_search_text(TRANSLATOR): (0.9, 0.1, 0.0),
+        skill_search_text(SQL_ANALYST): (0.0, 1.0, 0.0),
+        skill_search_text(TRIP_PLANNER): (0.3, 0.0, 1.0),
+    }
+)
+
+
+def _client(role: LitellmUserRoles) -> TestClient:
+    app = FastAPI()
+    app.include_router(router)
+    app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth(user_id="u", user_role=role)
+    return TestClient(app)
+
+
+@pytest.fixture
+def accessible_skills(monkeypatch: pytest.MonkeyPatch) -> AsyncMock:
+    list_for_search = AsyncMock(return_value=list(SKILLS))
+    monkeypatch.setattr(
+        "litellm.llms.litellm_proxy.skills.handler.LiteLLMSkillsHandler.list_skills_for_search", list_for_search
+    )
+    return list_for_search
+
+
+@pytest.fixture
+def key_limits(monkeypatch: pytest.MonkeyPatch) -> MagicMock:
+    limits = MagicMock()
+    limits.pre_call_hook = AsyncMock(side_effect=lambda user_api_key_dict, data, call_type: data)
+    monkeypatch.setattr("litellm.proxy.proxy_server.proxy_logging_obj", limits)
+    return limits
+
+
+@pytest.fixture
+def embedding_router(monkeypatch: pytest.MonkeyPatch, key_limits: MagicMock) -> MagicMock:
+    embedding_router = MagicMock()
+    embedding_router.aembedding = AsyncMock(
+        side_effect=lambda model, input, metadata: litellm.EmbeddingResponse(
+            model=model,
+            data=[{"object": "embedding", "index": i, "embedding": list(VECTORS[t])} for i, t in enumerate(input)],
+        )
+    )
+    monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", embedding_router)
+    monkeypatch.setattr(litellm, "skill_search_embedding_model", "text-embedding-3-small")
+    return embedding_router
+
+
+class TestGetSkillsQuery:
+    def test_query_ranks_and_scores_and_truncates(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "language translation", "top_k": 2},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 200
+        body = response.json()["data"]
+        assert [skill["id"] for skill in body] == ["translate-file", "trip-planner"]
+        assert body[0]["search_score"] > body[1]["search_score"]
+        assert embedding_router.aembedding.await_args.kwargs["metadata"]["user_api_key_user_id"] == "u"
+
+    def test_restricted_key_only_ranks_the_skills_it_can_access(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        accessible_skills.return_value = [SQL_ANALYST]
+        response = _client(LitellmUserRoles.INTERNAL_USER).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "language translation"},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 200
+        assert [skill["id"] for skill in response.json()["data"]] == ["warehouse-sql-analyst"]
+
+    def test_no_accessible_skills_is_a_no_match_empty_result(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        accessible_skills.return_value = []
+        response = _client(LitellmUserRoles.INTERNAL_USER).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "anything"},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 200
+        assert response.json()["data"] == []
+        embedding_router.aembedding.assert_not_awaited()
+
+    def test_query_is_unsupported_for_the_anthropic_passthrough_provider(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills", params={"query": "anything"}, headers={"Authorization": "Bearer k"}
+        )
+        assert response.status_code == 400
+        assert response.json()["detail"]["error"] == "skill_search_unsupported_provider"
+        accessible_skills.assert_not_awaited()
+
+    def test_missing_embedding_model_is_a_400(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        monkeypatch.setattr(litellm, "skill_search_embedding_model", None)
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "anything"},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 400
+        assert response.json()["detail"]["error"] == "skill_search_not_configured"
+
+    def test_embedding_provider_failure_is_a_503(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock
+    ) -> None:
+        embedding_router.aembedding = AsyncMock(side_effect=APIConnectionError(request=MagicMock()))
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "anything"},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 503
+        assert response.json()["detail"]["error"] == "skill_search_unavailable"
+
+    def test_a_key_over_its_rate_limit_gets_a_429_without_embedding(
+        self, accessible_skills: AsyncMock, embedding_router: MagicMock, key_limits: MagicMock
+    ) -> None:
+        key_limits.pre_call_hook = AsyncMock(side_effect=ProxyRateLimitError(detail="rpm exceeded"))
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "language translation"},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 429
+        embedding_router.aembedding.assert_not_awaited()
+
+    def test_top_k_is_validated(self, accessible_skills: AsyncMock, embedding_router: MagicMock) -> None:
+        response = _client(LitellmUserRoles.PROXY_ADMIN).get(
+            "/v1/skills",
+            params={"custom_llm_provider": "litellm_proxy", "query": "anything", "top_k": 0},
+            headers={"Authorization": "Bearer k"},
+        )
+        assert response.status_code == 422
diff --git a/tests/test_litellm/skills/test_skills_main.py b/tests/test_litellm/skills/test_skills_main.py
new file mode 100644
index 00000000000..e1c66c8d9ea
--- /dev/null
+++ b/tests/test_litellm/skills/test_skills_main.py
@@ -0,0 +1,57 @@
+from unittest.mock import MagicMock
+
+import litellm.skills.main as skills_main
+from litellm.types.utils import LlmProviders
+
+
+def test_create_skill_forwards_description_and_instructions_from_top_level_kwargs(
+    monkeypatch,
+) -> None:
+    """The REST /v1/skills form endpoint passes description/instructions as top-level
+    kwargs (not extra_body). Regression for a bug where the litellm_proxy dispatch
+    branch of create_skill() dropped both, so every LiteLLM-hosted skill was created
+    with description=None and instructions=None regardless of what the caller sent."""
