chore(qw): cleanup post-review (préfixes BUS, événements monitor, import io)

- safety_checks_provider : tous les logger.warning d'échec LLM préfixés
  [BUS] lea:safety_checks_llm_failed avec une raison spécifique
  (exception, http_status, timeout, network, json_decode).
- monitor_router : émission [BUS] lea:monitor_invalid_index si l'index
  explicite passé dans l'action est hors limites de monitors_geometry,
  et [BUS] lea:monitor_unavailable si focus actif demandé mais introuvable.
  Ces deux events permettent au bus de tracer chaque fallback de la cascade
  de routage QW1.
- safety_checks_provider : import io supprimé (inutilisé).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.env.example

@@ -46,6 +46,14 @@ LOGS_PATH=logs
 UPLOADS_PATH=data/training/uploads
 SESSIONS_PATH=data/training/sessions
 
+# ============================================================================
+# Feedback Bus (Léa parle pendant exécution)
+# ============================================================================
+# Bus SocketIO unifié 'lea:*' (action_started, action_done, need_confirm, paused).
+# Désactivé par défaut. Mettre à 1 pour activer les bulles temps réel dans ChatWindow.
+# Si la connexion bus échoue, l'exécution continue normalement (fail-safe).
+LEA_FEEDBACK_BUS=0
+
 # ============================================================================
 # FAISS
 # ============================================================================

agent_chat/app.py

@@ -133,6 +133,28 @@ def _streaming_headers() -> dict:
         headers["Authorization"] = f"Bearer {_STREAMING_API_TOKEN}"
     return headers
 
+
+# ============================================================
+# Feedback Bus — events 'lea:*' temps réel vers ChatWindow
+# ============================================================
+LEA_FEEDBACK_BUS = os.environ.get("LEA_FEEDBACK_BUS", "0").lower() in ("1", "true", "yes", "on")
+
+
+def _emit_lea(event: str, payload: Dict[str, Any]) -> None:
+    """Émet 'lea:{event}' sur le bus SocketIO. No-op silencieux si flag off ou erreur."""
+    if not LEA_FEEDBACK_BUS:
+        return
+    try:
+        socketio.emit(f"lea:{event}", payload)
+    except Exception:
+        logger.debug("_emit_lea silenced", exc_info=True)
+
+
+def _emit_dual(legacy_event: str, lea_event: str, payload: Dict[str, Any], **kwargs) -> None:
+    """Émet l'event legacy (compat dashboard) ET l'alias lea:* (ChatWindow tkinter)."""
+    socketio.emit(legacy_event, payload, **kwargs)
+    _emit_lea(lea_event, payload)
+
 execution_status = {
     "running": False,
     "workflow": None,
@@ -623,7 +645,7 @@ def api_execute():
     }
     
     # Notifier via WebSocket
-    socketio.emit('execution_started', {
+    _emit_dual('execution_started', 'action_started', {
         "workflow": match.workflow_name,
         "params": all_params
     })
@@ -1181,28 +1203,28 @@ def _execute_gesture(gesture):
         )
 
         if resp.status_code == 200:
-            socketio.emit('execution_completed', {
+            _emit_dual('execution_completed', 'done', {
                 "workflow": gesture.name,
                 "success": True,
                 "message": f"Geste '{gesture.name}' ({'+'.join(gesture.keys)}) envoyé",
             })
         else:
             error = resp.text[:200]
-            socketio.emit('execution_completed', {
+            _emit_dual('execution_completed', 'done', {
                 "workflow": gesture.name,
                 "success": False,
                 "message": f"Erreur: {error}",
             })
 
     except http_requests.ConnectionError:
-        socketio.emit('execution_completed', {
+        _emit_dual('execution_completed', 'done', {
             "workflow": gesture.name,
             "success": False,
             "message": "Serveur de streaming non disponible (port 5005).",
         })
     except Exception as e:
         logger.error(f"Gesture execution error: {e}")
-        socketio.emit('execution_completed', {
+        _emit_dual('execution_completed', 'done', {
             "workflow": gesture.name,
             "success": False,
             "message": f"Erreur: {str(e)}",
@@ -1661,6 +1683,52 @@ def handle_copilot_abort():
     })
 
