fix(stream): prevent non-JSON SSE lines and duplicate [DONE] from breaking clients

- Passthrough: skip non-JSON data lines instead of forwarding raw garbage
- Translate: stop emitting redundant [DONE] sentinel (message_stop terminates)
- Add streamDoneSent flag to prevent duplicate [DONE] across transform + flush
- Warn on unexpected upstream Content-Type for streaming responses

PR #2046 by @qianze0628

Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/handlers/chatCore/streamingHandler.js

@@ -52,6 +52,14 @@ export function handleStreamingResponse({ providerResponse, provider, model, sou
       });
   }
 
+  // Warn when upstream returns unexpected Content-Type for a streaming response.
+  // This often means the provider returned an HTML error page or plain-text error
+  // that the SSE transform stream would forward as garbage to the client.
+  const upstreamContentType = (providerResponse.headers.get('content-type') || '').toLowerCase();
+  if (upstreamContentType && !upstreamContentType.includes('text/event-stream') && !upstreamContentType.includes('application/json')) {
+    console.warn('[STREAM] ' + provider + ' | ' + model + ' | unexpected Content-Type: ' + upstreamContentType);
+  }
+
   const transformStream = buildTransformStream({ provider, sourceFormat, targetFormat, userAgent, reqLogger, toolNameMap, model, connectionId, body, onStreamComplete, apiKey });
 
   // Responses passthrough: synthesize response.failed + [DONE] if the stream aborts/stalls before a terminal event

open-sse/utils/stream.js

@@ -71,6 +71,7 @@ export function createSSEStream(options = {}) {
   let currentOpenAIResponsesEvent = null;
   let openAIResponsesTerminalSeen = false;
   let openAIResponsesDoneSent = false;
+  let streamDoneSent = false;  // track duplicate [DONE] across transform + flush
 
   return new TransformStream({
     transform(chunk, controller) {
@@ -166,7 +167,12 @@ export function createSSEStream(options = {}) {
                 output = `data: ${JSON.stringify(parsed)}\n`;
                 injectedUsage = true;
               }
-            } catch { }
+            } catch {
+              // Skip non-JSON data lines silently — don't forward garbage to clients.
+              // Upstream providers sometimes return plain-text errors (HTML, rate-limit
+              // messages) in the SSE stream that would break downstream JSON decoders.
+              continue;
+            }
           }
 
           if (!injectedUsage) {
@@ -211,9 +217,10 @@ export function createSSEStream(options = {}) {
             sseEmittedCount++;
           }
 
-          const output = "data: [DONE]\n\n";
-          reqLogger?.appendConvertedChunk?.(output);
-          controller.enqueue(sharedEncoder.encode(output));
+          // [DONE] not emitted in translate mode — some clients' SSE decoders
+          // fail to parse the OpenAI sentinel on Claude-format translated streams.
+          // message_stop already signals end-of-response; stream close handles it.
+          streamDoneSent = true;
           if (keepsOpenAIResponsesFormat) openAIResponsesDoneSent = true;
           continue;
         }
@@ -343,9 +350,11 @@ export function createSSEStream(options = {}) {
           // Some clients (e.g. OpenClaw) expect the OpenAI-style sentinel:
           //   data: [DONE]\n\n
           // Without it they can hang until timeout and trigger failover.
+          if (!streamDoneSent) {
             const doneOutput = "data: [DONE]\n\n";
             reqLogger?.appendConvertedChunk?.(doneOutput);
             controller.enqueue(sharedEncoder.encode(doneOutput));
+          }
 
           if (onStreamComplete) {
             onStreamComplete({
@@ -406,11 +415,8 @@ export function createSSEStream(options = {}) {
           openAIResponsesTerminalSeen = true;
         }
 
-        if (!keepsOpenAIResponsesFormat || !openAIResponsesDoneSent) {
-          const doneOutput = "data: [DONE]\n\n";
-          reqLogger?.appendConvertedChunk?.(doneOutput);
-          controller.enqueue(sharedEncoder.encode(doneOutput));
-        }
+        // [DONE] not emitted in translate mode — see comment above.
+        // Passthrough mode still emits it for standard OpenAI clients.
 
         if (!hasValidUsage(state?.usage) && totalContentLength > 0) {
           state.usage = estimateUsage(body, totalContentLength, sourceFormat);