From feeda1a36ce9f4155e443bbb264950ac0139f7a0 Mon Sep 17 00:00:00 2001
From: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com>
Date: Wed, 9 Sep 2026 13:32:06 -0700
Subject: [PATCH 1/4] fix(mcp): report resolved upstream authentication in
 debug headers

---
 .../_experimental/mcp_server/mcp_debug.py     | 161 ++++++++++++------
 .../mcp_server/mcp_server_manager.py          |  34 +++-
 .../outbound_credentials/resolver.py          |  31 ++++
 .../mcp_server/outbound_credentials/types.py  |  26 ++-
 .../proxy/_experimental/mcp_server/server.py  |  14 +-
 .../outbound_credentials/test_resolver.py     |  66 +++++++
 .../mcp_server/test_mcp_debug.py              | 161 ++++++++++--------
 .../mcp_server/test_mcp_server_manager.py     | 106 ++++++++++++
 8 files changed, 470 insertions(+), 129 deletions(-)

diff --git a/litellm/proxy/_experimental/mcp_server/mcp_debug.py b/litellm/proxy/_experimental/mcp_server/mcp_debug.py
index 6aaa00ae415..b25ed27becc 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_debug.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_debug.py
@@ -22,7 +22,16 @@ Response headers returned (all values are masked for safety):
     x-mcp-debug-auth-resolution
         Which auth priority was used for the outbound MCP call:
         ``per-request-header``, ``m2m-client-credentials``, ``static-token``,
-        ``oauth2-passthrough``, or ``no-auth``.
+        ``oauth2-passthrough``, ``stored-user-token``, ``token-exchange``,
+        ``id-jag``, ``aws-sigv4``, ``extra-headers``, or ``no-auth``.
+        ``unresolved`` means no outcome was available before the first response
+        frame; ``multiple`` means several servers resolved credentials;
+        ``not-applicable`` covers stdio; ``resolution-failed`` is a resolver error.
+
+    x-mcp-debug-auth-resolutions
+        For multiple servers, a JSON map of server IDs to resolution labels.
+        At most 32 entries are included; x-mcp-debug-auth-resolutions-truncated
+        is true when additional servers were omitted. No credentials are included.
 
     x-mcp-debug-outbound-url
         The upstream MCP server URL that will receive the request.
@@ -58,10 +67,16 @@ header is free for OAuth2 discovery::
 Symptom: ``x-mcp-debug-oauth2-token`` shows ``(none)`` and
 ``x-mcp-debug-auth-resolution`` shows ``no-auth``.
 
-This means the client didn't go through the OAuth2 flow. Check that:
-1. The ``Authorization`` header is NOT set as a static header in the client config.
-2. The ``.well-known/oauth-protected-resource`` endpoint returns valid metadata.
-3. The MCP server in LiteLLM config has ``auth_type: oauth2``.
+``no-auth`` means the resolved upstream client carries no authentication.
+An absent inbound OAuth2 token does not imply the user skipped OAuth: the gateway
+can retrieve a stored per-user token, reported as ``stored-user-token``.
+``unresolved`` is used when a stream starts before credential resolution, or a
+request (such as initialization or a cached tool listing) resolves no credential.
+Debug reporting does not fetch credentials or delay a streaming frame to resolve them.
+``extra-headers`` identifies supplied headers that won over the resolver or were
+the only headers supplied; their values are never inspected to guess a scheme.
+``per-request-header`` denotes a legacy credential override, including a BYOK
+credential supplied by the gateway; it does not imply a caller-supplied token.
 
 **Common issue: M2M token used instead of user token**
 
@@ -69,8 +84,8 @@ Symptom: ``x-mcp-debug-auth-resolution`` shows ``m2m-client-credentials``.
 
 This means the server has ``client_id``/``client_secret``/``token_url``
 configured and LiteLLM is fetching a machine-to-machine token instead of
-using the per-user OAuth2 token. If you want per-user tokens, remove the
-client credentials from the server config.
+using the per-user OAuth2 token. For gateway-stored per-user tokens,
+configure ``oauth2_flow: authorization_code``.
 
 Usage from Claude Code::
 
@@ -85,14 +100,16 @@ Usage with curl::
          http://localhost:4000/mcp/atlassian_mcp
 """
 
-from typing import TYPE_CHECKING, Final
+import json
+from collections.abc import Callable, Mapping
+from types import MappingProxyType
+from typing import Final
 
+from starlette.requests import HTTPConnection
 from starlette.types import Message, Send
 
 from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker
-
-if TYPE_CHECKING:
-    from litellm.types.mcp_server.mcp_server_manager import MCPServer
+from litellm.proxy._experimental.mcp_server.outbound_credentials.types import AuthResolution
 
 # Header the client sends to opt into debug mode
 MCP_DEBUG_REQUEST_HEADER: Final = "x-litellm-mcp-debug"
@@ -101,6 +118,83 @@ MCP_DEBUG_REQUEST_HEADER: Final = "x-litellm-mcp-debug"
 _RESPONSE_HEADER_PREFIX: Final = "x-mcp-debug"
 
 
+MCP_AUTH_DIAGNOSTICS_SCOPE_KEY: Final = "litellm.mcp.auth_diagnostics"
+
+
+def record_auth_resolution(server_id: str, source: AuthResolution) -> None:
+    from mcp.server.lowlevel.server import request_ctx
+
+    context: Final[object] = request_ctx.get(None)
+    request: Final[object] = getattr(context, "request", None)
+    if isinstance(request, HTTPConnection):
+        diagnostics: Final[object] = request.scope.get(MCP_AUTH_DIAGNOSTICS_SCOPE_KEY)
+        if isinstance(diagnostics, MCPAuthDiagnostics):
+            diagnostics.record(server_id, source)
+
+
+class MCPAuthDiagnostics:
+    def __init__(self) -> None:
+        self._outcomes: tuple[tuple[str, AuthResolution], ...] = ()
+
+    def record(self, server_id: str, resolution: AuthResolution) -> None:
+        self._outcomes = tuple(item for item in self._outcomes if item[0] != server_id) + ((server_id, resolution),)
+
+    def resolution(self) -> str:
+        match self._outcomes:
+            case ():
+                return AuthResolution.unresolved.value
+            case ((_, source),):
+                return source.value
+            case _:
+                return AuthResolution.multiple.value
+
+    def headers(self) -> Mapping[str, str]:
+        if len(self._outcomes) <= 1:
+            return MappingProxyType({"x-mcp-debug-auth-resolution": self.resolution()})
+        return MappingProxyType(
+            {
+                "x-mcp-debug-auth-resolution": AuthResolution.multiple.value,
+                "x-mcp-debug-auth-resolutions": json.dumps(
+                    {
+                        server_id: source.value for server_id, source in self._outcomes[:32]
+                    },  # mutable-ok: JSON encoder requires a concrete dict
+                    separators=(",", ":"),
+                    ensure_ascii=True,
+                ),
+                **(
+                    MappingProxyType({"x-mcp-debug-auth-resolutions-truncated": "true"})
+                    if len(self._outcomes) > 32
+                    else MappingProxyType({})
+                ),
+            }
+        )
+
+
+class _DiagnosticSend:
+    def __init__(self, send: Send, headers: Mapping[str, str], resolution: Callable[[], Mapping[str, str]]) -> None:
+        self._send = send
+        self._headers = headers
+        self._resolution = resolution
+        self._start: Message | None = None
+
+    async def __call__(self, message: Message) -> None:
+        if message["type"] == "http.response.start":
+            self._start = message
+            return
+        if self._start is not None:
+            start: Final = self._start
+            self._start = None
+            headers: Final = MappingProxyType({**self._headers, **self._resolution()})
+            await self._send(
+                {  # mutable-ok: ASGI send consumes a mutable message mapping
+                    **start,
+                    "headers": tuple(start.get("headers", ()))
+                    + tuple((key.encode(), value.encode()) for key, value in headers.items()),
+                }
+            )
+        await self._send(message)
+
+
 class MCPDebug:
     """
     Static helper class for MCP OAuth2 debug headers.
@@ -144,37 +238,6 @@ class MCPDebug:
                 return val.strip().lower() in ("true", "1", "yes")
         return False
 