 
+# =============================================================================
+# Bulle paused_need_help — handlers SocketIO depuis ChatWindow (J3.5)
+# =============================================================================
+
+@socketio.on('lea:replay_resume')
+def handle_lea_replay_resume(data):
+    """Bouton Continuer : relayer le resume vers le streaming server."""
+    replay_id = (data or {}).get("replay_id")
+    if not replay_id:
+        _emit_lea("resume_acked", {"status": "error", "detail": "replay_id manquant"})
+        return
+    try:
+        resp = http_requests.post(
+            f"{STREAMING_SERVER_URL}/api/v1/traces/stream/replay/{replay_id}/resume",
+            headers=_streaming_headers(),
+            timeout=5,
+        )
+        if resp.ok:
+            logger.info(f"Replay {replay_id} resume relayé OK")
+            _emit_lea("resume_acked", {"replay_id": replay_id, "status": "ok"})
+        else:
+            detail = resp.text[:200]
+            logger.warning(f"Resume échoué (HTTP {resp.status_code}): {detail}")
+            _emit_lea("resume_acked", {
+                "replay_id": replay_id, "status": "error",
+                "http_status": resp.status_code, "detail": detail,
+            })
+    except Exception as e:
+        logger.warning(f"Resume relay error: {e}")
+        _emit_lea("resume_acked", {
+            "replay_id": replay_id, "status": "error", "detail": str(e),
+        })
+
+
+@socketio.on('lea:replay_abort')
+def handle_lea_replay_abort(data):
+    """Bouton Annuler : arrêter le polling local. Le replay côté streaming sera
+    cleaned up naturellement au prochain replay (cf api_stream._replay_states stale)."""
+    global execution_status
+    replay_id = (data or {}).get("replay_id")
+    execution_status["running"] = False
+    execution_status["message"] = "Annulé par l'utilisateur"
+    logger.info(f"Replay {replay_id or '?'} abort par l'utilisateur (paused bubble)")
+    _emit_lea("abort_acked", {"replay_id": replay_id, "status": "ok"})
+
+
 # =============================================================================
 # Exécution de workflow
 # =============================================================================
@@ -1730,14 +1798,20 @@ def _poll_replay_progress(replay_id: str, workflow_name: str, total_actions: int
     """Suivre la progression d'un replay distant via polling."""
     import time
 
-    max_wait = 120  # 2 minutes max
+    max_wait_running = 120   # 2 min en exécution active
+    max_wait_paused = 600    # 10 min en pause supervisée (humain peut prendre son temps)
     poll_interval = 2.0
     elapsed = 0
+    was_paused = False
 
-    while elapsed < max_wait and execution_status.get("running"):
+    while execution_status.get("running"):
         time.sleep(poll_interval)
         elapsed += poll_interval
 
+        cap = max_wait_paused if was_paused else max_wait_running
+        if elapsed >= cap:
+            break
+
         try:
             resp = http_requests.get(
                 f"{STREAMING_SERVER_URL}/api/v1/traces/stream/replay/{replay_id}",
@@ -1753,7 +1827,26 @@ def _poll_replay_progress(replay_id: str, workflow_name: str, total_actions: int
             failed = data.get("failed_actions", 0)
             progress = int(10 + (completed / max(total_actions, 1)) * 80)
 
-            socketio.emit('execution_progress', {
+            if status == "paused_need_help" and not was_paused:
+                _emit_lea("paused", {
+                    "workflow": workflow_name,
+                    "replay_id": replay_id,
+                    "completed": completed,
+                    "total": total_actions,
+                    "failed_action": data.get("failed_action"),
+                    "reason": data.get("error") or "Action incertaine",
+                })
+                was_paused = True
+                elapsed = 0
+            elif was_paused and status != "paused_need_help":
+                _emit_lea("resumed", {
+                    "workflow": workflow_name,
+                    "replay_id": replay_id,
+                    "status_after": status,
+                })
+                was_paused = False
+
+            _emit_dual('execution_progress', 'action_progress', {
                 "progress": progress,
                 "step": f"Action {completed}/{total_actions} exécutée",
                 "current": completed,
@@ -1922,7 +2015,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
 
