feat(vwb): câblage 19 blocs, OCR réel, screenshots ancres, configs déploiement

Dispatch execute_action élargi de 12 à 19 blocs opérationnels :
- 4 blocs souris (hover, drag_drop, scroll, focus) avec pyautogui
- extract_text via Ollama VLM (remplace stub hardcodé)
- 5 blocs ai_* redirigés vers execute_ai_analyze avec prompts adaptés
- screenshot_evidence (capture + sauvegarde PNG)
- verify_element_exists (détection visuelle CLIP)

Import workflows Léa enrichi :
- Bridge extrait anchor_image_base64 des edges
- Import crée VisualAnchor en DB + fichiers thumbnail sur disque
- PropertiesPanel affiche automatiquement les screenshots

Frontend :
- visual_condition et loop_visual masqués (hidden: true)
- Filtre dans ToolPalette pour exclure les blocs cachés

Déploiement :
- 2 configs agent (TIM Pauline + Dev Windows) avec machine_id unique
- 2 workflows démo dans la BDD (batch factures + extraction IA)

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

deploy/configs/config_dev_windows.txt

@@ -0,0 +1,19 @@
+# ============================================================
+#  Configuration Lea — Poste Dev / Chef de projet (Windows)
+# ============================================================
+#
+#  Poste : PC dev chef de projet
+#  Objectif : enrichir connaissance Windows, evaluer robustesse
+#  Serveur : 192.168.1.40:5005 (RTX 5070)
+#
+# ============================================================
+
+RPA_SERVER_URL=http://192.168.1.40:5005/api/v1
+RPA_API_TOKEN=86031addb338e449fccdb1a983f61807aec15d42d482b9c7748ad607dc23caab
+RPA_MACHINE_ID=DEV_WINDOWS
+RPA_USER_LABEL=Dev
+
+# --- Parametres avances (ne pas modifier sauf indication) ---
+# RPA_OLLAMA_HOST=localhost
+RPA_BLUR_SENSITIVE=false
+RPA_LOG_RETENTION_DAYS=180

deploy/configs/config_tim_pauline.txt

@@ -0,0 +1,19 @@
+# ============================================================
+#  Configuration Lea — Poste TIM Pauline (LAN Anoust)
+# ============================================================
+#
+#  Poste : PC de Pauline (TIM urgences)
+#  Objectif : apprentissage outil metier (DPI OSIRIS)
+#  Serveur : 192.168.1.40:5005 (RTX 5070)
+#
+# ============================================================
+
+RPA_SERVER_URL=http://192.168.1.40:5005/api/v1
+RPA_API_TOKEN=86031addb338e449fccdb1a983f61807aec15d42d482b9c7748ad607dc23caab
+RPA_MACHINE_ID=TIM_PAULINE
+RPA_USER_LABEL=Pauline
+
+# --- Parametres avances (ne pas modifier sauf indication) ---
+# RPA_OLLAMA_HOST=localhost
+RPA_BLUR_SENSITIVE=true
+RPA_LOG_RETENTION_DAYS=180

visual_workflow_builder/backend/actions/vision_ui/extract_text.py

@@ -12,6 +12,7 @@ from datetime import datetime
 import time
 import traceback
 import re
+import os
 
 from ..base_action import BaseVWBAction, VWBActionResult, VWBActionStatus
 from ...contracts.error import VWBActionError, VWBErrorType, VWBErrorSeverity, create_vwb_error
@@ -435,14 +436,48 @@ class VWBExtractTextAction(BaseVWBAction):
         return None
     
     def _find_visual_element(self, screenshot, visual_anchor, threshold):
-        """Simulation de recherche d'élément visuel."""
+        """Recherche d'élément visuel via template matching."""
-        import random
+        try:
-        confidence = random.uniform(0.6, 0.95)
+            from ...catalog_routes import find_visual_anchor_on_screen
-        
+
-        if confidence >= threshold:
+            image_ancre = None
-            return True, {'x': 300, 'y': 200, 'width': 250, 'height': 80}, confidence
+            bounding_box = None
-        else:
+
-            return False, {}, confidence
+            if isinstance(visual_anchor, VWBVisualAnchor):
+                image_ancre = visual_anchor.screenshot_base64
+                if visual_anchor.has_bounding_box():
+                    bounding_box = visual_anchor.bounding_box
+            elif isinstance(visual_anchor, dict):
+                image_ancre = visual_anchor.get('screenshot') or visual_anchor.get('image_base64')
+                bounding_box = visual_anchor.get('bounding_box')
+
+            if image_ancre:
+                resultat = find_visual_anchor_on_screen(
+                    anchor_image_base64=image_ancre,
+                    confidence_threshold=threshold,
+                    bounding_box=bounding_box
+                )
+                if resultat and resultat.get('found'):
+                    coords = {
+                        'x': resultat.get('x', resultat.get('center_x', 0)),
+                        'y': resultat.get('y', resultat.get('center_y', 0)),
+                        'width': resultat.get('width', 200),
+                        'height': resultat.get('height', 80)
+                    }
+                    return True, coords, resultat.get('confidence', 0.9)
+
+            if bounding_box:
+                return True, bounding_box, 0.7
+
+            return False, {}, 0.0
+
+        except ImportError:
+            if hasattr(visual_anchor, 'bounding_box') and visual_anchor.bounding_box:
+                return True, visual_anchor.bounding_box, 0.7
+            return False, {}, 0.0
+        except Exception as e:
+            print(f"⚠️ Erreur recherche visuelle: {e}")
+            return False, {}, 0.0
     