-    @staticmethod
-    def resolve_auth_resolution(
-        server: "MCPServer",
-        mcp_auth_header: str | None,
-        mcp_server_auth_headers: dict[str, dict[str, str]] | None,
-        oauth2_headers: dict[str, str] | None,
-    ) -> str:
-        """
-        Determine which auth priority will be used for the outbound MCP call.
-
-        Returns one of: ``per-request-header``, ``m2m-client-credentials``,
-        ``static-token``, ``oauth2-passthrough``, or ``no-auth``.
-        """
-        from litellm.types.mcp import MCPAuth
-
-        has_server_specific: Final = bool(
-            mcp_server_auth_headers
-            and (
-                mcp_server_auth_headers.get(server.alias or "") or mcp_server_auth_headers.get(server.server_name or "")
-            )
-        )
-        if has_server_specific or mcp_auth_header:
-            return "per-request-header"
-        if server.has_client_credentials:
-            return "m2m-client-credentials"
-        if server.authentication_token:
-            return "static-token"
-        if oauth2_headers and server.auth_type == MCPAuth.oauth2:
-            return "oauth2-passthrough"
-        return "no-auth"
-
     @staticmethod
     def build_debug_headers(
         *,
@@ -244,12 +307,17 @@ class MCPDebug:
         return debug
 
     @staticmethod
-    def wrap_send_with_debug_headers(send: Send, debug_headers: dict[str, str]) -> Send:
+    def wrap_send_with_debug_headers(
+        send: Send, debug_headers: Mapping[str, str], resolution: Callable[[], Mapping[str, str]] | None = None
+    ) -> Send:
         """
         Return a new ASGI ``send`` callable that injects *debug_headers*
         into the ``http.response.start`` message.
         """
 
+        if resolution is not None:
+            return _DiagnosticSend(send, debug_headers, resolution)
+
         async def _send_with_debug(message: Message) -> None:
             if message["type"] == "http.response.start":
                 headers: Final = list(message.get("headers", []))
@@ -266,8 +334,6 @@ class MCPDebug:
         raw_headers: dict[str, str] | None,
         scope: dict,
         mcp_servers: list[str] | None,
-        mcp_auth_header: str | None,
-        mcp_server_auth_headers: dict[str, dict[str, str]] | None,
         oauth2_headers: dict[str, str] | None,
         client_ip: str | None,
     ) -> dict[str, str]:
@@ -288,16 +354,13 @@ class MCPDebug:
 
         server_url: str | None = None
         server_auth_type: str | None = None
-        auth_resolution = "no-auth"
+        auth_resolution: Final = AuthResolution.unresolved.value
 
         for server_name in mcp_servers or []:
             server = global_mcp_server_manager.get_mcp_server_by_name(server_name, client_ip=client_ip)
             if server:
                 server_url = server.url
                 server_auth_type = server.auth_type
-                auth_resolution = MCPDebug.resolve_auth_resolution(
-                    server, mcp_auth_header, mcp_server_auth_headers, oauth2_headers
-                )
                 break
 
         scope_headers: Final = MCPRequestHandler._safe_get_headers_from_scope(scope)
diff --git a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
index d7f238f142c..116bcc715d4 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_server_manager.py
@@ -80,6 +80,7 @@ from litellm.proxy._experimental.mcp_server.faults.list_outcomes import (
     raise_classified_list_failure,
     upstream_auth_challenge,
 )
+from litellm.proxy._experimental.mcp_server.mcp_debug import record_auth_resolution
 from litellm.proxy._experimental.mcp_server.oauth2_token_cache import (
     MCPPerUserTokenCache,
     mcp_per_user_token_cache,
@@ -108,12 +109,14 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.oauth_token_sto
 from litellm.proxy._experimental.mcp_server.outbound_credentials.per_user_oauth_store import (
     LazyPerUserOAuthTokenStore,
 )
+from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
 from litellm.proxy._experimental.mcp_server.outbound_credentials.token_exchange_provider import (
     build_token_exchanger,
 )
 from litellm.proxy._experimental.mcp_server.outbound_credentials.types import (
     DEFAULT_CREDENTIAL_HEADER,
     AuthorizationCodeConfig,
+    AuthResolution,
     ClientCredentialsConfig,
     CredError,
     IdJagConfig,
@@ -3832,13 +3835,21 @@ class MCPServerManager:
         (authorization_code's browser-OAuth 401, token_exchange's RFC 9728 challenge) or maps any
         other ``CredError`` onto its public HTTP status; it never returns an error as a value.
         """
-        match await provider.resolve_credentials(to_subject(user_api_key_auth, subject_token), spec):
-            case Ok(auth):
+        match await resolve_credentials_with_source(provider, to_subject(user_api_key_auth, subject_token), spec):
+            case Ok(credential):
+                auth: Final = credential.auth
                 # NoOpAuth has no header_name and so never conflicts.
                 header_name: Final[str | None] = getattr(auth, "header_name", None)
                 if header_name is None or not extra_headers:
+                    source: Final = (
+                        AuthResolution.extra_headers
+                        if credential.source == AuthResolution.no_auth and extra_headers
+                        else credential.source
+                    )
+                    record_auth_resolution(server.server_id, source)
                     return auth, extra_headers
                 if not has_header(extra_headers, header_name):
+                    record_auth_resolution(server.server_id, credential.source)
                     return auth, extra_headers
                 if isinstance(
                     spec.config,
@@ -3853,11 +3864,14 @@ class MCPServerManager:
                     # one-shot 401 refetch is lost with it). Drop only the header the resolved
                     # credential is about to occupy, so a static credential the operator aimed at a
                     # DIFFERENT header still reaches upstream.
+                    record_auth_resolution(server.server_id, credential.source)
                     return auth, without_header(extra_headers, header_name)
                 # Other modes: an Authorization already supplied via extra_headers (a forwarded caller
                 # header or static_headers) is intentional and wins; v1 applies those last.
+                record_auth_resolution(server.server_id, AuthResolution.extra_headers)
                 return None, extra_headers
             case Error(err):
+                record_auth_resolution(server.server_id, AuthResolution.failed)
                 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.
@@ -3960,6 +3974,7 @@ class MCPServerManager:
         Returns:
             Configured MCP client instance.
         """
+        record_auth_resolution(server.server_id, AuthResolution.unresolved)
         resolved_server: Final = await self.ensure_oauth_metadata_discovered(server)
         transport: Final = resolved_server.transport or MCPTransport.sse
         spec = None if transport == MCPTransport.stdio else _to_server_spec_fail_closed(resolved_server)
@@ -4032,6 +4047,7 @@ class MCPServerManager:
                     env=resolved_env,
                 )
 
