fix(mcp): return 503 on auth-backend outages instead of invalid_token

apps/mcp/src/server/auth/index.test.ts

@@ -1,6 +1,11 @@
 import { createLocalJWKSet, exportJWK, generateKeyPair, SignJWT } from "jose"
 import { afterEach, beforeAll, describe, expect, it, vi } from "vitest"
-import { fetchSession, validateApiKey, validateOAuthToken } from "./index"
+import {
+	fetchSession,
+	TransientAuthError,
+	validateApiKey,
+	validateOAuthToken,
+} from "./index"
 
 const API_URL = "https://api.example.com"
 const ISSUER = `${API_URL}/api/auth`
@@ -180,4 +185,32 @@ describe("MCP authentication", () => {
 		await expect(validateApiKey("not_a_key", API_URL)).resolves.toBeNull()
 		expect(fetchSpy).not.toHaveBeenCalled()
 	})
+
+	it("surfaces a 500 from the session endpoint as TransientAuthError, not invalid token", async () => {
+		vi.spyOn(console, "error").mockImplementation(() => {})
+		vi.stubGlobal(
+			"fetch",
+			vi.fn().mockResolvedValue(new Response(null, { status: 500 })),
+		)
+
+		await expect(
+			validateApiKey("sm_outage_key_0123456789abcdef", API_URL),
+		).rejects.toThrow(TransientAuthError)
+	})
+
+	it("surfaces a session-endpoint timeout as TransientAuthError", async () => {
+		vi.spyOn(console, "error").mockImplementation(() => {})
+		vi.stubGlobal(
+			"fetch",
+			vi.fn().mockRejectedValue(
+				Object.assign(new Error("The operation was aborted"), {
+					name: "TimeoutError",
+				}),
+			),
+		)
+
+		await expect(
+			validateApiKey("sm_timeout_key_0123456789abcd", API_URL),
+		).rejects.toThrow(TransientAuthError)
+	})
 })

apps/mcp/src/server/auth/index.ts

@@ -65,6 +65,46 @@ export function isApiKey(token: string): boolean {
 	return API_KEY_PATTERN.test(token)
 }
 
+// Upstream was unreachable, not the token being bad: reporting these as invalid_token makes clients discard working credentials.
+export class TransientAuthError extends Error {
+	readonly status?: number
+
+	constructor(message: string, status?: number) {
+		super(message)
+		this.name = "TransientAuthError"
+		this.status = status
+	}
+}
+
+const TRANSIENT_ERROR_NAMES = new Set([
+	"AbortError",
+	"TimeoutError",
+	"JWKSTimeout",
+])
+
+// ERR_JOSE_GENERIC is what jose throws when the JWKS endpoint answers non-200 or unparseable JSON.
+const TRANSIENT_JOSE_CODES = new Set(["ERR_JWKS_TIMEOUT", "ERR_JOSE_GENERIC"])
+
+function transientAuthErrorFor(error: unknown): TransientAuthError | null {
+	const status = (error as { status?: unknown } | null)?.status
+	if (typeof status === "number" && status !== 401 && status !== 403) {
+		return new TransientAuthError(`Session endpoint returned ${status}`, status)
+	}
+	if (error instanceof TypeError) {
+		return new TransientAuthError(`Auth backend unreachable: ${error.message}`)
+	}
+	if (error instanceof Error && TRANSIENT_ERROR_NAMES.has(error.name)) {
+		return new TransientAuthError(error.message)
+	}
+	const code = (error as { code?: unknown } | null)?.code
+	if (typeof code === "string" && TRANSIENT_JOSE_CODES.has(code)) {
+		return new TransientAuthError(
+			`JWKS fetch failed: ${(error as Error).message}`,
+		)
+	}
+	return null
+}
+
 export async function validateApiKey(
 	token: string,
 	apiUrl: string,
@@ -93,6 +133,8 @@ export async function validateApiKey(
 		return user
 	} catch (error) {
 		console.error("API key validation error:", error)
+		const transient = transientAuthErrorFor(error)
+		if (transient) throw transient
 		return null
 	}
 }
@@ -142,6 +184,8 @@ export async function validateOAuthToken(
 		}
 	} catch (error) {
 		console.error("OAuth token validation error:", error)
+		const transient = transientAuthErrorFor(error)
+		if (transient) throw transient
 		return null
 	}
 }

apps/mcp/src/server/index.ts

@@ -4,6 +4,7 @@ import { Hono, type Context } from "hono"
 import { cors } from "hono/cors"
 import {
 	isApiKey,
+	TransientAuthError,
 	validateApiKey,
 	validateOAuthToken,
 	type AuthUser,
@@ -47,7 +48,7 @@ app.use(
 		allowMethods: ["GET", "POST", "DELETE", "OPTIONS"],
 		// When omitted, Hono echoes Access-Control-Request-Headers. This keeps
 		// modern Mcp-Method/Mcp-Name/Mcp-Param-* routing forward-compatible.
-		exposeHeaders: ["WWW-Authenticate"],
+		exposeHeaders: ["WWW-Authenticate", "Retry-After"],
 	}),
 )
 
@@ -128,6 +129,30 @@ function authInfoFor(
 	}
 }
 
+type AuthResolution =
+	| { ok: true; user: AuthUser }
+	| { ok: false; reason: "invalid" }
+	| { ok: false; reason: "transient" }
+
+// Keeps a transient upstream failure distinct from an invalid token.
+async function resolveAuthUser(
+	token: string,
+	apiUrl: string,
+	mcpResource: string,
+): Promise<AuthResolution> {
+	try {
+		const user = isApiKey(token)
+			? await validateApiKey(token, apiUrl)
+			: await validateOAuthToken(token, apiUrl, mcpResource)
+		return user ? { ok: true, user } : { ok: false, reason: "invalid" }
+	} catch (error) {
+		if (error instanceof TransientAuthError) {
+			return { ok: false, reason: "transient" }
+		}
+		throw error
+	}
+}
+
 function unauthorizedResponse(
 	resourceMetadataUrl: string,
 	invalidToken = false,
@@ -181,10 +206,23 @@ async function handleMcpRequest(
 
 	if (!token) return unauthorizedResponse(resourceMetadataUrl)
 
-	const authUser = isApiKey(token)
-		? await validateApiKey(token, apiUrl)
-		: await validateOAuthToken(token, apiUrl, mcpResource)
-	if (!authUser) return unauthorizedResponse(resourceMetadataUrl, true)
+	const resolved = await resolveAuthUser(token, apiUrl, mcpResource)
+	if (!resolved.ok && resolved.reason === "transient") {
+		return Response.json(
+			{
+				jsonrpc: "2.0",
+				error: {
+					code: -32001,
+					message:
+						"Authentication backend temporarily unavailable, please retry",
+				},
+				id: null,
+			},
+			{ status: 503, headers: { "Retry-After": "5" } },
+		)
+	}
+	if (!resolved.ok) return unauthorizedResponse(resourceMetadataUrl, true)
+	const authUser = resolved.user
 
 	const actor: ActorContext = {
 		userId: authUser.userId,