     actions = _build_actions_from_workflow(match, params)
     if not actions:
-        socketio.emit('copilot_complete', {
+        _emit_dual('copilot_complete', 'done', {
             "workflow": workflow_name,
             "status": "error",
             "message": "Aucune action exécutable dans ce workflow.",
@@ -1959,7 +2052,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
             break
 
         copilot_state["status"] = "waiting_approval"
-        socketio.emit('copilot_step', {
+        _emit_dual('copilot_step', 'need_confirm', {
             "workflow": workflow_name,
             "step_index": idx,
             "total": total,
@@ -1982,7 +2075,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
 
         if waited >= max_wait:
             copilot_state["status"] = "aborted"
-            socketio.emit('copilot_complete', {
+            _emit_dual('copilot_complete', 'done', {
                 "workflow": workflow_name,
                 "status": "timeout",
                 "message": f"Timeout : pas de réponse après {max_wait}s.",
@@ -1999,7 +2092,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
         elif decision == "skipped":
             copilot_state["skipped"] += 1
             logger.info(f"Copilot skip étape {idx + 1}/{total}")
-            socketio.emit('copilot_step_result', {
+            _emit_dual('copilot_step_result', 'step_result', {
                 "step_index": idx,
                 "total": total,
                 "status": "skipped",
@@ -2034,7 +2127,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
 
                     if action_success:
                         copilot_state["completed"] += 1
-                        socketio.emit('copilot_step_result', {
+                        _emit_dual('copilot_step_result', 'step_result', {
                             "step_index": idx,
                             "total": total,
                             "status": "completed",
@@ -2042,7 +2135,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
                         })
                     else:
                         copilot_state["failed"] += 1
-                        socketio.emit('copilot_step_result', {
+                        _emit_dual('copilot_step_result', 'step_result', {
                             "step_index": idx,
                             "total": total,
                             "status": "failed",
@@ -2051,7 +2144,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
                 else:
                     error = resp.text[:200]
                     copilot_state["failed"] += 1
-                    socketio.emit('copilot_step_result', {
+                    _emit_dual('copilot_step_result', 'step_result', {
                         "step_index": idx,
                         "total": total,
                         "status": "failed",
@@ -2060,7 +2153,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
 
             except http_requests.ConnectionError:
                 copilot_state["failed"] += 1
-                socketio.emit('copilot_step_result', {
+                _emit_dual('copilot_step_result', 'step_result', {
                     "step_index": idx,
                     "total": total,
                     "status": "failed",
@@ -2070,7 +2163,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
             except Exception as e:
                 copilot_state["failed"] += 1
                 logger.error(f"Copilot action error: {e}")
-                socketio.emit('copilot_step_result', {
+                _emit_dual('copilot_step_result', 'step_result', {
                     "step_index": idx,
                     "total": total,
                     "status": "failed",
@@ -2098,7 +2191,7 @@ def execute_workflow_copilot(match, params: Dict[str, Any]):
             f"Copilot terminé : {completed} réussies, "
             f"{skipped} passées, {failed} échouées sur {total} étapes."
         )
-        socketio.emit('copilot_complete', {
+        _emit_dual('copilot_complete', 'done', {
             "workflow": workflow_name,
             "status": "completed" if success else "partial",
             "message": message,
@@ -2175,7 +2268,7 @@ def execute_workflow(match, params):
         execution_status["progress"] = 10
         execution_status["message"] = f"Envoyé à l'Agent V1 ({target_session})"
 
-        socketio.emit('execution_progress', {
+        _emit_dual('execution_progress', 'action_progress', {
             "progress": 10,
             "step": f"Replay envoyé à l'Agent V1 — {total_actions} actions en attente",
             "current": 0,
@@ -2523,7 +2616,7 @@ def update_progress(progress: int, message: str, current: int, total: int):
     execution_status["progress"] = progress
     execution_status["message"] = message
     
-    socketio.emit('execution_progress', {
+    _emit_dual('execution_progress', 'action_progress', {
         "progress": progress,
         "step": message,
         "current": current,
@@ -2543,7 +2636,7 @@ def finish_execution(workflow_name: str, success: bool, message: str):
     if command_history:
         command_history[-1]["status"] = "completed" if success else "failed"
     