+            record_auth_resolution(server.server_id, AuthResolution.not_applicable)
             return MCPClient(
                 server_url="",  # Not used for stdio
                 transport_type=transport,
@@ -4086,6 +4102,20 @@ class MCPServerManager:
                     aws_session_name=resolved_server.aws_session_name,
                 )
 
+            legacy_source: Final = (
+                AuthResolution.aws_sigv4
+                if aws_auth is not None
+                else AuthResolution.extra_headers
+                if extra_headers and has_header(extra_headers, auth_header_name or "Authorization")
+                else AuthResolution.per_request_header
+                if mcp_auth_header
+                else AuthResolution.static_token
+                if auth_value
+                else AuthResolution.extra_headers
+                if extra_headers
+                else AuthResolution.no_auth
+            )
+            record_auth_resolution(server.server_id, legacy_source)
             return MCPClient(
                 server_url=server_url,
                 transport_type=transport,
diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py
index 8328aae01ab..85c7f68719d 100644
--- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py
+++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/resolver.py
@@ -65,6 +65,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.token_exchanger
 from litellm.proxy._experimental.mcp_server.outbound_credentials.types import (
     ApiKeyConfig,
     AuthorizationCodeConfig,
+    AuthResolution,
     AuthSpecKind,
     AwsSigV4Config,
     Byok,
@@ -76,6 +77,7 @@ from litellm.proxy._experimental.mcp_server.outbound_credentials.types import (
     NoneConfig,
     PassthroughConfig,
     PrivateKeyJwtAuth,
+    ResolvedCredential,
     ServerSpec,
     SharedKey,
     Subject,
@@ -448,3 +450,32 @@ def _client_auth_fingerprint(client_auth: ClientAuth) -> str:
 
 def _not_implemented(kind: AuthSpecKind) -> Result[httpx.Auth, CredError]:
     return Error(CredError.of_not_implemented(f"{kind.value}: resolver arm not implemented yet"))
+
+
+async def resolve_credentials_with_source(
+    provider: UpstreamCredentialProvider, subject: Subject, server: ServerSpec
+) -> Result[ResolvedCredential, CredError]:
+    match await provider.resolve_credentials(subject, server):
+        case Error(err):
+            return Error(err)
+        case Ok(auth):
+            if isinstance(auth, NoOpAuth):
+                return Ok(ResolvedCredential(auth, AuthResolution.no_auth))
+            match server.config:
+                case NoneConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.no_auth))
+                case ApiKeyConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.static_token))
+                case PassthroughConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.oauth2_passthrough))
+                case ClientCredentialsConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.client_credentials))
+                case TokenExchangeConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.token_exchange))
+                case IdJagConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.id_jag))
+                case AuthorizationCodeConfig():
+                    return Ok(ResolvedCredential(auth, AuthResolution.stored_user_token))
+                case AwsSigV4Config():
+                    return Ok(ResolvedCredential(auth, AuthResolution.aws_sigv4))
+            assert_never(server.config)
diff --git a/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py b/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py
index 632dc57dcf6..d186724fd9f 100644
--- a/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py
+++ b/litellm/proxy/_experimental/mcp_server/outbound_credentials/types.py
@@ -26,10 +26,11 @@ union (see `result.py`), not `expression.Result`.
 from __future__ import annotations
 
 from collections.abc import Mapping
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from enum import Enum
 from typing import Annotated, Final, Literal
 
+import httpx
 from expression import case, tag, tagged_union
 from pydantic import BaseModel, ConfigDict, Field, SecretStr, field_validator
 from typing_extensions import assert_never
@@ -46,6 +47,29 @@ from litellm.types.mcp import (
 )
 
 
+class AuthResolution(str, Enum):
+    no_auth = "no-auth"
+    stored_user_token = "stored-user-token"
+    static_token = "static-token"
+    per_request_header = "per-request-header"
+    oauth2_passthrough = "oauth2-passthrough"
+    client_credentials = "m2m-client-credentials"
+    token_exchange = "token-exchange"
+    id_jag = "id-jag"
+    aws_sigv4 = "aws-sigv4"
+    extra_headers = "extra-headers"
+    not_applicable = "not-applicable"
+    unresolved = "unresolved"
+    failed = "resolution-failed"
+    multiple = "multiple"
+
+
+@dataclass(frozen=True, slots=True)
+class ResolvedCredential:
+    auth: httpx.Auth = field(repr=False)
+    source: AuthResolution
+
+
 class AuthSpecKind(str, Enum):
     """The server's statically-declared upstream-auth mode — derived from its `config`.
 
diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py
index 9a52da0cab1..b53c55ef5fa 100644
--- a/litellm/proxy/_experimental/mcp_server/server.py
+++ b/litellm/proxy/_experimental/mcp_server/server.py
@@ -49,7 +49,11 @@ from litellm.proxy._experimental.mcp_server.mcp_context import (
     _mcp_gateway_server_name,
     _mcp_proxy_mode,  # pyright: ignore[reportPrivateUsage]  # server-owned request mode
 )
-from litellm.proxy._experimental.mcp_server.mcp_debug import MCPDebug
+from litellm.proxy._experimental.mcp_server.mcp_debug import (
+    MCP_AUTH_DIAGNOSTICS_SCOPE_KEY,
+    MCPAuthDiagnostics,
+    MCPDebug,
+)
 from litellm.proxy._experimental.mcp_server.oauth_utils import (
     _redact_mcp_resource_url,
     get_passthrough_www_authenticate,
@@ -4472,13 +4476,13 @@ if MCP_AVAILABLE:
                 raw_headers=raw_headers,
                 scope=dict(scope),
                 mcp_servers=mcp_servers,
-                mcp_auth_header=mcp_auth_header,
-                mcp_server_auth_headers=mcp_server_auth_headers,
                 oauth2_headers=oauth2_headers,
                 client_ip=_client_ip,
             )
-            if _debug_headers:
-                send = MCPDebug.wrap_send_with_debug_headers(send, _debug_headers)
+            diagnostics: Final = MCPAuthDiagnostics() if _debug_headers else None
+            if diagnostics is not None:
+                scope[MCP_AUTH_DIAGNOSTICS_SCOPE_KEY] = diagnostics
+                send = MCPDebug.wrap_send_with_debug_headers(send, _debug_headers, diagnostics.headers)
 
             # Ensure session managers are initialized
             if not _SESSION_MANAGERS_INITIALIZED:
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
index 6b3098d9e60..2eab053ab73 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
@@ -1203,3 +1203,69 @@ async def test_passthrough_ignores_the_carrier_and_keeps_the_callers_slot():
     assert isinstance(result, Ok)
     headers, _ = await _emitted_async(result.ok)
     assert headers["Authorization"] == "caller-token"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("config", "subject", "expected_source", "expected_header"),
+    [
+        (NoneConfig(), _SUBJECT, "no-auth", None),
+        (PassthroughConfig(), _SUBJECT, "no-auth", None),
+        (PassthroughConfig(), _with_inbound("Bearer caller-token"), "oauth2-passthrough", "Bearer caller-token"),
+        (ApiKeyConfig(key_source=SharedKey(value=SecretStr("static-key"))), _SUBJECT, "static-token", "Bearer static-key"),
+        (AuthorizationCodeConfig(), Subject(tenant_id="", subject_id="alice"), "stored-user-token", "Bearer stored-alice"),
+    ],
+)
+async def test_resolved_source_matches_the_credential_sent_upstream(config, subject, expected_source, expected_header):
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
+
+    store = _FakeTokenStore({("alice", "s"): OAuthToken(access_token="stored-alice")})
+    provider = UpstreamCredentialProvider(oauth_token_store=store)
+    result = await resolve_credentials_with_source(provider, subject, _spec(config))
+    assert isinstance(result, Ok)
+    assert result.ok.source.value == expected_source
+    assert _emitted(result.ok.auth).get("Authorization") == expected_header
+    assert "stored-alice" not in repr(result.ok)
+    assert "static-key" not in repr(result.ok)
+
+
+@pytest.mark.asyncio
+async def test_resolved_source_preserves_missing_user_token_error():
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
+
+    result = await resolve_credentials_with_source(UpstreamCredentialProvider(), _SUBJECT, _spec(AuthorizationCodeConfig()))
+    assert isinstance(result, Error)
+    assert result.error.tag == "unauthorized"
+
+
+@pytest.mark.asyncio
+async def test_minted_token_sources_match_egress_and_do_not_fetch_twice():
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
+
+    source = _FakeM2MSource(Ok(OAuthToken(access_token="m2m-at")))
+    m2m = await resolve_credentials_with_source(UpstreamCredentialProvider(client_credentials_source=source), _SUBJECT, _spec(_M2M))
+    assert isinstance(m2m, Ok)
+    headers, _ = await _emitted_async(m2m.ok.auth)
+    assert headers["Authorization"] == "Bearer m2m-at"
+    assert m2m.ok.source.value == "m2m-client-credentials"
+    assert source.gets == ["s"]
+
+    exchanger = _FakeExchanger(Ok(OAuthToken(access_token="exchanged-at")))
+    exchanged = await resolve_credentials_with_source(UpstreamCredentialProvider(token_exchanger=exchanger), _with_inbound("subject"), _spec(_OBO))
+    assert isinstance(exchanged, Ok)
+    assert _emitted(exchanged.ok.auth)["Authorization"] == "Bearer exchanged-at"
+    assert exchanged.ok.source.value == "token-exchange"
+    assert len(exchanger.calls) == 1
+
+
+@pytest.mark.asyncio
+async def test_id_jag_source_describes_final_token_after_both_exchanges():
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
+
+    endpoint = _FakeTokenEndpoint(_two_leg_ok("resource-token"))
+    provider = UpstreamCredentialProvider(token_endpoint=endpoint)
+    result = await resolve_credentials_with_source(provider, _with_inbound("identity-token"), _spec(_id_jag_config()))
+    assert isinstance(result, Ok)
+    assert result.ok.source.value == "id-jag"
+    assert _emitted(result.ok.auth)["Authorization"] == "Bearer resource-token"
+    assert len(endpoint.calls) == 2
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
index d299239f68e..5039209c351 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
@@ -3,11 +3,17 @@ Tests for MCPDebug — MCP OAuth2 debug response headers.
 """
 
 import asyncio
-from unittest.mock import MagicMock
+from typing import Final
+
+import pytest
+from starlette.types import Message
+
+from litellm.proxy._experimental.mcp_server.outbound_credentials.types import AuthResolution
 
 from litellm.proxy._experimental.mcp_server.mcp_debug import (
     MCP_DEBUG_REQUEST_HEADER,
     MCPDebug,
+    MCPAuthDiagnostics,
 )
 
 
@@ -166,77 +172,6 @@ class TestBuildDebugHeaders:
         assert set(headers.keys()) == expected_keys
 
 
-class TestResolveAuthResolution:
-    def _make_server(self, **kwargs):
-        server = MagicMock()
-        server.alias = kwargs.get("alias", "test")
-        server.server_name = kwargs.get("server_name", "test")
-        server.has_client_credentials = kwargs.get("has_client_credentials", False)
-        server.authentication_token = kwargs.get("authentication_token", None)
-        server.auth_type = kwargs.get("auth_type", None)
-        return server
-
-    def test_per_request_header(self):
-        server = self._make_server()
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header="Bearer xxx",
-            mcp_server_auth_headers=None,
-            oauth2_headers=None,
-        )
-        assert result == "per-request-header"
-
-    def test_server_specific_header(self):
-        server = self._make_server(alias="atlas")
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header=None,
-            mcp_server_auth_headers={"atlas": {"Authorization": "Bearer xxx"}},
-            oauth2_headers=None,
-        )
-        assert result == "per-request-header"
-
-    def test_m2m(self):
-        server = self._make_server(has_client_credentials=True)
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header=None,
-            mcp_server_auth_headers=None,
-            oauth2_headers=None,
-        )
-        assert result == "m2m-client-credentials"
-
-    def test_static_token(self):
-        server = self._make_server(authentication_token="static-tok")
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header=None,
-            mcp_server_auth_headers=None,
-            oauth2_headers=None,
-        )
-        assert result == "static-token"
-
-    def test_oauth2_passthrough(self):
-        server = self._make_server(auth_type="oauth2")
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header=None,
-            mcp_server_auth_headers=None,
-            oauth2_headers={"Authorization": "Bearer eyJ..."},
-        )
-        assert result == "oauth2-passthrough"
-
-    def test_no_auth(self):
-        server = self._make_server()
-        result = MCPDebug.resolve_auth_resolution(
-            server,
-            mcp_auth_header=None,
-            mcp_server_auth_headers=None,
-            oauth2_headers=None,
-        )
-        assert result == "no-auth"
-
-
 class TestWrapSendWithDebugHeaders:
     def test_injects_headers(self):
         captured = []
@@ -269,3 +204,85 @@ class TestWrapSendWithDebugHeaders:
         asyncio.run(wrapped(body_msg))
 
