fix: prevent server crash on unauthenticated requests (ERR_HTTP_HEADERS_SENT)

.github/workflows/deploy.yml

@@ -7,17 +7,17 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Login to Registry
-        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121  # v4.1.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Docker metadata
         id: meta
-        uses: docker/metadata-action@030e881283bb7a6894de51c315a6bfe6a94e05cf  # v6.0.0
+        uses: docker/metadata-action@030e881283bb7a6894de51c315a6bfe6a94e05cf # v6.0.0
         with:
           images: |
             huggingface/responses-js
@@ -26,13 +26,13 @@ jobs:
             type=sha,enable=true,prefix=sha-,format=short,sha-len=8
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd  # v4.0.0
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4.0.0
 
       - name: Inject slug/short variables
-        uses: rlespinasse/github-slug-action@e6f261660910b273384c5c42b17a0217881b217a  # v5.6.0
+        uses: rlespinasse/github-slug-action@e6f261660910b273384c5c42b17a0217881b217a # v5.6.0
 
       - name: Build and Publish image
-        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f  # v7.1.0
+        uses: docker/build-push-action@bcafcacb16a39f128d818304e6c9c0c18556b85f # v7.1.0
         with:
           context: .
           file: Dockerfile
@@ -50,7 +50,7 @@ jobs:
     needs: ["build-and-publish"]
     steps:
       - name: Inject slug/short variables
-        uses: rlespinasse/github-slug-action@e6f261660910b273384c5c42b17a0217881b217a  # v5.6.0
+        uses: rlespinasse/github-slug-action@e6f261660910b273384c5c42b17a0217881b217a # v5.6.0
 
       - name: Gen values
         run: |
@@ -62,7 +62,7 @@ jobs:
           echo "VALUES=$(echo "$VALUES" | yq -o=json | jq tostring)" >> $GITHUB_ENV
 
       - name: Deploy on infra-deployments
-        uses: aurelien-baudet/workflow-dispatch@3133c5d135c7dbe4be4f9793872b6ef331b53bc7  # v4.0.0
+        uses: aurelien-baudet/workflow-dispatch@3133c5d135c7dbe4be4f9793872b6ef331b53bc7 # v4.0.0
         with:
           workflow: Update application single value
           repo: huggingface/infra-deployments

.github/workflows/lint.yml

@@ -12,15 +12,15 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
 
       - name: Setup Node.js
-        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
+        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
         with:
-          node-version: "18"
+          node-version: "22"
 
       - name: Setup pnpm
-        uses: pnpm/action-setup@0e279bb959325dab635dd2c09392533439d90093  # v6.0.8
+        uses: pnpm/action-setup@0e279bb959325dab635dd2c09392533439d90093 # v6.0.8
         with:
           version: 10.10.0
 
@@ -30,7 +30,7 @@ jobs:
           echo "STORE_PATH=$(pnpm store path --silent)" >> $GITHUB_ENV
 
       - name: Setup pnpm cache
-        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
+        uses: actions/cache@27d5ce7f107fe9357f9df03efb73ab90386fccae # v5.0.5
         with:
           path: ${{ env.STORE_PATH }}
           key: ${{ runner.os }}-pnpm-store-${{ hashFiles('**/pnpm-lock.yaml') }}

src/routes/responses.ts

@@ -56,30 +56,66 @@ export const postCreateResponse = async (
 	req: ValidatedRequest<CreateResponseParams>,
 	res: ExpressResponse
 ): Promise<void> => {
-	// To avoid duplicated code, we run all requests as stream.
-	const events = runCreateResponseStream(req, res);
-
-	// Then we return in the correct format depending on the user 'stream' flag.
-	if (req.body.stream) {
-		res.setHeader("Content-Type", "text/event-stream");
-		res.setHeader("Connection", "keep-alive");
-		console.debug("Stream request");
-		for await (const event of events) {
-			console.debug(`Event #${event.sequence_number}: ${event.type}`);
-			res.write(`data: ${JSON.stringify(event)}\n\n`);
-		}
-		res.end();
-	} else {
-		console.debug("Non-stream request");
-		for await (const event of events) {
-			if (event.type === "response.completed" || event.type === "response.failed") {
-				console.debug(event.type);
-				res.json(event.response);
+	// Auth is checked before any event is produced: once the stream has started
+	// writing, it is too late to send an error status.
+	const apiKey = req.headers.authorization?.split(" ")[1];
+	if (!apiKey) {
+		res.status(401).json({
+			success: false,
+			error: "Unauthorized",
+		});
+		return;
+	}
+
+	// Express 4 does not catch errors from async handlers: anything thrown here
+	// becomes an unhandled rejection and kills the process, so the whole handler
+	// is wrapped.
+	try {
+		// To avoid duplicated code, we run all requests as stream.
+		const events = runCreateResponseStream(req, apiKey);
+
+		// Then we return in the correct format depending on the user 'stream' flag.
+		if (req.body.stream) {
+			res.setHeader("Content-Type", "text/event-stream");
+			res.setHeader("Connection", "keep-alive");
+			console.debug("Stream request");
+			for await (const event of events) {
+				console.debug(`Event #${event.sequence_number}: ${event.type}`);
+				res.write(`data: ${JSON.stringify(event)}\n\n`);
+			}
+			res.end();
+		} else {
+			console.debug("Non-stream request");
+			for await (const event of events) {
+				if (event.type === "response.completed" || event.type === "response.failed") {
+					console.debug(event.type);
+					res.json(event.response);
+					break;
+				}
 			}
 		}
+	} catch (error) {
+		console.error("Error in postCreateResponse:", error);
+		if (!res.headersSent) {
+			res.status(500).json({
+				success: false,
+				error: errorMessage(error, "Internal server error"),
+			});
+		} else {
+			res.end();
+		}
 	}
 };
 
