Skip to content
Merged
Show file tree
Hide file tree
Changes from 3 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 8 additions & 6 deletions .github/actions/conformance/client.py
Original file line number Diff line number Diff line change
Expand Up @@ -7,8 +7,9 @@
- MCP_CONFORMANCE_SCENARIO env var -> scenario name
- MCP_CONFORMANCE_CONTEXT env var -> optional JSON (for client-credentials scenarios)
- MCP_CONFORMANCE_PROTOCOL_VERSION env var -> spec version the harness mock
server is speaking (e.g. "2025-11-25", "2026-07-28"). Always set; defaults
to the harness's LATEST_SPEC_VERSION when --spec-version is omitted.
server is speaking (e.g. "2025-11-25", "2026-07-28"). Always set; when
--spec-version is omitted the harness picks per-scenario (LATEST_SPEC_VERSION
for active scenarios, DRAFT_PROTOCOL_VERSION for draft-only ones).
- Server URL as last CLI argument (sys.argv[1])
- Must exit 0 within 30 seconds

Expand Down Expand Up @@ -54,10 +55,11 @@
logger = logging.getLogger(__name__)

#: Spec version the harness is running this scenario at (e.g. "2025-11-25",
#: "2026-07-28"). The harness always sets this (it falls back to its own
#: LATEST_SPEC_VERSION when --spec-version is omitted), so None means we were
#: invoked outside the harness. Handlers that need to take the stateless 2026
#: path will branch on this once the SDK has one; today it is logged only.
#: "2026-07-28"). The harness always sets this (when --spec-version is omitted
#: it picks per-scenario: LATEST_SPEC_VERSION for active scenarios,
#: DRAFT_PROTOCOL_VERSION for draft-only ones), so None means we were invoked
#: outside the harness. Handlers that need to take the stateless 2026 path will
#: branch on this once the SDK has one; today it is logged only.
PROTOCOL_VERSION: str | None = os.environ.get("MCP_CONFORMANCE_PROTOCOL_VERSION")

# Type for async scenario handler functions
Expand Down
48 changes: 42 additions & 6 deletions src/mcp/client/auth/oauth2.py
Original file line number Diff line number Diff line change
Expand Up @@ -424,6 +424,13 @@
# Parse and validate response with scope validation
token_response = await handle_token_response_scopes(response)

# RFC 6749 §5.1: an omitted scope means the granted scope equals the requested
# scope. Record it explicitly so the persisted token is self-describing — the
# SEP-2350 step-up union reads it after a restart, when client_metadata.scope
# has reverted to its constructor value.
if token_response.scope is None:
token_response.scope = self.context.client_metadata.scope

# Store tokens in context
self.context.current_tokens = token_response
self.context.update_token_expiry(token_response)
Expand Down Expand Up @@ -470,6 +477,12 @@
content = await response.aread()
token_response = OAuthToken.model_validate_json(content)

# RFC 6749 §6: an omitted scope on refresh means the scope is unchanged from
# the prior access token. Carry it forward so the persisted token stays
# self-describing for the SEP-2350 step-up union after a restart.
if token_response.scope is None and self.context.current_tokens is not None:
token_response.scope = self.context.current_tokens.scope

Comment on lines +480 to +485

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟣 Pre-existing issue in the block this PR touches: _handle_refresh_response now carries an omitted scope forward per RFC 6749 §6, but an omitted refresh_token (also allowed by §6 and common with non-rotating ASes) is not — the parsed token has refresh_token=None and unconditionally overwrites both context.current_tokens and persistent storage, so the next expiry forces a full interactive re-authorization (or fails for headless clients). The fix is the same one-line pattern this PR adds for scope: backfill token_response.refresh_token from self.context.current_tokens.refresh_token when the response omits it.

Extended reasoning...

The bug. _handle_refresh_response (src/mcp/client/auth/oauth2.py:469-495) parses the refresh response into a fresh OAuthToken and then unconditionally does self.context.current_tokens = token_response followed by await self.context.storage.set_tokens(token_response). OAuthToken.refresh_token is Optional with a default of None (src/mcp/shared/auth.py), so when the authorization server omits refresh_token from the refresh response, the still-valid stored refresh token is overwritten with None both in memory and in persistent storage. RFC 6749 §6 explicitly allows this server behavior — the AS MAY issue a new refresh token; when it does not, the client is expected to keep using the previously issued one. Non-rotating refresh tokens are common in practice (Google and many enterprise ASes omit refresh_token from refresh responses).

Code path. async_auth_flowcan_refresh_token() true → _refresh_token() builds the request → _handle_refresh_response handles the 200. The handler validates the JSON, applies the new scope carry-forward this PR adds (lines 480-484), and then replaces the stored token wholesale. Nothing anywhere in src/mcp/client/auth preserves the prior refresh_token — the only other uses are can_refresh_token() (which requires it) and building the refresh request itself. The only carry-forward logic in the module is the scope backfill introduced two lines above the overwrite.