         assert captured[0] == body_msg
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("source", tuple(AuthResolution))
+async def test_debug_uses_resolution_recorded_after_response_start(source: AuthResolution) -> None:
+    captured: Final[list[Message]] = []
+    diagnostics: Final = MCPAuthDiagnostics()
+
+    async def send(message: Message) -> None:
+        captured.append(message)
+
+    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(send, {}, diagnostics.headers)
+    await wrapped({"type": "http.response.start", "status": 200, "headers": []})
+    assert captured == []
+    diagnostics.record("s1", source)
+    body: Final[Message] = {"type": "http.response.body", "body": b"data: pong\n\n", "more_body": True}
+    await wrapped(body)
+    assert dict(captured[0]["headers"])[b"x-mcp-debug-auth-resolution"] == source.value.encode()
+    assert captured[1] == body
+
+
+@pytest.mark.asyncio
+async def test_early_stream_frame_reports_unresolved_without_waiting() -> None:
+    captured: Final[list[Message]] = []
+    diagnostics: Final = MCPAuthDiagnostics()
+
+    async def send(message: Message) -> None:
+        captured.append(message)
+
+    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(send, {}, diagnostics.headers)
+    await wrapped({"type": "http.response.start", "status": 200, "headers": []})
+    await wrapped({"type": "http.response.body", "body": b": ping\n\n", "more_body": True})
+    diagnostics.record("s1", AuthResolution.stored_user_token)
+    await wrapped({"type": "http.response.body", "body": b"data: pong\n\n", "more_body": False})
+    assert len(captured) == 3
+    assert dict(captured[0]["headers"])[b"x-mcp-debug-auth-resolution"] == b"unresolved"
+
+
+def test_diagnostics_keep_requests_separate_and_do_not_collapse_multiple_servers() -> None:
+    alice: Final = MCPAuthDiagnostics()
+    bob: Final = MCPAuthDiagnostics()
+    alice.record("s1", AuthResolution.stored_user_token)
+    assert bob.resolution() == "unresolved"
+    alice.record("s1", AuthResolution.token_exchange)
+    assert alice.resolution() == "token-exchange"
+    alice.record("s2", AuthResolution.static_token)
+    assert alice.resolution() == "multiple"
+    assert alice.headers()["x-mcp-debug-auth-resolutions"] == '{"s1":"token-exchange","s2":"static-token"}'
+
+
+@pytest.mark.asyncio
+async def test_concurrent_mcp_messages_record_on_their_own_http_scope() -> None:
+    from unittest.mock import MagicMock
+
+    from mcp.server.lowlevel.server import request_ctx
+    from mcp.shared.context import RequestContext
+    from starlette.requests import Request
+
+    from litellm.proxy._experimental.mcp_server.mcp_debug import (
+        MCP_AUTH_DIAGNOSTICS_SCOPE_KEY,
+        record_auth_resolution,
+    )
+
+    session: Final = MagicMock()
+    first: Final = MCPAuthDiagnostics()
+    second: Final = MCPAuthDiagnostics()
+
+    async def record(diagnostics: MCPAuthDiagnostics, source: AuthResolution) -> None:
+        context: Final = RequestContext(
+            request_id=1, meta=None, session=session, lifespan_context=None,
+            request=Request({"type": "http", MCP_AUTH_DIAGNOSTICS_SCOPE_KEY: diagnostics}),
+        )
+        token: Final = request_ctx.set(context)
+        try:
+            await asyncio.sleep(0)
+            record_auth_resolution("same-server", source)
+        finally:
+            request_ctx.reset(token)
+
+    await asyncio.gather(record(first, AuthResolution.stored_user_token), record(second, AuthResolution.per_request_header))
+    assert first.resolution() == "stored-user-token"
+    assert second.resolution() == "per-request-header"
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 46fef83092d..99447acd371 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
@@ -12567,3 +12567,109 @@ async def test_pre_call_tool_check_honors_guardrail_attached_to_key(monkeypatch,
     with pytest.raises(HTTPException) as exc_info:
         await call
     assert exc_info.value.status_code == 400
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("config", "extra_headers", "expected_source", "expected_authorization"),
+    [
+        ("stored", None, "stored-user-token", "Bearer stored-token"),
+        ("stored", {"aUtHoRiZaTiOn": "Bearer injected"}, "stored-user-token", "Bearer stored-token"),
+        ("static", None, "static-token", "Bearer static-token"),
+        ("static", {"authorization": "Bearer injected"}, "extra-headers", "Bearer injected"),
+        ("none", None, "no-auth", None),
+        ("none", {"Authorization": "Bearer injected"}, "extra-headers", "Bearer injected"),
+    ],
+)
+async def test_debug_resolution_matches_final_header_conflict_winner(
+    config, extra_headers, expected_source, expected_authorization
+):
+    from mcp.server.lowlevel.server import request_ctx
+    from mcp.shared.context import RequestContext
+    from starlette.requests import Request
+    from pydantic import SecretStr
+
+    from litellm.proxy._experimental.mcp_server.auth.litellm_auth_handler import MCPAuthenticatedUser
+    from litellm.proxy._experimental.mcp_server.mcp_debug import MCP_AUTH_DIAGNOSTICS_SCOPE_KEY, MCPAuthDiagnostics
+    from litellm.proxy._experimental.mcp_server.outbound_credentials import (
+        ApiKeyConfig, AuthorizationCodeConfig, NoneConfig, ServerSpec, SharedKey, UpstreamCredentialProvider,
+    )
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.oauth_token_store import OAuthToken
+    from litellm.types.mcp_server.mcp_server_manager import MCPServer
+
+    class Store:
+        def __init__(self):
+            self.calls = 0
+
+        async def fetch(self, user_id: str, server_id: str) -> OAuthToken | None:
+            self.calls += 1
+            return OAuthToken(access_token="stored-token") if user_id == "alice" else None
+
+    store = Store()
+    context = MCPAuthenticatedUser(UserAPIKeyAuth(user_id="alice"))
+    diagnostics = MCPAuthDiagnostics()
+    token = request_ctx.set(RequestContext(
+        request_id=1, meta=None, session=MagicMock(), lifespan_context=None,
+        request=Request({"type": "http", MCP_AUTH_DIAGNOSTICS_SCOPE_KEY: diagnostics}),
+    ))
+    selected = {
+        "stored": AuthorizationCodeConfig(),
+        "static": ApiKeyConfig(key_source=SharedKey(value=SecretStr("static-token"))),
+        "none": NoneConfig(),
+    }[config]
+    try:
+        auth, remaining = await MCPServerManager()._resolve_v2_auth(
+            server=MCPServer(
+                server_id="s", name="s", transport="http", url="https://up.example/mcp",
+                static_headers={"Authorization": "Bearer configured"},
+            ),
+            spec=ServerSpec(server_id="s", resource="https://up.example/mcp", config=selected),
+            provider=UpstreamCredentialProvider(oauth_token_store=store),
+            subject_token=None,
+            user_api_key_auth=context.user_api_key_auth,
+            extra_headers=extra_headers,
+        )
+        request = httpx.Request("GET", "https://up.example/mcp", headers=remaining)
+        if auth is not None:
+            next(auth.auth_flow(request))
+        assert diagnostics.resolution() == expected_source
+        assert request.headers.get("Authorization") == expected_authorization
+        assert store.calls == (1 if config == "stored" else 0)
+    finally:
+        request_ctx.reset(token)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("transport", ["http", "stdio"])
+async def test_debug_reports_legacy_signing_and_non_http_transport(transport):
+    from mcp.server.lowlevel.server import request_ctx
+    from mcp.shared.context import RequestContext
+    from starlette.requests import Request
+
+    from litellm.proxy._experimental.mcp_server.mcp_debug import MCP_AUTH_DIAGNOSTICS_SCOPE_KEY, MCPAuthDiagnostics
+    from litellm.types.mcp_server.mcp_server_manager import MCPServer
+
+    diagnostics = MCPAuthDiagnostics()
+    token = request_ctx.set(RequestContext(
+        request_id=1, meta=None, session=MagicMock(), lifespan_context=None,
+        request=Request({"type": "http", MCP_AUTH_DIAGNOSTICS_SCOPE_KEY: diagnostics}),
+    ))
+    try:
+        server = MCPServer(
+            server_id="signed", name="signed", transport=transport,
+            url="https://up.example/mcp", auth_type="aws_sigv4",
+            aws_access_key_id="AKIDEXAMPLE", aws_secret_access_key="test-signing-secret",
+            aws_region_name="us-east-1", aws_service_name="execute-api",
+            command="python", args=["-c", "pass"],
+        )
+        client = await MCPServerManager()._create_mcp_client(server)
+        if transport == "stdio":
+            assert diagnostics.resolution() == "not-applicable"
+        else:
+            assert diagnostics.resolution() == "aws-sigv4"
+            request = httpx.Request("POST", "https://up.example/mcp", content=b"{}")
+            next(client._aws_auth.auth_flow(request))
+            assert request.headers["Authorization"].startswith("AWS4-HMAC-SHA256 ")
+            assert "Credential=AKIDEXAMPLE/" in request.headers["Authorization"]
+    finally:
+        request_ctx.reset(token)

