perf: boucle fermée pHash (2s→150ms) + batch CLIP (90 appels→1)

Boucle fermée : time.sleep(2.0) remplacé par _wait_for_screen_change()
qui poll le pHash toutes les 150ms. Sort dès que l'écran change.
4 occurrences remplacées.

Batch CLIP : filtre par distance AVANT le CLIP (90→~20 éléments),
puis embed_image_batch() en un seul appel GPU + np.dot vectorisé.

Estimé : 42s→~20s total workflow.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

visual_workflow_builder/backend/api_v3/execute.py

@@ -24,6 +24,48 @@ from . import api_v3_bp
 logger = logging.getLogger(__name__)
 
 
+def _wait_for_screen_change(pre_screen, max_wait=2.0, poll_interval=0.15):
+    """
+    Attend que l'écran change après un clic, au lieu d'un sleep fixe.
+    Compare le pHash de l'écran actuel avec le screenshot pré-clic.
+    Sort dès que la distance de Hamming >= 5 (changement détecté).
+    Fallback: sort après max_wait secondes si aucun changement.
+
+    Args:
+        pre_screen: PIL Image du screenshot avant le clic (peut être None)
+        max_wait: Temps max d'attente en secondes (défaut 2.0)
+        poll_interval: Intervalle de polling en secondes (défaut 0.15)
+    """
+    if pre_screen is None:
+        # Pas de screenshot pré-clic, fallback sur sleep classique
+        time.sleep(max_wait)
+        return
+
+    try:
+        from core.analytics.screen_change_detector import compute_phash
+        import mss as _mss
+        from PIL import Image as _PILImage
+
+        _pre_hash = compute_phash(pre_screen)
+        _start = time.time()
+
+        while time.time() - _start < max_wait:
+            time.sleep(poll_interval)
+            try:
+                with _mss.mss() as _sct:
+                    _grab = _sct.grab(_sct.monitors[0])
+                    _post_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+                _post_hash = compute_phash(_post_screen)
+                _dist = _pre_hash - _post_hash
+                if _dist >= 5:  # Écran a changé
+                    break
+            except Exception:
+                break
+    except Exception:
+        # Si pHash/mss non disponible, fallback sur sleep
+        time.sleep(max_wait)
+
+
 from core.execution.input_handler import (
     safe_type_text as _shared_safe_type_text,
     check_screen_for_patterns as _shared_check_patterns,
@@ -722,6 +764,17 @@ def execute_action_with_coords(action_type: str, params: dict, coords: dict) ->
         x, y = coords['x'], coords['y']
         print(f"🖱️ [Self-Healing] Clic aux coordonnées choisies: ({x}, {y})")
 
+        # Capture pré-clic pour détection de changement
+        _pre_screen = None
+        try:
+            import mss as _mss
+            from PIL import Image as _PILImage
+            with _mss.mss() as _sct:
+                _grab = _sct.grab(_sct.monitors[0])
+                _pre_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+        except Exception:
+            pass
+
         if action_type in ['double_click_anchor']:
             pyautogui.doubleClick(x, y)
         elif action_type in ['right_click_anchor']:
@@ -729,7 +782,7 @@ def execute_action_with_coords(action_type: str, params: dict, coords: dict) ->
         else:
             pyautogui.click(x, y)
 
-        time.sleep(2.0)  # Délai après le clic
+        _wait_for_screen_change(_pre_screen)
 
         return {
             'success': True,
@@ -759,6 +812,17 @@ def execute_action_with_static_coords(action_type: str, params: dict) -> dict:
 
         print(f"🖱️ [Self-Healing] Clic aux coordonnées statiques: ({x}, {y})")
 
+        # Capture pré-clic pour détection de changement
+        _pre_screen = None
+        try:
+            import mss as _mss
+            from PIL import Image as _PILImage
+            with _mss.mss() as _sct:
+                _grab = _sct.grab(_sct.monitors[0])
+                _pre_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+        except Exception:
+            pass
+
         if action_type in ['double_click_anchor']:
             pyautogui.doubleClick(x, y)
         elif action_type in ['right_click_anchor']:
@@ -766,7 +830,7 @@ def execute_action_with_static_coords(action_type: str, params: dict) -> dict:
         else:
             pyautogui.click(x, y)
 
-        time.sleep(2.0)
+        _wait_for_screen_change(_pre_screen)
 
         return {
             'success': True,
@@ -844,6 +908,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                     _fc_target_desc = params.get('visual_anchor', {}).get('description', '')
 
                     x, y, confidence, method_used = None, None, 0, ''
+                    screen_img = None  # Screenshot pré-clic pour détection de changement
 
                     # === MÉTHODE 1 : Template matching direct (~1-10ms) ===
                     try:
@@ -922,7 +987,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                         else:
                             pyautogui.click(x, y)
 
-                        time.sleep(2.0)
+                        _wait_for_screen_change(screen_img)
 
                         return {
                             'success': True,
@@ -969,7 +1034,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                                 else:
                                     pyautogui.click(gx, gy)
 
-                                time.sleep(2.0)
+                                _wait_for_screen_change(screen_img)
 
                                 return {
                                     'success': True,

visual_workflow_builder/backend/services/intelligent_executor.py

@@ -218,6 +218,9 @@ class IntelligentExecutor:
         Matching par similarité d'embeddings CLIP + pondération par distance.
         Combine le score sémantique avec la proximité à la position originale.
 