     def _encode_screenshot(self, screenshot_data) -> str:
         """Encode un screenshot en base64."""
@@ -485,21 +520,28 @@ class VWBExtractTextAction(BaseVWBAction):
         }
     
     def _extract_image_region(self, screenshot_data, coords: Dict[str, int]):
-        """
+        """Extrait une région spécifique de l'image."""
-        Extrait une région spécifique de l'image.
-        
-        Args:
-            screenshot_data: Données de l'image complète
-            coords: Coordonnées de la région
-        
-        Returns:
-            Image de la région ou None
-        """
         try:
-            # Ici, on utiliserait PIL ou OpenCV pour extraire la région
+            from PIL import Image
-            # Pour la simulation, on retourne un objet factice
+            import numpy as np
-            print(f"✂️ Extraction région {coords['width']}x{coords['height']}")
+
-            return {"width": coords['width'], "height": coords['height'], "data": "simulated"}
+            x = int(coords.get('x', 0))
+            y = int(coords.get('y', 0))
+            w = int(coords.get('width', 100))
+            h = int(coords.get('height', 100))
+
+            if isinstance(screenshot_data, np.ndarray):
+                pil_image = Image.fromarray(screenshot_data)
+            elif isinstance(screenshot_data, Image.Image):
+                pil_image = screenshot_data
+            else:
+                print(f"⚠️ Type screenshot non supporté: {type(screenshot_data)}")
+                return None
+
+            cropped = pil_image.crop((x, y, x + w, y + h))
+            print(f"✂️ Extraction région {w}x{h}")
+            return cropped
+
         except Exception as e:
             print(f"❌ Erreur extraction région: {e}")
             return None
@@ -533,44 +575,77 @@ class VWBExtractTextAction(BaseVWBAction):
             return image_data
     
     def _perform_ocr_extraction(self, image_data) -> tuple[str, float, Dict[str, Any]]:
-        """
+        """Effectue l'extraction OCR via Ollama VLM."""
-        Effectue l'extraction OCR sur l'image.
-        
-        Args:
-            image_data: Image prétraitée
-        
-        Returns:
-            Tuple (texte, confiance, structure)
-        """
         try:
-            # Simulation d'extraction OCR
+            import requests
-            # En réalité, on utiliserait pytesseract ou une API OCR
+            import json
-            
+            import io
-            if self.extraction_mode == 'full':
+            import base64
-                extracted_text = "Texte exemple extrait par OCR\nLigne 2 du texte\nDernière ligne"
+            from PIL import Image
-            elif self.extraction_mode == 'numbers':
+
-                extracted_text = "123456 789 2026"
+            if isinstance(image_data, Image.Image):
-            elif self.extraction_mode == 'words':
+                buffer = io.BytesIO()
-                extracted_text = "mot1 mot2 mot3 mot4"
+                image_data.save(buffer, format='PNG')
-            elif self.extraction_mode == 'lines':
+                image_base64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
-                extracted_text = "Ligne 1\nLigne 2\nLigne 3"
+            elif isinstance(image_data, dict):
+                return "", 0.0, {}
             else:
-                extracted_text = "Texte personnalisé"
+                return "", 0.0, {}
-            
+
-            # Confiance simulée
+            prompt_map = {
-            confidence = 0.85
+                'full': "Extrais TOUT le texte visible dans cette image. Retourne uniquement le texte brut, sans commentaire.",
-            
+                'numbers': "Extrais uniquement les nombres et chiffres visibles. Retourne-les séparés par des espaces.",
-            # Structure simulée
+                'lines': "Extrais tout le texte visible ligne par ligne.",
-            structure = {
+                'words': "Extrais tous les mots visibles, séparés par des espaces.",
-                "lines": extracted_text.split('\n') if '\n' in extracted_text else [extracted_text],
-                "words": extracted_text.split(),
-                "characters": len(extracted_text),
-                "language_detected": self.ocr_language
             }
-            
+            prompt = prompt_map.get(self.extraction_mode, prompt_map['full'])
-            print(f"🔤 OCR terminé - Confiance: {confidence:.3f}")
+
-            return extracted_text, confidence, structure
+            ollama_url = os.environ.get("OLLAMA_URL", "http://localhost:11434")
-            
+            model = os.environ.get("RPA_VLM_MODEL", os.environ.get("VLM_MODEL", "gemma4:e4b"))
+
+            if 'qwen' in model.lower() and not prompt.startswith('/no_think'):
+                prompt = f"/no_think\n{prompt}"
+
+            print(f"🔤 OCR VLM avec {model} (mode: {self.extraction_mode})...")
+
+            payload = {
+                "model": model,
+                "prompt": prompt,
+                "images": [image_base64],
+                "stream": False,
+                "options": {"temperature": 0.1, "num_predict": 4000}
+            }
+
+            response = requests.post(
+                f"{ollama_url}/api/generate",
+                json=payload,
+                timeout=60
+            )
+
+            if response.status_code == 200:
+                result = response.json()
+                extracted_text = result.get('response', '').strip()
+                if not extracted_text and result.get('thinking'):
+                    extracted_text = result.get('thinking', '').strip()
+
+                confidence = 0.85 if extracted_text else 0.0
+
+                structure = {
+                    "lines": extracted_text.split('\n') if '\n' in extracted_text else [extracted_text],
+                    "words": extracted_text.split(),
+                    "characters": len(extracted_text),
+                    "language_detected": self.ocr_language
+                }
+
+                print(f"✅ OCR terminé - {len(extracted_text)} caractères")
+                return extracted_text, confidence, structure
+            else:
+                print(f"⚠️ Erreur Ollama: {response.status_code}")
+                return "", 0.0, {}
+
+        except requests.exceptions.ConnectionError:
+            print("⚠️ Ollama non accessible pour OCR")
+            return "", 0.0, {}
         except Exception as e:
             print(f"❌ Erreur OCR: {e}")
             return "", 0.0, {}