From c972bbe80fb575fb0a162e79547edbe1db22298d Mon Sep 17 00:00:00 2001
From: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com>
Date: Wed, 9 Sep 2026 16:22:41 -0700
Subject: [PATCH 2/4] fix(mcp): send debug headers immediately for GET streams

---
 .../proxy/_experimental/mcp_server/mcp_debug.py   |  8 ++++++--
 litellm/proxy/_experimental/mcp_server/server.py  |  4 +++-
 .../_experimental/mcp_server/test_mcp_debug.py    | 15 ++++++++++-----
 3 files changed, 19 insertions(+), 8 deletions(-)

diff --git a/litellm/proxy/_experimental/mcp_server/mcp_debug.py b/litellm/proxy/_experimental/mcp_server/mcp_debug.py
index b25ed27becc..ff30ef99ebd 100644
--- a/litellm/proxy/_experimental/mcp_server/mcp_debug.py
+++ b/litellm/proxy/_experimental/mcp_server/mcp_debug.py
@@ -308,14 +308,18 @@ class MCPDebug:
 
     @staticmethod
     def wrap_send_with_debug_headers(
-        send: Send, debug_headers: Mapping[str, str], resolution: Callable[[], Mapping[str, str]] | None = None
+        send: Send,
+        debug_headers: Mapping[str, str],
+        resolution: Callable[[], Mapping[str, str]] | None = None,
+        *,
+        request_method: str | None = None,
     ) -> Send:
         """
         Return a new ASGI ``send`` callable that injects *debug_headers*
         into the ``http.response.start`` message.
         """
 
-        if resolution is not None:
+        if resolution is not None and request_method == "POST":
             return _DiagnosticSend(send, debug_headers, resolution)
 
         async def _send_with_debug(message: Message) -> None:
diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py
index b53c55ef5fa..72d256e90fc 100644
--- a/litellm/proxy/_experimental/mcp_server/server.py
+++ b/litellm/proxy/_experimental/mcp_server/server.py
@@ -4482,7 +4482,9 @@ if MCP_AVAILABLE:
             diagnostics: Final = MCPAuthDiagnostics() if _debug_headers else None
             if diagnostics is not None:
                 scope[MCP_AUTH_DIAGNOSTICS_SCOPE_KEY] = diagnostics
-                send = MCPDebug.wrap_send_with_debug_headers(send, _debug_headers, diagnostics.headers)
+                send = MCPDebug.wrap_send_with_debug_headers(
+                    send, _debug_headers, diagnostics.headers, request_method=scope.get("method")
+                )
 
             # Ensure session managers are initialized
             if not _SESSION_MANAGERS_INITIALIZED:
diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
index 5039209c351..7d46bb0237a 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_debug.py
@@ -208,20 +208,25 @@ class TestWrapSendWithDebugHeaders:
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("source", tuple(AuthResolution))
-async def test_debug_uses_resolution_recorded_after_response_start(source: AuthResolution) -> None:
+@pytest.mark.parametrize("method", ("GET", "DELETE", "POST"))
+async def test_debug_defers_resolution_until_first_frame_only_for_post(source: AuthResolution, method: str) -> None:
     captured: Final[list[Message]] = []
     diagnostics: Final = MCPAuthDiagnostics()
 
     async def send(message: Message) -> None:
         captured.append(message)
 