+    handler = MagicMock()
+    monkeypatch.setattr(skills_main, "_get_litellm_skills_handler", lambda: handler)
+
+    skills_main.create_skill(
+        display_title="Document Translator",
+        description="Converts files from one language into another",
+        instructions="Take an uploaded document and produce it in the target language",
+        custom_llm_provider=LlmProviders.LITELLM_PROXY.value,
+    )
+
+    assert handler.create_skill_handler.call_args.kwargs["description"] == (
+        "Converts files from one language into another"
+    )
+    assert handler.create_skill_handler.call_args.kwargs["instructions"] == (
+        "Take an uploaded document and produce it in the target language"
+    )
+
+
+def test_create_skill_forwards_description_and_instructions_from_extra_body(monkeypatch) -> None:
+    """The SDK convention (see tests/proxy_unit_tests/test_skills_db.py) nests them under
+    extra_body instead of passing them as top-level kwargs; both paths must reach the DB."""
+    handler = MagicMock()
+    monkeypatch.setattr(skills_main, "_get_litellm_skills_handler", lambda: handler)
+
+    skills_main.create_skill(
+        display_title="Warehouse SQL Analyst",
+        extra_body={"description": "Runs SQL against the inventory database", "instructions": "Summarize results"},
+        custom_llm_provider=LlmProviders.LITELLM_PROXY.value,
+    )
+
+    assert handler.create_skill_handler.call_args.kwargs["description"] == (
+        "Runs SQL against the inventory database"
+    )
+    assert handler.create_skill_handler.call_args.kwargs["instructions"] == "Summarize results"
+
+
+def test_create_skill_without_description_or_instructions_passes_none(monkeypatch) -> None:
+    handler = MagicMock()
+    monkeypatch.setattr(skills_main, "_get_litellm_skills_handler", lambda: handler)
+
+    skills_main.create_skill(display_title="Bare Skill", custom_llm_provider=LlmProviders.LITELLM_PROXY.value)
+
+    assert handler.create_skill_handler.call_args.kwargs["description"] is None
+    assert handler.create_skill_handler.call_args.kwargs["instructions"] is None
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 842a3da4122..12e7eb801e6 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -19872,6 +19872,12 @@ export interface paths {
          *     curl "http://localhost:4000/v1/skills?beta=true&limit=10"       -H "Authorization: Bearer your-key"       -H "x-litellm-model: claude-account-1"
          *     ```
          *
+         *     Pass `?custom_llm_provider=litellm_proxy&query=<task>` to rank the LiteLLM-hosted skills you can
+         *     access by semantic similarity instead of paging through the whole registry:
+         *     ```bash
+         *     curl "http://localhost:4000/v1/skills?custom_llm_provider=litellm_proxy&query=summarize+a+pdf&top_k=5"       -H "Authorization: Bearer your-key"
+         *     ```
+         *
          *     Returns: ListSkillsResponse with list of skills
          */
         get: operations["list_skills_v1_skills_get"];
@@ -36050,12 +36056,16 @@ export interface components {
         Skill: {
             /** Created At */
             created_at: string;
+            /** Description */
+            description?: string | null;
             /** Display Title */
             display_title?: string | null;
             /** Id */
             id: string;
             /** Latest Version */
             latest_version?: string | null;
+            /** Search Score */
+            search_score?: number | null;
             /** Source */
             source: string;
             /**
@@ -64556,6 +64566,10 @@ export interface operations {
                 after_id?: string | null;
                 before_id?: string | null;
                 custom_llm_provider?: string | null;
+                /** @description Describe what you need in natural language to rank the skills you can access by semantic similarity over their title and description. Each result carries a search_score. Only supported for custom_llm_provider=litellm_proxy. Requires litellm_settings.skill_search_embedding_model. */
+                query?: string | null;
+                /** @description With query: the maximum number of ranked skills to return. */
+                top_k?: number;
             };
             header?: never;
             path?: never;