visual_workflow_builder/backend/actions/vision_ui/focus_anchor.py

@@ -198,23 +198,70 @@ class VWBFocusAnchorAction(BaseVWBAction):
             
             for attempt in range(self.max_attempts):
                 print(f"   Tentative {attempt + 1}/{self.max_attempts}")
-                
+
-                # Simulation de recherche d'ancre (à remplacer par vraie implémentation)
+                try:
-                import random
+                    from ...catalog_routes import find_visual_anchor_on_screen
-                confidence = random.uniform(0.6, 0.95)
+
-                
+                    image_ancre = None
-                if confidence >= self.confidence_threshold:
+                    bounding_box = None
-                    # Ancre trouvée
+                    if isinstance(self.visual_anchor, VWBVisualAnchor):
-                    match_found = True
+                        image_ancre = self.visual_anchor.screenshot_base64
-                    best_match = {
+                        if self.visual_anchor.has_bounding_box():
-                        'confidence': confidence,
+                            bounding_box = self.visual_anchor.bounding_box
-                        'bbox': {'x': 400, 'y': 300, 'width': 120, 'height': 30},
+                    elif isinstance(self.visual_anchor, dict):
-                        'center': {'x': 460, 'y': 315}
+                        image_ancre = self.visual_anchor.get('screenshot') or self.visual_anchor.get('image_base64')
-                    }
+                        bounding_box = self.visual_anchor.get('bounding_box')
-                    break
+
-                
+                    if image_ancre:
+                        resultat = find_visual_anchor_on_screen(
+                            anchor_image_base64=image_ancre,
+                            confidence_threshold=self.confidence_threshold,
+                            bounding_box=bounding_box
+                        )
+                        if resultat and resultat.get('found'):
+                            confidence = resultat.get('confidence', 0.9)
+                            cx = resultat.get('center_x', resultat.get('x', 460))
+                            cy = resultat.get('center_y', resultat.get('y', 315))
+                            match_found = True
+                            best_match = {
+                                'confidence': confidence,
+                                'bbox': {
+                                    'x': resultat.get('x', cx - 60),
+                                    'y': resultat.get('y', cy - 15),
+                                    'width': resultat.get('width', 120),
+                                    'height': resultat.get('height', 30)
+                                },
+                                'center': {'x': cx, 'y': cy}
+                            }
+                            break
+
+                    if bounding_box:
+                        match_found = True
+                        bx = bounding_box.get('x', 0)
+                        by = bounding_box.get('y', 0)
+                        bw = bounding_box.get('width', 120)
+                        bh = bounding_box.get('height', 30)
+                        best_match = {
+                            'confidence': 0.7,
+                            'bbox': bounding_box,
+                            'center': {'x': bx + bw // 2, 'y': by + bh // 2}
+                        }
+                        break
+
+                except ImportError:
+                    if hasattr(self.visual_anchor, 'bounding_box') and self.visual_anchor.bounding_box:
+                        bb = self.visual_anchor.bounding_box
+                        match_found = True
+                        best_match = {
+                            'confidence': 0.7,
+                            'bbox': bb,
+                            'center': {'x': bb.get('x', 0) + bb.get('width', 0) // 2,
+                                       'y': bb.get('y', 0) + bb.get('height', 0) // 2}
+                        }
+                        break
+
                 if attempt < self.max_attempts - 1:
-                    time.sleep(0.5)  # Attendre avant nouvelle tentative
+                    time.sleep(0.5)
             
             if not match_found:
                 # Ancre non trouvée
@@ -334,24 +381,23 @@ class VWBFocusAnchorAction(BaseVWBAction):
         try:
             center = match_info['center']
             
+            import pyautogui
+
             if self.focus_method == 'hover':
-                # Survol de l'élément
                 print(f"   Survol à ({center['x']}, {center['y']}) pendant {self.hover_duration_ms}ms")
-                # Simulation du survol
+                pyautogui.moveTo(center['x'], center['y'], duration=0.3)
                 time.sleep(self.hover_duration_ms / 1000.0)
                 return True
-                
+
             elif self.focus_method == 'click_light':
-                # Clic léger (sans appui prolongé)
                 print(f"   Clic léger à ({center['x']}, {center['y']})")
-                # Simulation du clic léger
+                pyautogui.click(center['x'], center['y'])
                 time.sleep(0.1)
                 return True
-                
+
             elif self.focus_method == 'tab':
-                # Navigation par tabulation (approximative)
                 print("   Navigation par tabulation")
-                # Simulation de la tabulation
+                pyautogui.press('tab')
                 time.sleep(0.2)
                 return True
                 

visual_workflow_builder/backend/actions/vision_ui/scroll_to_anchor.py

@@ -449,14 +449,48 @@ class VWBScrollToAnchorAction(BaseVWBAction):
         return None
     
