fix LLM stream

Author: toubatbrian

agents/src/llm/fallback_adapter.ts

@@ -209,6 +209,11 @@ class FallbackLLMStream extends LLMStream {
       extraKwargs: this.extraKwargs,
     });
 
+    // Suppress unhandled 'error' events from the child LLM — error detection
+    // is done via stream._runError to avoid cross-request contamination.
+    const noop = () => {};
+    llm.on('error', noop);
+
     try {
       let shouldSetCurrent = !checkRecovery;
       for await (const chunk of stream) {
@@ -218,6 +223,10 @@ class FallbackLLMStream extends LLMStream {
         }
         yield chunk;
       }
+
+      if (stream._runError) {
+        throw stream._runError;
+      }
     } catch (error) {
       if (error instanceof APIError) {
         if (checkRecovery) {
@@ -245,6 +254,8 @@ class FallbackLLMStream extends LLMStream {
         this._log.error({ llm: llm.label(), error }, 'unexpected error, switching to next LLM');
       }
       throw error;
+    } finally {
+      llm.off('error', noop);
     }
   }
 

agents/src/llm/llm.ts

@@ -115,6 +115,8 @@ export abstract class LLMStream implements AsyncIterableIterator<ChatChunk> {
   protected abortController = new AbortController();
   protected _connOptions: APIConnectOptions;
   protected logger = log();
+  /** @internal – Captured when run() fails so callers can inspect after iteration ends. */
+  _runError?: Error;
 
   #llm: LLM;
   #chatCtx: ChatContext;
@@ -148,7 +150,13 @@ export abstract class LLMStream implements AsyncIterableIterator<ChatChunk> {
     // is run **after** the constructor has finished. Otherwise we get
     // runtime error when trying to access class variables in the
     // `run` method.
-    startSoon(() => this.mainTask().finally(() => this.queue.close()));
+    startSoon(() =>
+      this.mainTask()
+        .catch((e) => {
+          this._runError = e instanceof Error ? e : new Error(String(e));
+        })
+        .finally(() => this.queue.close()),
+    );
   }
 
   private _mainTaskImpl = async (span: Span) => {