-    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(send, {}, diagnostics.headers)
+    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(
+        send, diagnostics.headers(), diagnostics.headers, request_method=method
+    )
     await wrapped({"type": "http.response.start", "status": 200, "headers": []})
-    assert captured == []
+    assert len(captured) == (0 if method == "POST" else 1)
     diagnostics.record("s1", source)
     body: Final[Message] = {"type": "http.response.body", "body": b"data: pong\n\n", "more_body": True}
     await wrapped(body)
-    assert dict(captured[0]["headers"])[b"x-mcp-debug-auth-resolution"] == source.value.encode()
+    assert dict(captured[0]["headers"])[b"x-mcp-debug-auth-resolution"] == (
+        source.value.encode() if method == "POST" else b"unresolved"
+    )
     assert captured[1] == body
 
 
@@ -233,7 +238,7 @@ async def test_early_stream_frame_reports_unresolved_without_waiting() -> None:
     async def send(message: Message) -> None:
         captured.append(message)
 
-    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(send, {}, diagnostics.headers)
+    wrapped: Final = MCPDebug.wrap_send_with_debug_headers(send, {}, diagnostics.headers, request_method="POST")
     await wrapped({"type": "http.response.start", "status": 200, "headers": []})
     await wrapped({"type": "http.response.body", "body": b": ping\n\n", "more_body": True})
     diagnostics.record("s1", AuthResolution.stored_user_token)

From fb9bb60e80baed40d80ef0cbd7c0482fbcd20e62 Mon Sep 17 00:00:00 2001
From: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com>
Date: Wed, 9 Sep 2026 16:55:52 -0700
Subject: [PATCH 3/4] test(mcp): cover auth diagnostics through HTTP handler

---
 .../mcp_server/test_mcp_server.py             | 162 ++++++++----------
 1 file changed, 76 insertions(+), 86 deletions(-)

diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py
index 9b6eaba3177..12805e355f9 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/test_mcp_server.py
@@ -3,6 +3,7 @@ import contextvars
 import os
 from datetime import datetime, timedelta
 from types import SimpleNamespace
+from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -1866,96 +1867,85 @@ async def test_streamable_http_session_manager_is_stateless():
 
 
 @pytest.mark.asyncio
-async def test_mcp_routing_initialize_to_stateful_no_session_to_stateless():
-    """
-    Test that routing correctly sends:
-    - initialize (no mcp-session-id) → stateful manager (so client gets mcp-session-id)
-    - tools/list (no mcp-session-id) → stateless manager (curl, Inspector)
-    """
-    try:
-        from litellm.proxy._experimental.mcp_server.server import (
-            handle_streamable_http_mcp,
-            session_manager_stateful,
-            session_manager_stateless,
+@pytest.mark.parametrize("debug", (False, True))
+@pytest.mark.parametrize(
+    ("method", "request_body", "stateful"),
+    (
+        ("POST", b'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}', True),
+        ("POST", b'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}', False),
+        ("GET", b"", False),
+        ("DELETE", b"", False),
+    ),
+)
+async def test_mcp_routing_initialize_to_stateful_no_session_to_stateless(
+    debug: bool, method: str, request_body: bytes, stateful: bool
+) -> None:
+    from mcp.server.lowlevel.server import request_ctx
+    from mcp.shared.context import RequestContext
+    from starlette.requests import Request
+    from starlette.types import Message, Receive, Scope, Send
+
+    from litellm.proxy._experimental.mcp_server import server as mcp_module
+    from litellm.proxy._experimental.mcp_server.mcp_debug import record_auth_resolution
+    from litellm.proxy._experimental.mcp_server.outbound_credentials.types import AuthResolution
+
+    scope: Final[Scope] = {"type": "http", "method": method, "path": "/mcp", "headers": []}
+    receive: Final = AsyncMock(return_value={"type": "http.request", "body": request_body, "more_body": False})
+    send: Final = AsyncMock()
+    observe_start: Final = AsyncMock()
+    body: Final[Message] = {"type": "http.response.body", "body": b"data: pong\n\n", "more_body": True}
+
+    async def handle_request(request_scope: Scope, receive: Receive, outgoing: Send) -> None:
+        await outgoing({"type": "http.response.start", "status": 200, "headers": []})
+        await observe_start(send.await_count)
+        context: Final = RequestContext(
+            request_id=1, meta=None, session=MagicMock(), lifespan_context=None, request=Request(request_scope)
         )
-    except ImportError:
-        pytest.skip("MCP server not available")
+        token: Final = request_ctx.set(context)
+        try:
+            record_auth_resolution("s1", AuthResolution.stored_user_token)
+        finally:
+            request_ctx.reset(token)
+        await outgoing(body)
 
-    async def make_request(method_body: bytes, path: str = "/mcp/progress_test"):
-        scope = {
-            "type": "http",
-            "method": "POST",
-            "path": path,
-            "headers": [
-                (b"content-type", b"application/json"),
-                (b"authorization", b"Bearer test-key"),
-            ],
-        }
-        receive = AsyncMock(
-            return_value={
-                "type": "http.request",
-                "body": method_body,
-                "more_body": False,
-            }
-        )
-        send = AsyncMock()
-
-        stateless_called = []
-        stateful_called = []
-
-        async def stateless_handle(s, r, se):
-            stateless_called.append(1)
-
-        async def stateful_handle(s, r, se):
-            stateful_called.append(1)
-
-        with (
-            patch(
-                "litellm.proxy._experimental.mcp_server.server.extract_mcp_auth_context",
-                new_callable=AsyncMock,
-                return_value=(MagicMock(), None, ["progress_test"], None, None, None),
+    stateless_handle: Final = AsyncMock(side_effect=handle_request)
+    stateful_handle: Final = AsyncMock(side_effect=handle_request)
+    with (
+        patch(
+            "litellm.proxy._experimental.mcp_server.server.extract_mcp_auth_context",
+            new_callable=AsyncMock,
+            return_value=(
+                UserAPIKeyAuth(user_id="debug-user"),
+                None,
+                None,
+                None,
+                None,
+                {"x-litellm-mcp-debug": "true"} if debug else {},
             ),
-            patch(
-                "litellm.proxy._experimental.mcp_server.server.set_auth_context",
-            ),
-            patch(
-                "litellm.proxy._experimental.mcp_server.server._SESSION_MANAGERS_INITIALIZED",
-                True,
-            ),
-            patch.object(
-                session_manager_stateless,
-                "handle_request",
-                side_effect=stateless_handle,
-            ),
-            patch.object(
-                session_manager_stateful,
-                "handle_request",
-                side_effect=stateful_handle,
-            ),
-            patch.object(
-                session_manager_stateless,
-                "_server_instances",
-                {},
-            ),
-            patch.object(
-                session_manager_stateful,
-                "_server_instances",
-                {},
-            ),
-        ):
-            await handle_streamable_http_mcp(scope, receive, send)
+        ),
+        patch("litellm.proxy._experimental.mcp_server.server._SESSION_MANAGERS_INITIALIZED", True),
+        patch(
+            "litellm.proxy._experimental.mcp_server.server.session_manager_stateless",
+            SimpleNamespace(handle_request=stateless_handle),
+        ),
+        patch(
+            "litellm.proxy._experimental.mcp_server.server.session_manager_stateful",
+            SimpleNamespace(handle_request=stateful_handle),
+        ),
+    ):
+        await mcp_module.handle_streamable_http_mcp(scope, receive, send)
 