-    socketio.emit('execution_completed', {
+    _emit_dual('execution_completed', 'done', {
         "workflow": workflow_name,
         "success": success,
         "message": message

agent_v0/agent_v1/main.py

@@ -448,6 +448,12 @@ class AgentV1:
                         window_title = self.vision.get_active_window_title()
                         if window_title:
                             heartbeat_event["active_window_title"] = window_title
+                        # QW1 — enrichissement multi-écrans (additif, fallback gracieux)
+                        try:
+                            from .vision.capturer import _enrich_with_monitor_info
+                            _enrich_with_monitor_info(heartbeat_event)
+                        except Exception:
+                            pass
                         self.streamer.push_event(heartbeat_event)
             except Exception as e:
                 logger.error(f"Heartbeat error: {e}")

agent_v0/agent_v1/network/feedback_bus.py

@@ -0,0 +1,149 @@
+# agent_v1/network/feedback_bus.py
+"""Client SocketIO pour le bus feedback Léa.
+
+Consomme les events 'lea:*' émis par agent_chat (port 5004) et les dispatche
+vers ChatWindow pour affichage en bulles temps réel.
+
+Events écoutés :
+    lea:action_started   — début d'un workflow ou d'une action
+    lea:action_progress  — progression dans le workflow
+    lea:done             — fin d'un workflow ou d'un copilot
+    lea:need_confirm     — étape copilot en attente de validation
+    lea:step_result      — résultat d'une étape copilot
+    lea:paused           — basculement en paused_need_help (asset démo)
+    lea:resumed          — sortie de pause supervisée
+
+Fail-safe : toute erreur de connexion ou de dispatch est silencieusement
+loggée. Le ChatWindow continue de fonctionner même si le bus est mort
+(comportement strictement identique au pré-J3).
+
+Usage :
+    bus = FeedbackBusClient(
+        server_url="http://localhost:5004",
+        token=os.environ.get("RPA_API_TOKEN", ""),
+        on_event=lambda event, payload: print(event, payload),
+    )
+    bus.start()  # connexion en arrière-plan, non-bloquant
+    # ... ChatWindow tourne ...
+    bus.stop()
+"""
+
+import logging
+import threading
+from typing import Callable, Optional
+
+import socketio
+
+logger = logging.getLogger(__name__)
+
+LEA_EVENTS = (
+    'lea:action_started',
+    'lea:action_progress',
+    'lea:done',
+    'lea:need_confirm',
+    'lea:step_result',
+    'lea:paused',
+    'lea:resumed',
+)
+
+EventCallback = Callable[[str, dict], None]
+
+
+class FeedbackBusClient:
+    """Client SocketIO non-bloquant pour le bus 'lea:*'."""
+
+    def __init__(
+        self,
+        server_url: str,
+        token: Optional[str] = None,
+        on_event: Optional[EventCallback] = None,
+    ):
+        self._url = server_url.rstrip('/')
+        self._token = token or None
+        self._on_event: EventCallback = on_event or (lambda e, p: None)
+        self._sio = socketio.Client(
+            reconnection=True,
+            reconnection_attempts=0,   # 0 = illimité
+            reconnection_delay=2,
+            reconnection_delay_max=30,
+            logger=False,
+            engineio_logger=False,
+        )
+        self._thread: Optional[threading.Thread] = None
+        self._register_handlers()
+
+    def _register_handlers(self) -> None:
+        @self._sio.event
+        def connect():
+            logger.info("FeedbackBus connecté à %s", self._url)
+
+        @self._sio.event
+        def disconnect():
+            logger.info("FeedbackBus déconnecté")
+
+        for ev in LEA_EVENTS:
+            self._sio.on(ev, lambda data, e=ev: self._dispatch(e, data))
+
+    def _dispatch(self, event: str, payload: Optional[dict]) -> None:
+        try:
+            self._on_event(event, payload or {})
+        except Exception:
+            logger.debug("FeedbackBus dispatch silenced", exc_info=True)
+
+    def start(self) -> None:
+        """Démarrer la connexion en arrière-plan (idempotent, non-bloquant)."""
+        if self._thread is not None and self._thread.is_alive():
+            return
+        self._thread = threading.Thread(
+            target=self._run, daemon=True, name="LeaFeedbackBus",
+        )
+        self._thread.start()
+
+    def _run(self) -> None:
+        headers = {}
+        if self._token:
+            headers['Authorization'] = f'Bearer {self._token}'
+        try:
+            self._sio.connect(self._url, headers=headers, wait=True)
+            self._sio.wait()
+        except Exception as e:
+            logger.warning(
+                "FeedbackBus connect échoué (%s) — ChatWindow continue normalement", e,
+            )
+
+    def stop(self) -> None:
+        """Arrêter proprement la connexion (idempotent, fail-safe)."""
+        try:
+            if self._sio.connected:
+                self._sio.disconnect()
+        except Exception:
+            logger.debug("FeedbackBus stop silenced", exc_info=True)
+
+    @property
+    def connected(self) -> bool:
+        return bool(self._sio.connected)
+
+    # ------------------------------------------------------------------
+    # Actions utilisateur depuis la bulle paused_need_help (J3.5)
+    # ------------------------------------------------------------------
+
+    def resume_replay(self, replay_id: str) -> bool:
+        """Bouton Continuer : émet 'lea:replay_resume' vers agent_chat.
+
+        Retourne True si l'event a pu être émis, False sinon (déconnecté/erreur).
+        """
+        return self._safe_emit("lea:replay_resume", {"replay_id": replay_id})
+
+    def abort_replay(self, replay_id: str) -> bool:
+        """Bouton Annuler : émet 'lea:replay_abort' vers agent_chat."""
+        return self._safe_emit("lea:replay_abort", {"replay_id": replay_id})
+
+    def _safe_emit(self, event: str, payload: dict) -> bool:
+        try:
+            if not self._sio.connected:
+                return False
+            self._sio.emit(event, payload)
+            return True
+        except Exception:
+            logger.debug("FeedbackBus _safe_emit silenced", exc_info=True)
+            return False