Step-by-step proof.

  1. Initial authorization grants {access_token: A1, refresh_token: R1, expires_in: 3600}; both are persisted.
  2. An hour later the access token expires. async_auth_flow sees can_refresh_token() is true (R1 is stored) and issues the refresh.
  3. The AS responds 200 {"access_token": "A2", "token_type": "Bearer", "expires_in": 3600} — no refresh_token, which §6 permits and non-rotating ASes routinely do.
  4. OAuthToken.model_validate_json produces a token with refresh_token=None. The handler stores it in context.current_tokens and calls storage.set_tokens, wiping R1 from memory and from persistent storage.
  5. At the next expiry, can_refresh_token() (oauth2.py:142-144) returns False because current_tokens.refresh_token is None. The client silently falls back to a full interactive 401 re-authorization (redirect + callback) — an unnecessary user interaction for interactive clients, and an outright failure for headless or long-running ones. After a process restart the persisted record has also lost the refresh token, so the degradation survives restarts.

Why nothing prevents it. The existing tests always include refresh_token in mocked refresh responses, so the omitted case is uncovered, and there is no other code path that carries the prior value forward.

Severity / scope. This is pre-existing — the overwrite line predates this PR and the PR does not change runtime behavior here. It is flagged because the PR edits this exact handler and implements the identical RFC 6749 §6 carry-forward principle for the sibling scope field two lines above, so this is the natural place to fix it.

Fix. Mirror the scope backfill the PR adds:

if token_response.refresh_token is None and self.context.current_tokens is not None:
    token_response.refresh_token = self.context.current_tokens.refresh_token

placed alongside the new scope carry-forward, before current_tokens is replaced and the token is persisted. A test analogous to test_handle_refresh_response_carries_prior_scope_when_response_omits_it (asserting the stored token keeps the old refresh_token) would cover it.

self.context.current_tokens = token_response
self.context.update_token_expiry(token_response)
await self.context.storage.set_tokens(token_response)
Expand Down Expand Up @@ -578,6 +591,9 @@
logger.debug("Authorization server changed; discarding bound credentials and re-registering")
self.context.client_info = None
self.context.clear_tokens()
# Any cached AS metadata is for the old server; drop it so a failed
# rediscovery cannot leak the old registration/token endpoints into Step 4.
self.context.oauth_metadata = None

asm_discovery_urls = build_oauth_authorization_server_metadata_discovery_urls(
self.context.auth_server_url, self.context.server_url
Comment thread
claude[bot] marked this conversation as resolved.
Expand All @@ -600,6 +616,23 @@
else:
logger.debug(f"OAuth metadata discovery failed: {url}")

# SEP-2352: on the legacy no-PRM path the issuer is only known after ASM
# discovery, so re-evaluate the binding here using the discovered metadata
# issuer (mirroring the bound_issuer fallback in Step 4).
if (
self.context.client_info is not None
and self.context.auth_server_url is None
and self.context.oauth_metadata is not None
and not credentials_match_issuer(
self.context.client_info,
str(self.context.oauth_metadata.issuer),
self.context.client_metadata_url,
)
):
logger.debug("Authorization server changed; discarding bound credentials and re-registering")
self.context.client_info = None
self.context.clear_tokens()

# Step 3: Apply scope selection strategy
self.context.client_metadata.scope = get_client_metadata_scopes(
extract_scope_from_www_auth(response),
Expand All @@ -609,13 +642,16 @@
)

# Step 4: Register client or use URL-based client ID (CIMD)
if not self.context.client_info:
# SEP-2352: bind the credentials to the issuing AS. Prefer the PRM-advertised
# authorization server; on the legacy no-PRM path fall back to the issuer from
# the discovered metadata so the binding is still recorded.
bound_issuer = self.context.auth_server_url
if bound_issuer is None and self.context.oauth_metadata is not None:
bound_issuer = str(self.context.oauth_metadata.issuer)
# SEP-2352: bind the credentials to the issuing AS — but only when ASM
# discovery succeeded, so the registration request below actually targets
# that issuer's registration_endpoint. With no metadata (discovery failed),
# DCR falls back to the resource-server origin's /register; recording that
# as bound to a PRM-advertised AS we never reached would persist a
# mis-bound record that the binding check then accepts indefinitely.
bound_issuer: str | None = None
if self.context.oauth_metadata is not None:
bound_issuer = self.context.auth_server_url or str(self.context.oauth_metadata.issuer)

