fix(oauth): narrow async_auth_flow lock scope to avoid blocking long-poll requests

Previously the entire `OAuthClientProvider.async_auth_flow` body ran under
`self.context.lock`, including the `yield request` that hands the request
off to httpx. For requests that complete quickly this is fine, but a GET
SSE long-poll holds the lock for the full SSE lifetime — which means any
concurrent POST (e.g. `tools/call`) is blocked waiting for the lock,
producing a ~16s perceived stall on lazy MCP connections that use OAuth.

This commit splits the single coarse lock into purpose-specific scopes:

  Phase 1 (context.lock): initialize state, capture protocol_version, and
    decide whether a refresh is needed. Short-held; no HTTP I/O.
  Phase 2 (refresh_lock, new): single-flight token refresh. The refresh
    request `yield` happens outside any lock. A double-check inside
    `context.lock` ensures concurrent waiters do not redundantly refresh
    after another coroutine completed one.
  Phase 3 (no lock): add the auth header and yield the actual request.
    GET SSE long-polls and other long-running requests no longer block
    unrelated traffic.
  Phase 4 (context.lock): 401 / 403 full OAuth re-auth path. Conservatively
    kept under lock because this path is rare and its yielded sub-requests
    (metadata discovery, registration, token exchange) hit the AS, not the
    resource server. A future refactor can narrow this further.

Lock additions:
  - `OAuthContext.refresh_lock: anyio.Lock` provides single-flight refresh
    so concurrent requests trigger at most one token refresh.

Behavior changes:
  - Concurrent requests through the same `OAuthClientProvider` no longer
    serialize at the lock. GET SSE long-polls and POSTs now proceed in
    parallel.
  - Token refresh remains serialized (via `refresh_lock`), preserving the
    invariant that only one refresh request is in flight at a time.
  - Public API and behavior are otherwise unchanged.

Related upstream issue:
  #1326

Author: peisuke

src/mcp/client/auth/oauth2.py

@@ -114,7 +114,18 @@ class OAuthContext:
     token_expiry_time: float | None = None
 
     # State
+    #
+    # `lock` guards short-lived reads/writes of provider state (initialization
+    # flag, token cache mutation, protocol_version assignment). It is held only
+    # while mutating state and is released before any HTTP request is yielded
+    # so a long-running request (e.g. GET SSE long-poll) does not block
+    # unrelated concurrent requests.
+    #
+    # `refresh_lock` provides single-flight semantics for token refresh: only
+    # one concurrent refresh fires; other waiters block on this lock, then
+    # re-check the token cache and proceed without re-refreshing.
     lock: anyio.Lock = field(default_factory=anyio.Lock)
+    refresh_lock: anyio.Lock = field(default_factory=anyio.Lock)
 
     def get_authorization_base_url(self, server_url: str) -> str:
         """Extract base URL by removing path component."""
@@ -504,7 +515,17 @@ async def _validate_resource_match(self, prm: ProtectedResourceMetadata) -> None
             raise OAuthFlowError(f"Protected resource {prm_resource} does not match expected {default_resource}")
 
     async def async_auth_flow(self, request: httpx.Request) -> AsyncGenerator[httpx.Request, httpx.Response]:
-        """HTTPX auth flow integration."""
+        """HTTPX auth flow integration.
+
+        Lock scope:
+          ``self.context.lock`` is held only while reading/mutating provider
+          state. The actual HTTP request yield (which may be a long-poll GET
+          SSE stream) runs outside any lock so concurrent unrelated requests
+          are not blocked. ``self.context.refresh_lock`` provides
+          single-flight semantics for token refresh.
+        """
+        # === Phase 1: state read + refresh decision (brief context.lock) ===
+        needs_refresh = False
         async with self.context.lock:
             if not self._initialized:
                 await self._initialize()  # pragma: no cover
@@ -513,20 +534,43 @@ async def async_auth_flow(self, request: httpx.Request) -> AsyncGenerator[httpx.
             self.context.protocol_version = request.headers.get(MCP_PROTOCOL_VERSION)
 
             if not self.context.is_token_valid() and self.context.can_refresh_token():
-                # Try to refresh token
-                refresh_request = await self._refresh_token()  # pragma: no cover
-                refresh_response = yield refresh_request  # pragma: no cover
-
-                if not await self._handle_refresh_response(refresh_response):  # pragma: no cover
-                    # Refresh failed, need full re-authentication
-                    self._initialized = False
-
-            if self.context.is_token_valid():
-                self._add_auth_header(request)
-
-            response = yield request
-
-            if response.status_code == 401:
+                needs_refresh = True
+
+        # === Phase 2: single-flight token refresh (yield outside context.lock) ===
+        if needs_refresh:
+            async with self.context.refresh_lock:
+                # Re-check under context.lock: another coroutine may already have
+                # refreshed while we were waiting on refresh_lock.
+                async with self.context.lock:
+                    still_invalid = (
+                        not self.context.is_token_valid()
+                        and self.context.can_refresh_token()
+                    )
+                    if still_invalid:
+                        refresh_request = await self._refresh_token()  # pragma: no cover
+                if still_invalid:
+                    # yield runs outside any lock so a long network round trip
+                    # does not block unrelated concurrent requests.
+                    refresh_response = yield refresh_request  # pragma: no cover
+                    async with self.context.lock:
+                        if not await self._handle_refresh_response(refresh_response):  # pragma: no cover
+                            # Refresh failed; fall through to 401 handling below.
+                            self._initialized = False
+
+        # === Phase 3: send request (no lock; safe for long-poll GET SSE) ===
+        if self.context.is_token_valid():
+            self._add_auth_header(request)
+
+        response = yield request
+
+        # === Phase 4: 401 / 403 full OAuth flow ===
+        # NOTE: Phase 4 yields multiple sub-requests (discovery, registration,
+        # token exchange) under context.lock. This is the existing behavior and
+        # is acceptable because the 401 path is exceptional and not concurrent
+        # with steady-state traffic. A future refactor could narrow the lock
+        # here in the same pattern as Phase 1-2.
+        if response.status_code == 401:
+            async with self.context.lock:
                 # Perform full OAuth flow
                 try:
                     # OAuth flow must be inline due to generator constraints
@@ -619,7 +663,8 @@ async def async_auth_flow(self, request: httpx.Request) -> AsyncGenerator[httpx.
                 # Retry with new tokens
                 self._add_auth_header(request)
                 yield request
-            elif response.status_code == 403:
+        elif response.status_code == 403:
+            async with self.context.lock:
                 # Step 1: Extract error field from WWW-Authenticate header
                 error = extract_field_from_www_auth(response, "error")