+        Utilise embed_image_batch() pour encoder tous les éléments en un seul
+        appel GPU au lieu de ~90 appels individuels.
+
         SEUILS STRICTS pour éviter les faux positifs:
         - MAX_DISTANCE_PX: Distance maximale absolue (500px)
         - MIN_CLIP_SCORE: Score CLIP minimum (0.50)
@@ -253,31 +256,22 @@ class IntelligentExecutor:
             # Obtenir l'embedding de l'ancre
             anchor_embedding = self._clip_model.embed_image(anchor_image)
 
-            best_match = None
-            best_combined_score = 0.0
-            candidates = []
-            rejected_candidates = []  # Pour debug: garder trace des rejetés
-
             print(f"🔍 [CLIP] {len(elements)} éléments détectés par UI-DETR-1")
 
+            # === ÉTAPE 1 : Filtrer par distance et préparer les crops ===
+            nearby_elements = []       # Éléments gardés (distance OK)
+            nearby_crops = []          # Crops PIL correspondants
+            nearby_distances = []      # Distances pré-calculées
+            nearby_distance_factors = []  # Facteurs de pondération
+            rejected_candidates = []   # Pour debug: garder trace des rejetés
+
             for elem in elements:
-                # Extraire la région de l'élément
                 x1, y1 = elem.bbox['x1'], elem.bbox['y1']
                 x2, y2 = elem.bbox['x2'], elem.bbox['y2']
 
-                elem_crop = screen_image.crop((x1, y1, x2, y2))
-
-                # Obtenir l'embedding de l'élément
-                elem_embedding = self._clip_model.embed_image(elem_crop)
-
-                # Calculer la similarité cosinus (score sémantique CLIP)
-                clip_score = float(np.dot(anchor_embedding, elem_embedding) /
-                            (np.linalg.norm(anchor_embedding) * np.linalg.norm(elem_embedding)))
-
-                # Calculer la pondération par distance si position originale connue
-                distance_factor = 1.0
+                # Calculer la distance si position originale connue
                 distance = None
-                rejected_reason = None
+                distance_factor = 1.0
 
                 if anchor_center_x is not None and anchor_center_y is not None:
                     elem_center_x = (x1 + x2) // 2
@@ -287,30 +281,63 @@ class IntelligentExecutor:
                         (elem_center_y - anchor_center_y) ** 2
                     )
 
-                    # Pondération par distance
-                    normalized_distance = distance / screen_diagonal
-                    distance_factor = max(0.2, 1.0 - (normalized_distance * 5.0))
-
                     # REJET STRICT: distance > MAX_DISTANCE_PX
                     if distance > MAX_DISTANCE_PX:
-                        rejected_reason = f"distance {distance:.0f}px > {MAX_DISTANCE_PX}px"
                         rejected_candidates.append({
                             'element_id': elem.id,
-                            'clip_score': clip_score,
+                            'clip_score': 0.0,
                             'distance': distance,
-                            'reason': rejected_reason,
+                            'reason': f"distance {distance:.0f}px > {MAX_DISTANCE_PX}px",
                             'center': {'x': elem_center_x, 'y': elem_center_y}
                         })
                         continue
 
+                    # Pondération par distance
+                    normalized_distance = distance / screen_diagonal
+                    distance_factor = max(0.2, 1.0 - (normalized_distance * 5.0))
+
+                # Cropper l'élément
+                elem_crop = screen_image.crop((x1, y1, x2, y2))
+
+                nearby_elements.append(elem)
+                nearby_crops.append(elem_crop)
+                nearby_distances.append(distance)
+                nearby_distance_factors.append(distance_factor)
+
+            print(f"🔍 [CLIP] {len(nearby_elements)} éléments après filtre distance "
+                  f"({len(rejected_candidates)} rejetés par distance)")
+
+            # === ÉTAPE 2 : Batch CLIP — un seul appel GPU ===
+            best_match = None
+            best_combined_score = 0.0
+            candidates = []
+
+            if nearby_crops:
+                # Encoder tous les crops en batch (1 appel GPU au lieu de N)
+                all_embeddings = self._clip_model.embed_image_batch(nearby_crops)
+
+                # === ÉTAPE 3 : Similarités vectorisées avec numpy ===
+                # anchor_embedding shape: (dim,), all_embeddings shape: (N, dim)
+                anchor_norm = np.linalg.norm(anchor_embedding)
+                elem_norms = np.linalg.norm(all_embeddings, axis=1)
+                # Similarité cosinus vectorisée
+                clip_scores = np.dot(all_embeddings, anchor_embedding) / (elem_norms * anchor_norm)
+
+                # === ÉTAPE 4 : Appliquer seuils et construire les candidats ===
+                for i, elem in enumerate(nearby_elements):
+                    clip_score = float(clip_scores[i])
+                    distance = nearby_distances[i]
+                    distance_factor = nearby_distance_factors[i]
+
                     # REJET STRICT: score CLIP < MIN_CLIP_SCORE
                     if clip_score < MIN_CLIP_SCORE:
-                    rejected_reason = f"CLIP {clip_score:.2f} < {MIN_CLIP_SCORE}"
+                        x1, y1 = elem.bbox['x1'], elem.bbox['y1']
+                        x2, y2 = elem.bbox['x2'], elem.bbox['y2']
                         rejected_candidates.append({
                             'element_id': elem.id,
                             'clip_score': clip_score,
                             'distance': distance,
-                        'reason': rejected_reason,
+                            'reason': f"CLIP {clip_score:.2f} < {MIN_CLIP_SCORE}",
                             'center': {'x': (x1+x2)//2, 'y': (y1+y2)//2}
                         })
                         continue