Check warning on line 654 in src/mcp/client/auth/oauth2.py

View check run for this annotation

Claude / Claude Code Review

Issuer still stamped when discovered metadata lacks registration_endpoint

When the discovered AS metadata omits the optional `registration_endpoint`, `create_client_registration_request` still falls back to the resource origin's `/register`, yet the new gate (`oauth_metadata is not None`) stamps the PRM-advertised issuer on that fallback registration — so the new comment's guarantee that "the registration request below actually targets that issuer's registration_endpoint" doesn't hold on this path. Non-blocking nit (the runtime behavior here is unchanged from before t
Comment thread
claude[bot] marked this conversation as resolved.
Outdated

if should_use_client_metadata_url(
self.context.oauth_metadata, self.context.client_metadata_url
Expand Down
196 changes: 194 additions & 2 deletions tests/client/test_auth.py
Original file line number Diff line number Diff line change
Expand Up @@ -51,13 +51,13 @@ def __init__(self):
self._client_info: OAuthClientInformationFull | None = None

async def get_tokens(self) -> OAuthToken | None:
return self._tokens # pragma: no cover
return self._tokens

async def set_tokens(self, tokens: OAuthToken) -> None:
self._tokens = tokens

async def get_client_info(self) -> OAuthClientInformationFull | None:
return self._client_info # pragma: no cover
return self._client_info

async def set_client_info(self, client_info: OAuthClientInformationFull) -> None:
self._client_info = client_info
Expand Down Expand Up @@ -2833,3 +2833,195 @@ def test_credentials_match_issuer_url_shaped_dcr_id_is_not_portable():
issuer="https://as.example.com",
)
assert credentials_match_issuer(info, "https://other", "https://client.example/metadata.json") is False


@pytest.mark.anyio
async def test_handle_token_response_backfills_omitted_scope_from_request(
oauth_provider: OAuthClientProvider, mock_storage: MockTokenStorage
):
"""RFC 6749 §5.1: an omitted token-response scope means granted == requested.

The token is stored with the requested scope filled in so it remains self-describing
after a restart, when the SEP-2350 step-up union reads it but ``client_metadata.scope``
has reverted to its constructor value.
"""
oauth_provider.context.client_metadata.scope = "read admin"
response = httpx.Response(
200,
json={"access_token": "t", "token_type": "Bearer", "expires_in": 3600},
request=httpx.Request("POST", "https://auth.example.com/token"),
)
await oauth_provider._handle_token_response(response)

assert oauth_provider.context.current_tokens is not None
assert oauth_provider.context.current_tokens.scope == "read admin"
stored = await mock_storage.get_tokens()
assert stored is not None
assert stored.scope == "read admin"


@pytest.mark.anyio
async def test_handle_refresh_response_carries_prior_scope_when_response_omits_it(
oauth_provider: OAuthClientProvider, mock_storage: MockTokenStorage
):
"""RFC 6749 §6: an omitted refresh-response scope means scope is unchanged from the prior token."""
oauth_provider.context.current_tokens = OAuthToken(access_token="old", scope="read write")
response = httpx.Response(
200,
json={"access_token": "new", "token_type": "Bearer", "expires_in": 3600},
request=httpx.Request("POST", "https://auth.example.com/token"),
)
ok = await oauth_provider._handle_refresh_response(response)

assert ok is True
assert oauth_provider.context.current_tokens is not None
assert oauth_provider.context.current_tokens.access_token == "new"
assert oauth_provider.context.current_tokens.scope == "read write"
stored = await mock_storage.get_tokens()
assert stored is not None
assert stored.scope == "read write"


@pytest.mark.anyio
async def test_issuer_binding_re_evaluated_after_asm_when_prm_discovery_failed(
oauth_provider: OAuthClientProvider,
):
"""SEP-2352: on the legacy no-PRM path the binding check uses the ASM-discovered issuer.

PRM discovery fails (404) so ``auth_server_url`` stays ``None`` and the post-PRM check is
skipped; when ASM discovery then succeeds via the root well-known fallback, the discovered
metadata's issuer is compared against the stored credentials' bound issuer and a mismatch
triggers re-registration.
"""
oauth_provider.context.current_tokens = None
oauth_provider.context.token_expiry_time = None
oauth_provider._initialized = True
oauth_provider.context.client_info = OAuthClientInformationFull(
client_id="stale-client",
redirect_uris=[AnyUrl("http://localhost:3030/callback")],
issuer="https://old-as.example.com",
)

auth_flow = oauth_provider.async_auth_flow(httpx.Request("GET", "https://api.example.com/v1/mcp"))
request = await auth_flow.__anext__()
response_401 = httpx.Response(401, request=request)

