Surface GEPA optimization: prompt history, live banner, smart retry

Backend:
- Add /api/prompt-history endpoint (pareto front + query count + nextOptimizeAt)
- Emit gepa_start/gepa_done SSE events so banner shows during optimization
- Accept previousSql in execute-query to tell LLM to try a different approach

Frontend:
- Retry after marking "wrong" auto-executes with previous SQL as context
(button changes label to "Retry differently")
- gepa_done event refreshes PromptEvolution in right sidebar immediately
- PromptEvolution shows query count + progress bar toward next optimization
- PromptEvolution polls /api/prompt-history every 30s

backend/api/demo.py

@@ -119,11 +119,39 @@ async def connect_db(req: ConnectDbRequest):
     return {"success": False, "message": message, "tables": [], "dbLabel": get_active_db_label()}
 
 
+# ─── /api/prompt-history ─────────────────────────────────────────
+
+@router.get("/prompt-history")
+async def get_prompt_history():
+    import datetime
+    gepa = get_gepa()
+    pareto = gepa.get_pareto_front()
+    history = [
+        {
+            "generation": c.generation,
+            "prompt": c.system_prompt,
+            "score": c.score,
+            "summary": c.feedback[0][:200] if c.feedback else "Seed prompt",
+            "timestamp": datetime.datetime.utcnow().strftime("%Y-%m-%d"),
+        }
+        for c in sorted(pareto, key=lambda x: x.generation)
+    ]
+    return {
+        "prompt": gepa.get_current_prompt(),
+        "generation": gepa.current_generation,
+        "history": history,
+        "queryCount": len(gepa.get_history()),
+        "nextOptimizeAt": (len(gepa.get_history()) // 4 + 1) * 4,
+    }
+
+
 # ─── /api/execute-query ───────────────────────────────────────────
 
 class ExecuteQueryRequest(BaseModel):
     question: str
     task_id: str = "simple_queries"
+    previousSql: Optional[str] = None   # SQL from a prior attempt user marked wrong
+    previousFeedback: Optional[str] = None  # "wrong" context message
 
 
 @router.post("/execute-query")
@@ -160,8 +188,17 @@ async def execute_query_stream(req: ExecuteQueryRequest):
 
             if attempt == 1 or ep.current_sql is None:
                 system_prompt = BASE_SYSTEM_PROMPT
+                # Include previous wrong SQL if user retried after marking wrong
+                prev_context = ""
+                if req.previousSql:
+                    prev_context = (
+                        f"\nNOTE: A previous attempt generated the following SQL which was marked INCORRECT:\n"
+                        f"```sql\n{req.previousSql}\n```\n"
+                        f"You MUST try a completely different approach.\n"
+                    )
                 user_msg = (
-                    f"Schema:\n{obs.schema_info}\n\nQuestion: {req.question}\n\n"
+                    f"Schema:\n{obs.schema_info}\n\nQuestion: {req.question}\n"
+                    f"{prev_context}\n"
                     "Write a SQL query to answer this question."
                 )
             else:
@@ -376,12 +413,19 @@ async def execute_query_stream(req: ExecuteQueryRequest):
             env._episode.done = True
             env._episode.success = success
 
-        # Trigger GEPA if needed
+        # Trigger GEPA if needed — emit events so frontend shows banner
         if gepa.should_optimize():
+            yield {"data": json.dumps({"type": "gepa_start"})}
             try:
-                await gepa.run_optimization_cycle()
-            except Exception:
-                pass
+                gepa_result = await gepa.run_optimization_cycle()
+                yield {"data": json.dumps({
+                    "type": "gepa_done",
+                    "generation": gepa.current_generation,
+                    "reflection": gepa_result.get("reflection", "")[:300] if gepa_result else "",
+                })}
+            except Exception as e:
+                logger.error("GEPA optimization failed: %s", e)
+                yield {"data": json.dumps({"type": "gepa_done", "generation": gepa.current_generation, "reflection": ""})}
 
     return EventSourceResponse(event_generator())
 

frontend/src/components/ChatPanel.tsx

@@ -5,7 +5,7 @@ import {
   Loader2, MessageSquare, Zap, RefreshCw, Trash2,
 } from 'lucide-react'
 import { useStore } from '../store/useStore'
-import { streamExecuteQuery, submitFeedback } from '../lib/api'
+import { streamExecuteQuery, submitFeedback, fetchPromptHistory } from '../lib/api'
 import { ResultsTable } from './ResultsTable'
 import type { ChatMessage, AttemptStep } from '../lib/types'
 
@@ -202,7 +202,7 @@ function MessageCard({
 }: {
   msg: ChatMessage
   onFeedback: (id: string, correct: boolean) => Promise<void>
-  onRetry: (q: string) => void
+  onRetry: (q: string, previousSql?: string) => void
 }) {
   const [sqlOpen, setSqlOpen] = useState(true)
 