agent_v0/agent_v1/requirements.txt

@@ -3,7 +3,9 @@ mss>=9.0.1              # Capture d'écran haute performance
 pynput>=1.7.7           # Clavier/Souris Cross-plateforme
 Pillow>=10.0.0          # Crops et processing image
 requests>=2.31.0        # Streaming réseau
+python-socketio[client]>=5.10,<6.0  # Bus feedback Léa 'lea:*' (compat Flask-SocketIO 5.3.x serveur)
 psutil>=5.9.0           # Monitoring CPU/RAM
+screeninfo>=0.8         # QW1 — détection des monitors physiques + offsets
 pystray>=0.19.5         # Icône Tray UI
 plyer>=2.1.0            # Notifications toast natives (remplace PyQt5)
 pywebview>=5.0          # Fenêtre de chat Léa intégrée (Edge WebView2 sur Windows)

agent_v0/agent_v1/ui/chat_window.py

@@ -16,6 +16,15 @@ from typing import Any, Callable, Dict, Optional
 
 logger = logging.getLogger(__name__)
 
+# FeedbackBus : import fail-safe (le ChatWindow doit tourner même si python-socketio
+# n'est pas installé sur le poste client, par exemple ancienne installation Pauline)
+try:
+    from ..network.feedback_bus import FeedbackBusClient
+    _HAS_FEEDBACK_BUS = True
+except Exception:
+    FeedbackBusClient = None  # type: ignore
+    _HAS_FEEDBACK_BUS = False
+
 # ---------------------------------------------------------------------------
 # Theme — palette professionnelle claire
 # ---------------------------------------------------------------------------
@@ -42,6 +51,25 @@ SCROLLBAR_BG = "#E5E7EB"     # Fond scrollbar
 SCROLLBAR_FG = "#9CA3AF"     # Curseur scrollbar
 MSG_BORDER_COLOR = "#D1D5DB"  # Bordure subtile des bulles de messages
 