     def _find_visual_element(self, screenshot, visual_anchor, threshold):
-        """Simulation de recherche d'élément visuel."""
+        """Recherche d'élément visuel via template matching."""
-        import random
+        try:
-        confidence = random.uniform(0.6, 0.95)
+            from ...catalog_routes import find_visual_anchor_on_screen
-        
+
-        if confidence >= threshold:
+            image_ancre = None
-            return True, {'x': 400, 'y': 300, 'width': 200, 'height': 50}, confidence
+            bounding_box = None
-        else:
+
-            return False, {}, confidence
+            if isinstance(visual_anchor, VWBVisualAnchor):
+                image_ancre = visual_anchor.screenshot_base64
+                if visual_anchor.has_bounding_box():
+                    bounding_box = visual_anchor.bounding_box
+            elif isinstance(visual_anchor, dict):
+                image_ancre = visual_anchor.get('screenshot') or visual_anchor.get('image_base64')
+                bounding_box = visual_anchor.get('bounding_box')
+
+            if image_ancre:
+                resultat = find_visual_anchor_on_screen(
+                    anchor_image_base64=image_ancre,
+                    confidence_threshold=threshold,
+                    bounding_box=bounding_box
+                )
+                if resultat and resultat.get('found'):
+                    coords = {
+                        'x': resultat.get('x', resultat.get('center_x', 0)),
+                        'y': resultat.get('y', resultat.get('center_y', 0)),
+                        'width': resultat.get('width', 200),
+                        'height': resultat.get('height', 50)
+                    }
+                    return True, coords, resultat.get('confidence', 0.9)
+
+            if bounding_box:
+                return True, bounding_box, 0.7
+
+            return False, {}, 0.0
+
+        except ImportError:
+            if hasattr(visual_anchor, 'bounding_box') and visual_anchor.bounding_box:
+                return True, visual_anchor.bounding_box, 0.7
+            return False, {}, 0.0
+        except Exception as e:
+            print(f"⚠️ Erreur recherche visuelle: {e}")
+            return False, {}, 0.0
     
     def _encode_screenshot(self, screenshot_data) -> str:
         """Encode un screenshot en base64."""
@@ -492,19 +526,18 @@ class VWBScrollToAnchorAction(BaseVWBAction):
         scroll_y = 0
         
         try:
+            import pyautogui
+
             if self.scroll_direction in ['vertical', 'both']:
-                # Défilement vertical vers le bas
                 scroll_y = self.scroll_step_pixels
                 print(f"   ⬇️ Défilement vertical: {scroll_y}px")
-                # En réalité: pyautogui.scroll(-scroll_y)
+                pyautogui.scroll(-scroll_y // 100)
-            
+
             if self.scroll_direction in ['horizontal', 'both']:
-                # Défilement horizontal vers la droite
                 scroll_x = self.scroll_step_pixels
                 print(f"   ➡️ Défilement horizontal: {scroll_x}px")
-                # En réalité: pyautogui.hscroll(scroll_x)
+                pyautogui.hscroll(scroll_x // 100)
-            
+
-            # Simuler le délai de défilement
             time.sleep(0.1)
             
         except Exception as e:

visual_workflow_builder/backend/api_v3/execute.py

@@ -388,7 +388,7 @@ def execute_ai_analyze(params: dict) -> dict:
 
     try:
         prompt = params.get('analysis_prompt', params.get('prompt', ''))
-        model = params.get('model', params.get('ollama_model', 'qwen3-vl:8b'))
+        model = params.get('model', params.get('ollama_model', os.environ.get("RPA_VLM_MODEL", os.environ.get("VLM_MODEL", "gemma4:e4b"))))
         output_variable = params.get('output_variable', 'resultat_analyse')
         timeout_ms = params.get('timeout_ms', 120000)  # 2 minutes par défaut
         temperature = params.get('temperature', 0.7)    # Même défaut que CLI Ollama
@@ -532,6 +532,125 @@ def execute_ai_analyze(params: dict) -> dict:
         return {'success': False, 'error': str(e)}
 