# PRM discovery: path-based then root, both 404.
prm_req = await auth_flow.asend(response_401)
assert str(prm_req.url) == "https://api.example.com/.well-known/oauth-protected-resource/v1/mcp"
prm_req = await auth_flow.asend(httpx.Response(404, request=prm_req))
assert str(prm_req.url) == "https://api.example.com/.well-known/oauth-protected-resource"

# ASM discovery via root fallback (no auth_server_url) succeeds with a different issuer.
asm_req = await auth_flow.asend(httpx.Response(404, request=prm_req))
assert str(asm_req.url) == "https://api.example.com/.well-known/oauth-authorization-server"
asm_response = httpx.Response(
200,
content=(
b'{"issuer": "https://api.example.com", '
b'"authorization_endpoint": "https://api.example.com/authorize", '
b'"token_endpoint": "https://api.example.com/token", '
b'"registration_endpoint": "https://api.example.com/register"}'
),
request=asm_req,
)

# The stale bound credentials are discarded, so the next yield is a DCR request
# rather than the authorize redirect.
next_req = await auth_flow.asend(asm_response)
assert oauth_provider.context.auth_server_url is None
assert next_req.method == "POST"
assert str(next_req.url) == "https://api.example.com/register"
await auth_flow.aclose()


@pytest.mark.anyio
async def test_issuer_is_not_stamped_when_registration_falls_back_after_asm_discovery_fails(
oauth_provider: OAuthClientProvider, mock_storage: MockTokenStorage
):
"""SEP-2352: a fallback registration is not recorded as bound to an undiscovered AS.

PRM advertises a new authorization server, so the stored credentials (bound to the old
issuer) are discarded. ASM discovery for the new server then fails, so DCR falls back to
the resource-server origin's ``/register``. That registration was not derived from the new
AS's metadata, so persisting it as bound to the new AS would wedge the binding check on
later flows; instead the issuer is left unset.
"""
oauth_provider.context.current_tokens = None
oauth_provider.context.token_expiry_time = None
oauth_provider._initialized = True
oauth_provider.context.client_info = OAuthClientInformationFull(
client_id="stale-client",
redirect_uris=[AnyUrl("http://localhost:3030/callback")],
issuer="https://api.example.com/",
)

captured_state: str | None = None

async def capture_redirect(url: str) -> None:
nonlocal captured_state
captured_state = parse_qs(urlparse(url).query).get("state", [None])[0]

async def echo_callback() -> AuthorizationCodeResult:
return AuthorizationCodeResult(code="auth_code", state=captured_state)

oauth_provider.context.redirect_handler = capture_redirect
oauth_provider.context.callback_handler = echo_callback

auth_flow = oauth_provider.async_auth_flow(httpx.Request("GET", "https://api.example.com/v1/mcp"))
request = await auth_flow.__anext__()
response_401 = httpx.Response(
401,
headers={
"WWW-Authenticate": (
'Bearer resource_metadata="https://api.example.com/.well-known/oauth-protected-resource"'
)
},
request=request,
)

# PRM succeeds and advertises a new AS — the discard block fires.
prm_req = await auth_flow.asend(response_401)
assert str(prm_req.url) == "https://api.example.com/.well-known/oauth-protected-resource"
prm_response = httpx.Response(
200,
content=(
b'{"resource": "https://api.example.com/v1/mcp", "authorization_servers": ["https://new-as.example.com"]}'
),
request=prm_req,
)

# ASM discovery for the new AS fails on every well-known URL.
asm_req = await auth_flow.asend(prm_response)
assert oauth_provider.context.client_info is None
assert oauth_provider.context.oauth_metadata is None
assert str(asm_req.url) == "https://new-as.example.com/.well-known/oauth-authorization-server"
asm_req = await auth_flow.asend(httpx.Response(404, request=asm_req))
assert str(asm_req.url) == "https://new-as.example.com/.well-known/openid-configuration"

# Step 4 falls back to the resource-server origin's /register.
dcr_req = await auth_flow.asend(httpx.Response(404, request=asm_req))
assert dcr_req.method == "POST"
assert str(dcr_req.url) == "https://api.example.com/register"
dcr_response = httpx.Response(
201,
json={"client_id": "fallback-client", "redirect_uris": ["http://localhost:3030/callback"]},
request=dcr_req,
)
token_req = await auth_flow.asend(dcr_response)

# The persisted record carries no issuer binding — not the PRM-advertised AS we never reached.
stored = await mock_storage.get_client_info()
assert stored is not None
assert stored.client_id == "fallback-client"
assert stored.issuer is None

# Drive the flow to completion so the context lock is released cleanly.
token_response = httpx.Response(
200, json={"access_token": "t", "token_type": "Bearer", "expires_in": 3600}, request=token_req
)
final_req = await auth_flow.asend(token_response)
try:
await auth_flow.asend(httpx.Response(200, request=final_req))
except StopAsyncIteration:
pass
Loading