+# Bulle paused_need_help (J3.5) — alerte non bloquante, asset démo majeur
+PAUSED_BG = "#FEF3C7"             # Jaune pâle
+PAUSED_BORDER = "#F59E0B"         # Orange ambré
+PAUSED_FG = "#92400E"             # Brun foncé (lisible sur fond jaune)
+PAUSED_BTN_RESUME_BG = "#22C55E"  # Vert
+PAUSED_BTN_RESUME_HOVER = "#16A34A"
+PAUSED_BTN_ABORT_BG = "#9CA3AF"   # Gris neutre (pas dramatique)
+PAUSED_BTN_ABORT_HOVER = "#6B7280"
+
+# Bulle "Léa exécute" (J3.4) — distincte des bulles chat normales
+ACTION_BG = "#F1F5F9"             # Gris très clair (différencie d'une réponse chat)
+ACTION_BORDER = "#CBD5E1"         # Gris pâle
+ACTION_FG = "#1E293B"             # Gris foncé
+ACTION_META_FG = "#94A3B8"        # Métadonnées en gris discret
+ACTION_ICON_RUN = "#3B82F6"       # Bleu (en cours)
+ACTION_ICON_OK = "#22C55E"        # Vert (succès)
+ACTION_ICON_ERR = "#EF4444"       # Rouge (échec)
+ACTION_ICON_INFO = "#64748B"      # Gris (neutre)
+
 # Dimensions — confortables
 WIN_WIDTH = 600
 WIN_HEIGHT = 800
@@ -62,6 +90,80 @@ FONT_SEND_BTN = ("Segoe UI", 13)
 FONT_RESIZE_GRIP = ("Segoe UI", 10)
 
 
+# ---------------------------------------------------------------------------
+# Templates de bulles "Léa exécute" (J3.4)
+# Chaque template prend un payload et retourne (icon, icon_color, title).
+# Les libellés sont volontairement neutres : le contexte métier vient du
+# payload (workflow, action, message), pas de hardcoding.
+# ---------------------------------------------------------------------------
+
+def _tpl_action_started(payload: Dict[str, Any]) -> tuple:
+    wf = payload.get("workflow") or "?"
+    return ("▶", ACTION_ICON_RUN, f"Démarrage : {wf}")
+
+
+def _tpl_action_progress(payload: Dict[str, Any]) -> tuple:
+    cur = payload.get("current", "?")
+    tot = payload.get("total", "?")
+    step = payload.get("step")
+    title = step if step else f"Étape {cur}/{tot}"
+    return ("⋯", ACTION_ICON_RUN, str(title))
+
+
+def _tpl_done(payload: Dict[str, Any]) -> tuple:
+    success = bool(payload.get("success", True))
+    msg = payload.get("message") or ("Terminé" if success else "Échec")
+    if success:
+        return ("✓", ACTION_ICON_OK, str(msg))
+    return ("✗", ACTION_ICON_ERR, str(msg))
+
+
+def _tpl_need_confirm(payload: Dict[str, Any]) -> tuple:
+    action = payload.get("action") or {}
+    desc = action.get("description") if isinstance(action, dict) else None
+    title = desc or "Validation requise"
+    return ("?", ACTION_ICON_RUN, str(title))
+
+
+def _tpl_step_result(payload: Dict[str, Any]) -> tuple:
+    status = (payload.get("status") or "").lower()
+    msg = payload.get("message") or status or "Étape terminée"
+    if status in ("ok", "success", "approved"):
+        return ("✓", ACTION_ICON_OK, str(msg))
+    if status in ("error", "failed"):
+        return ("✗", ACTION_ICON_ERR, str(msg))
+    return ("·", ACTION_ICON_INFO, str(msg))
+
+
+def _tpl_resumed(payload: Dict[str, Any]) -> tuple:
+    return ("→", ACTION_ICON_OK, "Reprise")
+
+
+_ACTION_TEMPLATES = {
+    "lea:action_started": _tpl_action_started,
+    "lea:action_progress": _tpl_action_progress,
+    "lea:done": _tpl_done,
+    "lea:need_confirm": _tpl_need_confirm,
+    "lea:step_result": _tpl_step_result,
+    "lea:resumed": _tpl_resumed,
+}
+
+
+def _extract_meta(payload: Dict[str, Any]) -> str:
+    """Métadonnées techniques en pied de bulle (workflow, étape, replay_id court)."""
+    parts = []
+    wf = payload.get("workflow")
+    if wf:
+        parts.append(str(wf))
+    cur, tot = payload.get("current"), payload.get("total")
+    if cur is not None and tot is not None:
+        parts.append(f"étape {cur}/{tot}")
+    rid = payload.get("replay_id")
+    if rid:
+        parts.append(f"#{str(rid)[-6:]}")
+    return " • ".join(parts)
+
+
 class ChatWindow:
     """Fenetre de chat Lea en tkinter natif.
 