 
+def execute_extract_text(params: dict) -> dict:
+    """
+    Extrait du texte depuis l'écran via Ollama VLM.
+    Capture la zone de l'ancre (ou l'écran entier) et demande au VLM d'extraire le texte.
+    """
+    import requests
+    import re
+    global _execution_state
+
+    try:
+        anchor = params.get('visual_anchor', {})
+        model = params.get('model', os.environ.get("RPA_VLM_MODEL", os.environ.get("VLM_MODEL", "gemma4:e4b")))
+        output_variable = params.get('output_variable', 'texte_extrait')
+        timeout_ms = params.get('timeout_ms', 60000)
+        extraction_mode = params.get('extraction_mode', 'full')
+        text_filters = params.get('text_filters', [])
+
+        screenshot_base64 = anchor.get('screenshot') if anchor else None
+
+        if not screenshot_base64:
+            try:
+                from PIL import ImageGrab
+                import io
+
+                bbox = anchor.get('bounding_box', {}) if anchor else {}
+
+                if bbox:
+                    x, y = int(bbox.get('x', 0)), int(bbox.get('y', 0))
+                    w, h = int(bbox.get('width', 100)), int(bbox.get('height', 100))
+                    print(f"📸 [OCR] Capture zone: ({x}, {y}) -> ({x+w}, {y+h})")
+                    screenshot = ImageGrab.grab(bbox=(x, y, x + w, y + h))
+                else:
+                    print(f"📸 [OCR] Capture écran complet")
+                    screenshot = ImageGrab.grab()
+
+                buffer = io.BytesIO()
+                screenshot.save(buffer, format='PNG')
+                screenshot_base64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
+            except Exception as cap_err:
+                return {'success': False, 'error': f"Erreur capture: {cap_err}"}
+
+        if not screenshot_base64:
+            return {'success': False, 'error': "Pas d'image à analyser"}
+
+        prompt_map = {
+            'full': "Extrais TOUT le texte visible dans cette image. Retourne uniquement le texte brut, sans commentaire.",
+            'numbers': "Extrais uniquement les nombres et chiffres visibles dans cette image. Retourne-les séparés par des espaces.",
+            'lines': "Extrais tout le texte visible ligne par ligne. Une ligne par ligne de texte visible.",
+            'words': "Extrais tous les mots visibles dans cette image, séparés par des espaces.",
+        }
+        prompt = prompt_map.get(extraction_mode, prompt_map['full'])
+
+        if 'qwen' in model.lower() and not prompt.startswith('/no_think'):
+            prompt = f"/no_think\n{prompt}"
+
+        print(f"📝 [OCR] Extraction texte avec {model} (mode: {extraction_mode})...")
+
+        ollama_url = params.get('ollama_url', 'http://localhost:11434')
+        payload = {
+            "model": model,
+            "prompt": prompt,
+            "images": [screenshot_base64],
+            "stream": False,
+            "options": {"temperature": 0.1, "num_predict": 4000}
+        }
+
+        response = requests.post(
+            f"{ollama_url}/api/generate",
+            json=payload,
+            timeout=timeout_ms / 1000
+        )
+
+        if response.status_code != 200:
+            return {'success': False, 'error': f"Erreur Ollama: {response.status_code}"}
+
+        result = response.json()
+        extracted_text = result.get('response', '').strip()
+
+        if not extracted_text and result.get('thinking'):
+            extracted_text = result.get('thinking', '').strip()
+
+        for f in text_filters:
+            if f == 'digits_only':
+                extracted_text = re.sub(r'[^\d\s]', '', extracted_text)
+            elif f == 'letters_only':
+                extracted_text = re.sub(r'[^a-zA-ZÀ-ÿ\s]', '', extracted_text)
+            elif f == 'trim_whitespace':
+                extracted_text = extracted_text.strip()
+            elif f == 'uppercase':
+                extracted_text = extracted_text.upper()
+            elif f == 'lowercase':
+                extracted_text = extracted_text.lower()
+
+        print(f"✅ [OCR] Texte extrait ({len(extracted_text)} caractères)")
+        if extracted_text:
+            print(f"   Résultat: {extracted_text[:150]}...")
+
+        _execution_state['variables'][output_variable] = extracted_text
+
+        return {
+            'success': True,
+            'output': {
+                'extracted_text': extracted_text,
+                'variable': output_variable,
+                'character_count': len(extracted_text),
+                'word_count': len(extracted_text.split()) if extracted_text else 0,
+                'mode': extraction_mode,
+                'model': model
+            }
+        }
+
+    except requests.exceptions.Timeout:
+        return {'success': False, 'error': f"Timeout Ollama après {timeout_ms}ms"}
+    except requests.exceptions.ConnectionError:
+        return {'success': False, 'error': "Ollama non accessible"}
+    except Exception as e:
+        return {'success': False, 'error': str(e)}
+
+
 def execute_action_with_coords(action_type: str, params: dict, coords: dict) -> dict:
     """
     Exécute une action avec des coordonnées spécifiées par l'utilisateur (self-healing).
@@ -792,6 +911,167 @@ def execute_action(action_type: str, params: dict) -> dict:
             # Analyse de texte avec IA (Ollama)
             return execute_ai_analyze(params)
 
