Surface real error messages and fix LLM exception formatting

- LLM exceptions (401, rate limit, etc.) now show clean message instead
of raw HTML or full stack trace
- Backend error events now consistently use 'message' field
- Frontend reads event.message ?? event.error so actual errors are shown
- Bail early if LLM returns empty response

backend/api/demo.py

@@ -170,10 +170,20 @@ async def execute_query_stream(req: ExecuteQueryRequest):
                         chunks.append(delta)
                         yield {"data": json.dumps({"type": "sql_chunk", "chunk": delta})}
             except Exception as e:
-                yield {"data": json.dumps({"type": "error", "error": str(e), "error_class": "other"})}
+                # Format LLM exception concisely (avoid dumping full HTML 401 pages)
+                err_str = str(e)
+                if len(err_str) > 300 or '<html' in err_str.lower():
+                    err_str = f"LLM API error: {type(e).__name__} (check HF_TOKEN / model availability)"
+                yield {"data": json.dumps({"type": "error", "message": err_str, "error_class": "other"})}
                 break
 
             generated_sql = _clean_sql("".join(chunks))
+
+            # If LLM returned nothing useful, bail early
+            if not generated_sql.strip():
+                yield {"data": json.dumps({"type": "error", "message": "LLM returned empty response", "error_class": "other"})}
+                break
+
             yield {"data": json.dumps({"type": "sql_complete", "sql": generated_sql})}
             yield {"data": json.dumps({"type": "executing"})}
 
@@ -228,7 +238,7 @@ async def execute_query_stream(req: ExecuteQueryRequest):
                 except Exception:
                     pass
 
-                yield {"data": json.dumps({"type": "error", "error": error, "error_class": error_class_name})}
+                yield {"data": json.dumps({"type": "error", "message": error, "error_class": error_class_name})}
 
             # Grader + RL update
             grader_in = GraderInput(

frontend/src/components/ChatPanel.tsx

@@ -451,7 +451,7 @@ export function ChatPanel() {
           } else if (event.type === 'error') {
             updateMessage(msgId, {
               status: 'error',
-              errorMsg: event.message as string,
+              errorMsg: ((event.message ?? event.error) as string | undefined) ?? 'Agent exhausted all repair attempts',
             })
           } else if (event.type === 'gepa_start') {
             setOptimizingBanner(true)