@@ -91,6 +193,8 @@ class ChatWindow:
         self._root = None
         self._ready = threading.Event()
         self._messages = []  # historique local
+        self._bus: Optional[Any] = None  # FeedbackBusClient (J3.3, peut rester None)
+        self._active_paused_bubble: Optional[Dict[str, Any]] = None  # bulle paused active (J3.5)
 
         # S'abonner aux changements de l'etat partage
         if self._shared_state is not None:
@@ -266,6 +370,9 @@ class ChatWindow:
             # Signaler que la fenetre est prete
             self._ready.set()
 
+            # Demarrer le bus feedback Lea (events 'lea:*' temps reel)
+            self._start_feedback_bus()
+
             # Boucle tkinter
             root.mainloop()
 
@@ -608,6 +715,12 @@ class ChatWindow:
 
     def _do_destroy(self) -> None:
         """Detruit la fenetre (appele dans le thread tkinter)."""
+        if self._bus is not None:
+            try:
+                self._bus.stop()
+            except Exception:
+                pass
+            self._bus = None
         if self._root is not None:
             try:
                 self._root.quit()
@@ -617,6 +730,232 @@ class ChatWindow:
             self._root = None
         self._visible = False
 
+    # ======================================================================
+    # FeedbackBus — bulles temps reel pendant l'execution (J3.3)
+    # ======================================================================
+
+    def _start_feedback_bus(self) -> None:
+        """Demarrer la connexion au bus 'lea:*' si flag actif et lib disponible."""
+        if not _HAS_FEEDBACK_BUS:
+            logger.debug("FeedbackBus non disponible (python-socketio manquant)")
+            return
+        flag = os.environ.get("LEA_FEEDBACK_BUS", "0").lower()
+        if flag not in ("1", "true", "yes", "on"):
+            return
+        try:
+            url = f"http://{self._server_host}:{self._chat_port}"
+            token = os.environ.get("RPA_API_TOKEN", "") or None
+            self._bus = FeedbackBusClient(url, token=token, on_event=self._on_lea_event)
+            self._bus.start()
+            logger.info("FeedbackBus demarre : %s", url)
+        except Exception:
+            logger.debug("FeedbackBus init silenced", exc_info=True)
+            self._bus = None
+
+    def _on_lea_event(self, event: str, payload: Dict[str, Any]) -> None:
+        """Callback bus → bulle Lea. Thread-safe : helpers utilisent root.after."""
+        payload = payload or {}
+
+        # J3.5 : la pause supervisée a sa propre bulle interactive
+        if event == "lea:paused":
+            self._add_paused_bubble(payload)
+            return
+        if event in ("lea:resumed", "lea:done"):
+            self._close_active_paused_bubble(reason=event)
+            # on continue pour afficher la bulle d'action (cf. dispatch ci-dessous)
+
+        # Acks bus (resume_acked, abort_acked) : silencieux côté UI
+        if event in ("lea:resume_acked", "lea:abort_acked"):
+            return
+
+        # J3.4 : bulle "Léa exécute" stylisée (séparée des bulles chat normales)
+        rendered = _ACTION_TEMPLATES.get(event)
+        if rendered is None:
+            # Event inconnu : on affiche en bulle d'action neutre
+            self._add_action_bubble(
+                icon="·", icon_color=ACTION_ICON_INFO,
+                title=event.removeprefix("lea:"),
+                meta=_extract_meta(payload),
+            )
+            return
+        icon, icon_color, title = rendered(payload)
+        self._add_action_bubble(
+            icon=icon, icon_color=icon_color, title=title,
+            meta=_extract_meta(payload),
+        )
+
+    # ------------------------------------------------------------------
+    # Bulle "Léa exécute" stylisée (J3.4)
