improve token error handling; remove old tokens

src/lib/state/token.svelte.ts

@@ -18,7 +18,7 @@ class Token {
 		const parsed = safeParse(storedHfToken ?? "");
 		const storedToken = typia.is<string>(parsed) ? parsed : "";
 
-		if (storedToken) {
+		if (storedToken && storedToken.startsWith("hf_jwt")) {
 			this.#value = storedToken;
 		} else {
 			this.requestTokenFromParent();
@@ -36,18 +36,23 @@ class Token {
 		this.#value = token;
 	}
 
-	requestTokenFromParent = (): Promise<void> => {
-		if (typeof window === "undefined") return Promise.resolve();
+	requestTokenFromParent = (): Promise<boolean> => {
+		if (typeof window === "undefined") return Promise.resolve(false);
 
 		return new Promise(resolve => {
+			const timeout = window.setTimeout(() => {
+				window.removeEventListener("message", handleMessage);
+				resolve(false);
+			}, 5000);
+
 			const handleMessage = (event: MessageEvent) => {
 				if (event.data.type === "INFERENCE_JWT_RESPONSE") {
 					const token = event.data.token;
-					if (token && typeof token === "string") {
-						this.value = token;
-						window.removeEventListener("message", handleMessage);
-						resolve();
-					}
+					if (!token || typeof token !== "string") return resolve(false);
+					this.value = token;
+					window.removeEventListener("message", handleMessage);
+					resolve(true);
+					window.clearTimeout(timeout);
 				}
 			};
 

src/lib/utils/business.svelte.ts

@@ -32,6 +32,7 @@ import { mcpServers } from "$lib/state/mcps.svelte.js";
 import { modifySnippet } from "$lib/utils/snippets.js";
 import { models } from "$lib/state/models.svelte";
 import { StreamReader } from "$lib/utils/stream.js";
+import { dev } from "$app/environment";
 
 type ChatCompletionInputMessageChunk =
 	NonNullable<ChatCompletionInputMessage["content"]> extends string | (infer U)[] ? U : never;
@@ -118,6 +119,10 @@ function getResponseFormatObj(conversation: ConversationClass | Conversation) {
 	}
 }
 
+const tokenErrMessage = dev
+	? "Please set your Hugging Face token in the .env file"
+	: "Failed to connect to inference providers. Are you logged in?";
+
 export async function handleStreamingResponse(
 	conversation: ConversationClass | Conversation,
 	onChunk: (content: string) => void,
@@ -171,7 +176,8 @@ export async function handleStreamingResponse(
 		}
 	} catch (error) {
 		if (error instanceof Error && error.message.includes("401") && retryCount === 0) {
-			await token.requestTokenFromParent();
+			const ok = await token.requestTokenFromParent();
+			if (!ok) throw new Error(tokenErrMessage);
 			return handleStreamingResponse(conversation, onChunk, abortController, retryCount + 1);
 		}
 		throw error;
@@ -218,7 +224,8 @@ export async function handleNonStreamingResponse(
 
 	if (!response.ok) {
 		if (response.status === 401 && retryCount === 0) {
-			await token.requestTokenFromParent();
+			const ok = await token.requestTokenFromParent();
+			if (!ok) throw new Error(tokenErrMessage);
 			return handleNonStreamingResponse(conversation, retryCount + 1);
 		}
 		const error = await response.json();

src/routes/api/generate/+server.ts

@@ -144,7 +144,16 @@ export const POST: RequestHandler = async ({ request }) => {
 						}
 					} catch (error) {
 						console.error("stream error", error);
-						writer.error(error instanceof Error ? error : new Error(String(error)));
+
+						let status = 500;
+						if (error instanceof InferenceClientProviderApiError || error instanceof InferenceClientHubApiError) {
+							status = error.httpResponse.status;
+						}
+
+						const errorMessage = error instanceof Error ? error.message : String(error);
+						const errorWithStatus = new Error(status === 401 ? "401 Unauthorized" : errorMessage);
+
+						writer.error(errorWithStatus);
 						res.finish_reason = "stop";
 						return res;
 					}