fix: properly close SDK message on error, show tool errorText

agent/core/agent_loop.py

@@ -240,6 +240,7 @@ class Handlers:
         # Agentic loop - continue until model doesn't call tools or max iterations is reached
         iteration = 0
         final_response = None
+        errored = False
 
         while iteration < max_iterations:
             # ── Cancellation check: before LLM call ──
@@ -517,11 +518,12 @@ class Handlers:
                         data={"error": str(e) + "\n" + traceback.format_exc()},
                     )
                 )
+                errored = True
                 break
 
         if session.is_cancelled:
             await session.send_event(Event(event_type="interrupted"))
-        else:
+        elif not errored:
             await session.send_event(
                 Event(
                     event_type="turn_complete",

eval/hf_agent_connector.py

@@ -30,7 +30,7 @@ class AgentResponseGenerator:
     returns the assistant's final message.
     """
 
-    def __init__(self, config_path: str | Path, max_iterations: int = 10) -> None:
+    def __init__(self, config_path: str | Path, max_iterations: int = 300) -> None:
         self.config_path = _resolve_project_path(config_path)
         self.config: Config = load_config(str(self.config_path))
         self.max_iterations = max_iterations

frontend/src/components/Chat/ToolCallGroup.tsx

@@ -371,13 +371,14 @@ export default function ToolCallGroup({ tools, approveTools }: ToolCallGroupProp
       const displayName = toolDisplayMap[tool.toolCallId] || tool.toolName;
 
       if (tool.toolName === 'hf_jobs' && args?.script) {
-        const hasOutput = (tool.state === 'output-available' || tool.state === 'output-error') && tool.output;
+        const jobOutput = tool.output ?? (tool.state === 'output-error' ? (tool as Record<string, unknown>).errorText : undefined);
+        const hasOutput = (tool.state === 'output-available' || tool.state === 'output-error') && jobOutput;
         const scriptContent = getEditedScript(tool.toolCallId) || String(args.script);
         setPanel(
           {
             title: displayName,
             script: { content: scriptContent, language: 'python' },
-            ...(hasOutput ? { output: { content: String(tool.output), language: 'markdown' } } : {}),
+            ...(hasOutput ? { output: { content: String(jobOutput), language: 'markdown' } } : {}),
             parameters: { tool_call_id: tool.toolCallId },
           },
           hasOutput ? 'output' : 'script',
@@ -389,9 +390,10 @@ export default function ToolCallGroup({ tools, approveTools }: ToolCallGroupProp
 
       const inputSection = args ? { content: JSON.stringify(args, null, 2), language: 'json' } : undefined;
 
-      if ((tool.state === 'output-available' || tool.state === 'output-error') && tool.output) {
+      const outputText = tool.output ?? (tool.state === 'output-error' ? (tool as Record<string, unknown>).errorText : undefined);
+      if ((tool.state === 'output-available' || tool.state === 'output-error') && outputText) {
         let language = 'text';
-        const content = String(tool.output);
+        const content = String(outputText);
         if (content.trim().startsWith('{') || content.trim().startsWith('[')) language = 'json';
         else if (content.includes('```')) language = 'markdown';
 

frontend/src/lib/sse-chat-transport.ts

@@ -94,6 +94,9 @@ function createEventToChunkStream(sideChannel: SideChannelCallbacks): TransformS
           break;
 
         case 'shutdown':
+          endTextPart(controller);
+          controller.enqueue({ type: 'finish-step' });
+          controller.enqueue({ type: 'finish', finishReason: 'stop' });
           sideChannel.onShutdown();
           break;
 
@@ -243,8 +246,10 @@ function createEventToChunkStream(sideChannel: SideChannelCallbacks): TransformS
 
         case 'error': {
           const errorMsg = (event.data?.error as string) || 'Unknown error';
+          endTextPart(controller);
+          controller.enqueue({ type: 'finish-step' });
+          controller.enqueue({ type: 'finish', finishReason: 'error' });
           sideChannel.onError(errorMsg);
-          controller.enqueue({ type: 'error', errorText: errorMsg });
           sideChannel.onProcessingDone();
           break;
         }