-        return bool(stateless_called), bool(stateful_called)
-
-    # initialize → stateful
-    init_body = b'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05"}}'
-    stateless_called, stateful_called = await make_request(init_body)
-    assert stateful_called and not stateless_called, "initialize (no session) should route to stateful, not stateless"
-
-    # tools/list → stateless
-    tools_body = b'{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'
-    stateless_called, stateful_called = await make_request(tools_body)
-    assert stateless_called and not stateful_called, "tools/list (no session) should route to stateless, not stateful"
+    assert stateful_handle.await_count == (1 if stateful else 0)
+    assert stateless_handle.await_count == (0 if stateful else 1)
+    observe_start.assert_awaited_once_with(0 if debug and method == "POST" else 1)
+    assert send.await_count == 2
+    assert send.call_args_list[0].args[0]["status"] == 200
+    assert send.call_args_list[1].args[0] == body
+    headers: Final = dict(send.call_args_list[0].args[0]["headers"])
+    if debug:
+        assert headers[b"x-mcp-debug-auth-resolution"] == (b"stored-user-token" if method == "POST" else b"unresolved")
+    else:
+        assert not any(name.startswith(b"x-mcp-debug") for name in headers)
 
 
 @pytest.mark.asyncio

From 95f6e96ef982d0244cfa8607d5c4fd1f2703867c Mon Sep 17 00:00:00 2001
From: Joshua Valluru <326636767+joshua-berri@users.noreply.github.com>
Date: Wed, 9 Sep 2026 20:57:33 -0700
Subject: [PATCH 4/4] test(mcp): type auth diagnostics regression parameters

---
 .../outbound_credentials/test_resolver.py           | 11 +++++++----
 .../mcp_server/test_mcp_server_manager.py           | 13 ++++++++-----
 2 files changed, 15 insertions(+), 9 deletions(-)

diff --git a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
index 2eab053ab73..5fab4ceec72 100644
--- a/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
+++ b/tests/test_litellm/proxy/_experimental/mcp_server/outbound_credentials/test_resolver.py
@@ -19,6 +19,7 @@ from pydantic import SecretStr
 
 from litellm.proxy._experimental.mcp_server.outbound_credentials import (
     ApiKeyConfig,
+    AuthConfig,
     AuthorizationCodeConfig,
     AwsSigV4Config,
     Byok,
@@ -1216,7 +1217,9 @@ async def test_passthrough_ignores_the_carrier_and_keeps_the_callers_slot():
         (AuthorizationCodeConfig(), Subject(tenant_id="", subject_id="alice"), "stored-user-token", "Bearer stored-alice"),
     ],
 )
-async def test_resolved_source_matches_the_credential_sent_upstream(config, subject, expected_source, expected_header):
+async def test_resolved_source_matches_the_credential_sent_upstream(
+    config: AuthConfig, subject: Subject, expected_source: str, expected_header: str | None
+) -> None:
     from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
 
     store = _FakeTokenStore({("alice", "s"): OAuthToken(access_token="stored-alice")})
@@ -1230,7 +1233,7 @@ async def test_resolved_source_matches_the_credential_sent_upstream(config, subj
 
 
 @pytest.mark.asyncio
-async def test_resolved_source_preserves_missing_user_token_error():
+async def test_resolved_source_preserves_missing_user_token_error() -> None:
     from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
 
     result = await resolve_credentials_with_source(UpstreamCredentialProvider(), _SUBJECT, _spec(AuthorizationCodeConfig()))
@@ -1239,7 +1242,7 @@ async def test_resolved_source_preserves_missing_user_token_error():
 
 
 @pytest.mark.asyncio
-async def test_minted_token_sources_match_egress_and_do_not_fetch_twice():
+async def test_minted_token_sources_match_egress_and_do_not_fetch_twice() -> None:
     from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
 
     source = _FakeM2MSource(Ok(OAuthToken(access_token="m2m-at")))
@@ -1259,7 +1262,7 @@ async def test_minted_token_sources_match_egress_and_do_not_fetch_twice():
 
 
 @pytest.mark.asyncio
-async def test_id_jag_source_describes_final_token_after_both_exchanges():
+async def test_id_jag_source_describes_final_token_after_both_exchanges() -> None:
     from litellm.proxy._experimental.mcp_server.outbound_credentials.resolver import resolve_credentials_with_source
 
     endpoint = _FakeTokenEndpoint(_two_leg_ok("resource-token"))
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 99447acd371..adf985e9a21 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
@@ -5,7 +5,7 @@ import logging
 import os
 import sys
 from datetime import datetime
-from typing import Any, Dict, Final, Optional
+from typing import Any, Dict, Final, Literal, Optional
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
@@ -12582,8 +12582,11 @@ async def test_pre_call_tool_check_honors_guardrail_attached_to_key(monkeypatch,
     ],
 )
 async def test_debug_resolution_matches_final_header_conflict_winner(
-    config, extra_headers, expected_source, expected_authorization
-):
+    config: Literal["stored", "static", "none"],
+    extra_headers: dict[str, str] | None,
+    expected_source: str,
+    expected_authorization: str | None,
+) -> None:
     from mcp.server.lowlevel.server import request_ctx
     from mcp.shared.context import RequestContext
     from starlette.requests import Request
@@ -12598,7 +12601,7 @@ async def test_debug_resolution_matches_final_header_conflict_winner(
     from litellm.types.mcp_server.mcp_server_manager import MCPServer
 
     class Store:
-        def __init__(self):
+        def __init__(self) -> None:
             self.calls = 0
 
         async def fetch(self, user_id: str, server_id: str) -> OAuthToken | None:
@@ -12641,7 +12644,7 @@ async def test_debug_resolution_matches_final_header_conflict_winner(
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize("transport", ["http", "stdio"])
-async def test_debug_reports_legacy_signing_and_non_http_transport(transport):
+async def test_debug_reports_legacy_signing_and_non_http_transport(transport: Literal["http", "stdio"]) -> None:
     from mcp.server.lowlevel.server import request_ctx
     from mcp.shared.context import RequestContext
     from starlette.requests import Request