+        elif action_type in ['hover_anchor', 'hover']:
+            anchor = params.get('visual_anchor', {})
+            bbox = anchor.get('bounding_box', {})
+            if not bbox:
+                return {'success': False, 'error': 'Pas de bounding_box dans visual_anchor'}
+
+            x = bbox.get('x', 0) + bbox.get('width', 0) / 2
+            y = bbox.get('y', 0) + bbox.get('height', 0) / 2
+            duration_ms = params.get('hover_duration_ms', params.get('duration_ms', 1000))
+
+            print(f"🖱️ [Action] Survol à ({x}, {y}) pendant {duration_ms}ms")
+            pyautogui.moveTo(x, y, duration=0.3)
+            time.sleep(duration_ms / 1000)
+            return {'success': True, 'output': {'hovered_at': {'x': x, 'y': y}, 'duration_ms': duration_ms}}
+
+        elif action_type in ['drag_drop_anchor', 'drag_drop']:
+            source_anchor = params.get('source_anchor', params.get('visual_anchor', {}))
+            dest_anchor = params.get('destination_anchor', {})
+            source_bbox = source_anchor.get('bounding_box', {})
+            dest_bbox = dest_anchor.get('bounding_box', {})
+
+            if not source_bbox or not dest_bbox:
+                return {'success': False, 'error': 'bounding_box source et destination requis'}
+
+            src_x = source_bbox.get('x', 0) + source_bbox.get('width', 0) / 2
+            src_y = source_bbox.get('y', 0) + source_bbox.get('height', 0) / 2
+            dst_x = dest_bbox.get('x', 0) + dest_bbox.get('width', 0) / 2
+            dst_y = dest_bbox.get('y', 0) + dest_bbox.get('height', 0) / 2
+            duration_ms = params.get('drag_duration_ms', 500)
+
+            print(f"🖱️ [Action] Glisser de ({src_x}, {src_y}) vers ({dst_x}, {dst_y})")
+            pyautogui.moveTo(src_x, src_y, duration=0.2)
+            time.sleep(0.1)
+            pyautogui.drag(dst_x - src_x, dst_y - src_y, duration=duration_ms / 1000, button='left')
+            return {'success': True, 'output': {'from': {'x': src_x, 'y': src_y}, 'to': {'x': dst_x, 'y': dst_y}}}
+
+        elif action_type in ['scroll_to_anchor', 'scroll']:
+            direction = params.get('scroll_direction', 'down')
+            amount = params.get('scroll_amount', params.get('scroll_step_pixels', 3))
+            anchor = params.get('visual_anchor', {})
+            bbox = anchor.get('bounding_box', {})
+
+            if bbox:
+                x = bbox.get('x', 0) + bbox.get('width', 0) / 2
+                y = bbox.get('y', 0) + bbox.get('height', 0) / 2
+                pyautogui.moveTo(x, y, duration=0.1)
+
+            scroll_value = amount if direction in ['up', 'left'] else -amount
+
+            print(f"📜 [Action] Scroll {direction} ({amount})")
+            if direction in ['left', 'right']:
+                pyautogui.hscroll(scroll_value)
+            else:
+                pyautogui.scroll(scroll_value)
+
+            time.sleep(0.5)
+            return {'success': True, 'output': {'direction': direction, 'amount': amount}}
+
+        elif action_type in ['focus_anchor', 'focus']:
+            anchor = params.get('visual_anchor', {})
+            bbox = anchor.get('bounding_box', {})
+            if not bbox:
+                return {'success': False, 'error': 'Pas de bounding_box dans visual_anchor'}
+
+            x = bbox.get('x', 0) + bbox.get('width', 0) / 2
+            y = bbox.get('y', 0) + bbox.get('height', 0) / 2
+
+            print(f"🎯 [Action] Focus à ({x}, {y})")
+            pyautogui.click(x, y)
+            time.sleep(0.3)
+            return {'success': True, 'output': {'focused_at': {'x': x, 'y': y}}}
+
+        elif action_type == 'extract_text':
+            return execute_extract_text(params)
+
+        elif action_type == 'ai_ocr':
+            params.setdefault('analysis_prompt', "Extrais TOUT le texte visible dans cette image. Retourne uniquement le texte brut, ligne par ligne, sans commentaire.")
+            return execute_ai_analyze(params)
+
+        elif action_type == 'ai_summarize':
+            params.setdefault('analysis_prompt', "Résume le contenu visible dans cette image en 3-5 phrases concises. Identifie les informations clés.")
+            return execute_ai_analyze(params)
+
+        elif action_type == 'ai_extract':
+            params.setdefault('analysis_prompt', "Extrais les données structurées visibles (noms, dates, montants, identifiants). Retourne un JSON structuré.")
+            return execute_ai_analyze(params)
+
+        elif action_type == 'ai_classify':
+            categories = params.get('categories', [])
+            cats_str = ', '.join(categories) if categories else 'les catégories pertinentes'
+            params.setdefault('analysis_prompt', f"Classe le contenu visible parmi : {cats_str}. Retourne la catégorie et un score de confiance.")
+            return execute_ai_analyze(params)
+
+        elif action_type == 'ai_custom':
+            system_prompt = params.get('system_prompt', '')
+            if system_prompt and 'analysis_prompt' not in params:
+                params['analysis_prompt'] = system_prompt
+            return execute_ai_analyze(params)
+
+        elif action_type == 'screenshot_evidence':
+            import pyautogui
+            from PIL import Image
+            from pathlib import Path
+            import io
+
+            label = params.get('label', params.get('description', 'evidence'))
+            output_variable = params.get('output_variable', 'screenshot_evidence')
+
+            screenshot = pyautogui.screenshot()
+
+            # Sauvegarder la preuve
+            evidence_dir = Path('data/evidence')
+            evidence_dir.mkdir(parents=True, exist_ok=True)
+            timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
+            filepath = evidence_dir / f"evidence_{timestamp}_{label[:30]}.png"
+            screenshot.save(str(filepath))
+
+            # Encoder en base64 pour la variable
+            buffer = io.BytesIO()
+            screenshot.save(buffer, format='PNG')
+            screenshot_b64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
+
+            _execution_state['variables'][output_variable] = screenshot_b64
+
+            print(f"📸 [Evidence] Capture sauvegardée: {filepath}")
+            return {'success': True, 'output': {'filepath': str(filepath), 'variable': output_variable}}
+
+        elif action_type in ['verify_element_exists', 'verify_element']:
+            anchor = params.get('visual_anchor', {})
+            screenshot_base64 = anchor.get('screenshot')
+            bbox = anchor.get('bounding_box', {})
+            expected = params.get('expected', True)
+            output_variable = params.get('output_variable', 'element_exists')
+
+            found = False
+            confidence = 0.0
+
+            if screenshot_base64 and execution_mode in ['intelligent', 'debug']:
+                try:
+                    from services.intelligent_executor import find_and_click
+                    result = find_and_click(
+                        anchor_image_base64=screenshot_base64,
+                        anchor_bbox=bbox,
+                        method='clip',
+                        detection_threshold=0.35
+                    )
+                    found = result.get('found', False)
+                    confidence = result.get('confidence', 0.0)
+                except Exception as e:
+                    print(f"⚠️ [Verify] Erreur vision: {e}")
+            elif bbox:
+                found = True
+                confidence = 0.5
+
+            match = (found == expected)
+            _execution_state['variables'][output_variable] = found
+
+            status = "trouvé" if found else "absent"
+            print(f"🔍 [Verify] Élément {status} (confiance: {confidence:.2f}, attendu: {expected})")
+            return {'success': match, 'output': {'found': found, 'confidence': confidence, 'expected': expected, 'match': match}}
+
         else:
             return {'success': False, 'error': f"Type d'action non supporté: {action_type}"}
 