@@ -343,11 +343,14 @@ function MessageCard({
             )}
             {(msg.status === 'done' || msg.status === 'error') && (
               <button
-                onClick={() => onRetry(msg.question)}
+                onClick={() => onRetry(
+                  msg.question,
+                  msg.feedback === 'wrong' ? msg.sql : undefined
+                )}
                 className="ml-auto flex items-center gap-1 text-[10px] text-gray-600 hover:text-gray-400 transition-colors"
               >
                 <RefreshCw size={10} />
-                Retry
+                {msg.feedback === 'wrong' ? 'Retry differently' : 'Retry'}
               </button>
             )}
           </div>
@@ -392,7 +395,7 @@ export function ChatPanel() {
   )
 
   const execute = useCallback(
-    async (question: string) => {
+    async (question: string, previousSql?: string) => {
       if (!question.trim() || isExecuting) return
       setIsExecuting(true)
 
@@ -412,7 +415,7 @@ export function ChatPanel() {
       addMessage(newMsg)
 
       try {
-        for await (const event of streamExecuteQuery(question, taskId)) {
+        for await (const event of streamExecuteQuery(question, taskId, previousSql)) {
           if (event.type === 'sql') {
             updateMessage(msgId, { sql: event.sql as string })
           } else if (event.type === 'sql_chunk') {
@@ -457,6 +460,10 @@ export function ChatPanel() {
             setOptimizingBanner(true)
           } else if (event.type === 'gepa_done') {
             setOptimizingBanner(false)
+            // Refresh prompt history in right sidebar
+            fetchPromptHistory()
+              .then((data) => useStore.getState().setPromptData(data))
+              .catch(() => { /* noop */ })
           }
         }
       } catch (err) {
@@ -524,7 +531,7 @@ export function ChatPanel() {
                 key={msg.id}
                 msg={msg}
                 onFeedback={handleFeedback}
-                onRetry={(q) => { setInput(q); inputRef.current?.focus() }}
+                onRetry={(q, prevSql) => void execute(q, prevSql)}
               />
             ))}
             <div ref={bottomRef} />

frontend/src/components/PromptEvolution.tsx

@@ -21,11 +21,18 @@ export function PromptEvolution() {
   const prompt = currentPrompt || SEED_PROMPT
   const generation = promptGeneration
 
+  const [queryCount, setQueryCount] = useState(0)
+  const [nextAt, setNextAt] = useState(4)
+
   const loadHistory = async () => {
     setLoading(true)
     try {
       const data = await fetchPromptHistory()
       setPromptData(data)
+      if ((data as Record<string, unknown>).queryCount !== undefined) {
+        setQueryCount((data as Record<string, unknown>).queryCount as number)
+        setNextAt((data as Record<string, unknown>).nextOptimizeAt as number)
+      }
     } catch {
       // noop
     } finally {
@@ -35,6 +42,9 @@ export function PromptEvolution() {
 
   useEffect(() => {
     void loadHistory()
+    // Poll for updates every 30s
+    const interval = setInterval(() => void loadHistory(), 30000)
+    return () => clearInterval(interval)
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [])
 
@@ -65,6 +75,22 @@ export function PromptEvolution() {
         )}
       </button>
 
+      {/* Progress toward next optimization */}
+      {queryCount > 0 && (
+        <div className="flex flex-col gap-1">
+          <div className="flex items-center justify-between text-[9px] text-gray-600">
+            <span>{queryCount} queries processed</span>
+            <span>{nextAt - queryCount} until next optimization</span>
+          </div>
+          <div className="h-1 bg-white/5 rounded-full overflow-hidden">
+            <div
+              className="h-full rounded-full bg-violet-500/50 transition-all duration-500"
+              style={{ width: `${((queryCount % 4) / 4) * 100}%` }}
+            />
+          </div>
+        </div>
+      )}
+
       <AnimatePresence>
         {expanded && (
           <motion.div

frontend/src/lib/api.ts

@@ -29,12 +29,15 @@ async function* parseSSE(response: Response): AsyncGenerator<SSEEvent> {
 
 export async function* streamExecuteQuery(
   question: string,
-  taskId: string
+  taskId: string,
+  previousSql?: string,
 ): AsyncGenerator<SSEEvent> {
+  const body: Record<string, unknown> = { question, task_id: taskId }
+  if (previousSql) body.previousSql = previousSql
   const res = await fetch(`${BASE_URL}/api/execute-query`, {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({ question, task_id: taskId }),
+    body: JSON.stringify(body),
   })
   if (!res.ok) {
     throw new Error(`HTTP ${res.status}: ${res.statusText}`)