+function errorMessage(error: unknown, fallback: string): string {
+	return typeof error === "object" &&
+		error &&
+		"message" in error &&
+		typeof (error as { message: unknown }).message === "string"
+		? (error as { message: string }).message
+		: fallback;
+}
+
 /*
  * Top-level stream.
  *
@@ -88,7 +124,7 @@ export const postCreateResponse = async (
  */
 async function* runCreateResponseStream(
 	req: ValidatedRequest<CreateResponseParams>,
-	res: ExpressResponse
+	apiKey: string
 ): AsyncGenerator<PatchedResponseStreamEvent> {
 	let sequenceNumber = 0;
 	// Prepare response object that will be iteratively populated
@@ -134,25 +170,17 @@ async function* runCreateResponseStream(
 
 	// Any events (LLM call, MCP call, list tools, etc.)
 	try {
-		for await (const event of innerRunStream(req, res, responseObject)) {
+		for await (const event of innerRunStream(req, apiKey, responseObject)) {
 			yield { ...event, sequence_number: sequenceNumber++ };
 		}
 	} catch (error) {
 		// Error event => stop
 		console.error("Error in stream:", error);
 
-		const message =
-			typeof error === "object" &&
-			error &&
-			"message" in error &&
-			typeof (error as { message: unknown }).message === "string"
-				? (error as { message: string }).message
-				: "An error occurred in stream";
-
 		responseObject.status = "failed";
 		responseObject.error = {
 			code: "server_error",
-			message,
+			message: errorMessage(error, "An error occurred in stream"),
 		};
 		yield {
 			type: "response.failed",
@@ -173,20 +201,10 @@ async function* runCreateResponseStream(
 
 async function* innerRunStream(
 	req: ValidatedRequest<CreateResponseParams>,
-	res: ExpressResponse,
+	apiKey: string,
 	responseObject: IncompleteResponse
 ): AsyncGenerator<PatchedResponseStreamEvent> {
-	// Retrieve API key from headers
-	const apiKey = req.headers.authorization?.split(" ")[1];
-	if (!apiKey) {
-		res.status(401).json({
-			success: false,
-			error: "Unauthorized",
-		});
-		return;
-	}
-
-	// Forward headers (except authorization handled separately)
+	// Forward headers (the exclusion set covers authorization)
 	const defaultHeaders = Object.fromEntries(
 		Object.entries(req.headers).filter(([key]) => !NOT_FORWARDED_HEADERS.has(key.toLowerCase()))
 	) as Record<string, string>;
@@ -518,7 +536,7 @@ async function* listMcpToolsStream(
  * Call LLM and stream the response.
  */
 async function* handleOneTurnStream(
-	apiKey: string | undefined,
+	apiKey: string,
 	payload: ChatCompletionCreateParamsStreaming,
 	responseObject: IncompleteResponse,
 	mcpToolsMapping: Record<string, McpServerParams>,

tests/responses.test.js

@@ -778,6 +778,18 @@ describe("responses.js", function () {
 		assert.equal(output[1].content[0].type, "output_text");
 		assert.equal(typeof output[1].content[0].text, "string");
 	});
+
+	it("unauthenticated request gets a 401 and does not crash the server", async function () {
+		const response = await fetch("http://localhost:3000/v1/responses", {
+			method: "POST",
+			headers: { "Content-Type": "application/json" },
+			body: JSON.stringify({ model: "Qwen/Qwen2.5-VL-7B-Instruct", input: "hi" }),
+		});
+		assert.equal(response.status, 401);
+
+		const health = await fetch("http://localhost:3000/health");
+		assert.equal(health.status, 200);
+	});
 });
 
 function dropConsecutiveEvents(events) {