visual_workflow_builder/backend/api_v3/learned_workflows.py

@@ -27,7 +27,7 @@ from flask import jsonify, request
 
 from . import api_v3_bp
 from .workflow import generate_id
-from db.models import db, Workflow, Step
+from db.models import db, Workflow, Step, VisualAnchor
 
 logger = logging.getLogger(__name__)
 
@@ -303,7 +303,7 @@ def import_learned_workflow(workflow_id: str):
 
         db.session.add(workflow)
 
-        # Créer les steps
+        # Créer les steps (avec sauvegarde des screenshots d'ancres)
         for step_data in steps_list:
             step = Step(
                 id=generate_id("step"),
@@ -314,7 +314,57 @@ def import_learned_workflow(workflow_id: str):
                 position_y=step_data.get("position_y", 200),
                 label=step_data.get("label", step_data["action_type"]),
             )
-            step.parameters = step_data.get("parameters", {})
+            params = dict(step_data.get("parameters", {}))
+
+            # Extraire et sauvegarder le screenshot d'ancre si présent
+            anchor_b64 = params.pop("_anchor_image_base64", None)
+            params.pop("_anchor_bbox", None)
+            if anchor_b64:
+                try:
+                    from services.anchor_image_service import (
+                        save_anchor_image, generate_anchor_id
+                    )
+                    from PIL import Image
+                    from io import BytesIO
+                    import base64 as b64mod
+
+                    if ',' in anchor_b64:
+                        anchor_b64 = anchor_b64.split(',', 1)[1]
+                    img_data = b64mod.b64decode(anchor_b64)
+                    img = Image.open(BytesIO(img_data))
+                    bbox = {
+                        "x": 0, "y": 0,
+                        "width": img.width, "height": img.height
+                    }
+                    anchor_id = generate_anchor_id()
+                    result = save_anchor_image(
+                        anchor_id=anchor_id,
+                        image_base64=anchor_b64,
+                        bounding_box=bbox,
+                        metadata={"source": "learned_import", "workflow_id": wf_id}
+                    )
+                    if result.get("success"):
+                        from services.anchor_image_service import (
+                            get_original_path, get_thumbnail_path
+                        )
+                        va = VisualAnchor(
+                            id=anchor_id,
+                            image_path=str(get_original_path(anchor_id) or ""),
+                            thumbnail_path=str(get_thumbnail_path(anchor_id) or ""),
+                            bbox_x=0, bbox_y=0,
+                            bbox_width=img.width, bbox_height=img.height,
+                            description=step_data.get("label", ""),
+                            capture_method="learned_import",
+                        )
+                        db.session.add(va)
+                        step.anchor_id = anchor_id
+                        logger.info("Ancre sauvegardée: %s pour step %s",
+                                    anchor_id, step.id)
+                except Exception as e:
+                    logger.warning("Échec sauvegarde ancre pour step %s: %s",
+                                   step_data.get("order"), e)
+
+            step.parameters = params
             db.session.add(step)
 
         db.session.commit()

visual_workflow_builder/backend/services/learned_workflow_bridge.py

@@ -218,6 +218,20 @@ def convert_learned_to_vwb_steps(
             if target.get("by_text"):
                 vwb_params["target_text"] = target["by_text"]
 
+            # Extraire le screenshot de l'ancre pour la preview dans le VWB
+            anchor_b64 = (
+                target.get("anchor_image_base64")
+                or target.get("screenshot")
+                or action_params.get("anchor_image_base64")
+            )
+            if anchor_b64:
+                vwb_params["_anchor_image_base64"] = anchor_b64
+                bbox = target.get("by_position")
+                if bbox and isinstance(bbox, (list, tuple)) and len(bbox) >= 2:
+                    vwb_params["_anchor_bbox"] = {
+                        "x_pct": bbox[0], "y_pct": bbox[1]
+                    }
+
             label = _build_step_label(vwb_action_type, vwb_params, from_name, to_name)
             steps.append({
                 "action_type": vwb_action_type,
@@ -229,6 +243,10 @@ def convert_learned_to_vwb_steps(
                 "metadata": edge_meta,
             })
 
+    # Fusionner les type_text consécutifs et les key_press en combos
+    steps = _merge_consecutive_text_inputs(steps)
+    steps = _merge_consecutive_key_presses(steps)
+
     # Appliquer le layout serpentin à tous les steps
     _compute_layout(steps)
 
@@ -298,6 +316,79 @@ def _convert_compound_substep(
     return vwb_type, vwb_params
 
 
+def _merge_consecutive_text_inputs(
+    steps: List[Dict[str, Any]],
+) -> List[Dict[str, Any]]:
+    """
+    Fusionne les steps type_text consécutifs en un seul.
+
+    Quand un compound est décomposé lettre par lettre (ex: "bonjour" → 7 steps),
+    cette fonction les recombine en un seul step "Saisir : bonjour".
+    """
+    if not steps:
+        return steps
+
+    merged = [steps[0]]
+    for step in steps[1:]:
+        prev = merged[-1]
+        if (prev["action_type"] == "type_text"
+                and step["action_type"] == "type_text"):
+            # Concaténer le texte
+            prev_text = prev.get("parameters", {}).get("text", "")
+            curr_text = step.get("parameters", {}).get("text", "")
+            prev["parameters"]["text"] = prev_text + curr_text
+            # Mettre à jour le label
+            combined = prev["parameters"]["text"]
+            prev["label"] = f'Saisir : "{combined}"'
+        else:
+            merged.append(step)
+
+    # Réindexer les ordres
+    for idx, step in enumerate(merged):
+        step["order"] = idx
+
+    return merged
+
+
+def _merge_consecutive_key_presses(
+    steps: List[Dict[str, Any]],
+) -> List[Dict[str, Any]]:
+    """
+    Fusionne les key_press / keyboard_shortcut consécutifs portant une seule touche
+    en un seul keyboard_shortcut combo (ex: ctrl puis s → ctrl+s).
+
+    Ne fusionne que les steps keyboard_shortcut consécutifs dont chacun ne porte
+    qu'une seule touche (signe d'un combo décomposé). Les raccourcis déjà composés
+    (keys avec 2+ éléments) ne sont pas touchés.
+    """
+    if not steps:
+        return steps
+
+    merged = [steps[0]]
+    for step in steps[1:]:
+        prev = merged[-1]
+        if (prev["action_type"] == "keyboard_shortcut"
+                and step["action_type"] == "keyboard_shortcut"):
+            prev_keys = prev.get("parameters", {}).get("keys", [])
+            curr_keys = step.get("parameters", {}).get("keys", [])
+            # Ne fusionner que si chaque step porte exactement 1 touche
+            # (un combo déjà composé comme ["ctrl", "s"] ne doit pas absorber le suivant)
+            if len(curr_keys) == 1 and len(prev_keys) >= 1:
+                # Vérifier que le prev est lui-même issu d'une fusion ou d'une seule touche
+                # On fusionne tant que c'est un enchaînement de touches simples
+                prev["parameters"]["keys"] = prev_keys + curr_keys
+                combo_str = "+".join(prev["parameters"]["keys"])
+                prev["label"] = f"Raccourci : {combo_str}"
+                continue
+        merged.append(step)
+
+    # Réindexer les ordres
+    for idx, step in enumerate(merged):
+        step["order"] = idx
+
+    return merged
+
+
 def _compute_layout(
     steps: List[Dict[str, Any]],
     cols: int = 3,

visual_workflow_builder/frontend_v4/src/components/ToolPalette.tsx

@@ -27,7 +27,7 @@ export default function ToolPalette() {
       <div className="tool-categories">
         {categories.map((catKey) => {
           const cat = ACTION_CATEGORIES[catKey];
-          const tools = ACTIONS.filter(a => a.category === catKey);
+          const tools = ACTIONS.filter(a => a.category === catKey && !a.hidden);
           const isExpanded = expandedCategories.includes(catKey);
 
           if (tools.length === 0) return null;

visual_workflow_builder/frontend_v4/src/types.ts

@@ -71,6 +71,7 @@ export interface ActionDefinition {
   category: 'mouse' | 'keyboard' | 'wait' | 'data' | 'logic' | 'ai' | 'llm' | 'validation' | 'files';
   needsAnchor: boolean;
   params: { name: string; type: string; description: string }[];
+  hidden?: boolean;
 }
 
 export const ACTIONS: ActionDefinition[] = [
@@ -116,11 +117,11 @@ export const ACTIONS: ActionDefinition[] = [
   ] },
 
   // === LOGIQUE ===
-  { type: 'visual_condition', label: 'Condition visuelle', icon: '🔀', description: 'Branchement conditionnel : si l\'ancre est trouvée, suit la sortie bas ; sinon, la sortie droite.', category: 'logic', needsAnchor: true, params: [
+  { type: 'visual_condition', label: 'Condition visuelle', icon: '🔀', description: 'Branchement conditionnel : si l\'ancre est trouvée, suit la sortie bas ; sinon, la sortie droite.', category: 'logic', needsAnchor: true, hidden: true, params: [
     { name: 'on_found', type: 'string', description: 'ID de l\'étape si l\'élément est trouvé' },
     { name: 'on_not_found', type: 'string', description: 'ID de l\'étape si l\'élément n\'est pas trouvé' }
   ] },
-  { type: 'loop_visual', label: 'Boucle visuelle', icon: '🔁', description: 'Répète les étapes connectées tant que l\'ancre est visible.', category: 'logic', needsAnchor: true, params: [
+  { type: 'loop_visual', label: 'Boucle visuelle', icon: '🔁', description: 'Répète les étapes connectées tant que l\'ancre est visible.', category: 'logic', needsAnchor: true, hidden: true, params: [
     { name: 'max_iterations', type: 'number', description: 'Nombre maximum d\'itérations' }
   ] },