+    # ------------------------------------------------------------------
+
+    def _add_action_bubble(
+        self, icon: str, icon_color: str, title: str, meta: str = "",
+    ) -> None:
+        if self._root is None:
+            return
+        self._root.after(0, lambda: self._render_action_bubble(icon, icon_color, title, meta))
+
+    def _render_action_bubble(
+        self, icon: str, icon_color: str, title: str, meta: str,
+    ) -> None:
+        tk = self._tk
+        if getattr(self, "_msg_frame", None) is None:
+            return
+        now = datetime.now().strftime("%H:%M")
+
+        container = tk.Frame(self._msg_frame, bg=BG_COLOR)
+        container.pack(fill=tk.X, padx=MARGIN, pady=3)
+
+        inner = tk.Frame(
+            container, bg=ACTION_BG, padx=10, pady=6,
+            highlightbackground=ACTION_BORDER, highlightthickness=1,
+        )
+        inner.pack(anchor=tk.W, padx=(0, 70), fill=tk.X)
+
+        row = tk.Frame(inner, bg=ACTION_BG)
+        row.pack(fill=tk.X, anchor=tk.W)
+
+        tk.Label(
+            row, text=icon, bg=ACTION_BG, fg=icon_color,
+            font=("Segoe UI", 13, "bold"), padx=4,
+        ).pack(side=tk.LEFT)
+
+        tk.Label(
+            row, text=title, bg=ACTION_BG, fg=ACTION_FG,
+            font=FONT_MSG, anchor="w", justify=tk.LEFT,
+            wraplength=MSG_WRAP_WIDTH - 60,
+        ).pack(side=tk.LEFT, fill=tk.X, expand=True, padx=(2, 0))
+
+        if meta:
+            tk.Label(
+                inner, text=f"{meta}  •  {now}",
+                bg=ACTION_BG, fg=ACTION_META_FG,
+                font=FONT_TIMESTAMP, anchor="w",
+            ).pack(fill=tk.X, anchor=tk.W, pady=(2, 0))
+
+    # ------------------------------------------------------------------
+    # Bulle paused_need_help interactive (J3.5)
+    # ------------------------------------------------------------------
+
+    def _add_paused_bubble(self, payload: Dict[str, Any]) -> None:
+        """Ajouter une bulle paused interactive (asset démo : Léa demande de l'aide)."""
+        if self._root is None:
+            return
+        self._root.after(0, lambda: self._render_paused_bubble(payload))
+
+    def _render_paused_bubble(self, payload: Dict[str, Any]) -> None:
+        tk = self._tk
+        if getattr(self, "_msg_frame", None) is None:
+            return
+
+        replay_id = str(payload.get("replay_id", "") or "")
+        workflow = payload.get("workflow", "?")
+        reason = payload.get("reason") or "Action incertaine — j'ai besoin de votre validation."
+        completed = payload.get("completed", 0)
+        total = payload.get("total", "?")
+        now = datetime.now().strftime("%H:%M")
+
+        container = tk.Frame(self._msg_frame, bg=BG_COLOR)
+        container.pack(fill=tk.X, padx=MARGIN, pady=6)
+
+        inner = tk.Frame(
+            container, bg=PAUSED_BG, padx=14, pady=12,
+            highlightbackground=PAUSED_BORDER, highlightthickness=2,
+        )
+        inner.pack(anchor=tk.W, padx=(0, 50), fill=tk.X)
+
+        tk.Label(
+            inner, text=f"⏸  Pause supervisée • {now}",
+            bg=PAUSED_BG, fg=PAUSED_FG,
+            font=("Segoe UI", 12, "bold"), anchor="w",
+        ).pack(fill=tk.X, anchor=tk.W)
+
+        tk.Label(
+            inner, text=reason, bg=PAUSED_BG, fg=PAUSED_FG,
+            font=FONT_MSG, wraplength=MSG_WRAP_WIDTH - 30,
+            anchor="w", justify=tk.LEFT,
+        ).pack(fill=tk.X, anchor=tk.W, pady=(6, 0))
+
+        tk.Label(
+            inner, text=f"{workflow} — étape {completed}/{total}",
+            bg=PAUSED_BG, fg=TIMESTAMP_FG, font=FONT_TIMESTAMP, anchor="w",
+        ).pack(fill=tk.X, anchor=tk.W, pady=(4, 8))
+
+        btn_frame = tk.Frame(inner, bg=PAUSED_BG)
+        btn_frame.pack(fill=tk.X, anchor=tk.W)
+
+        btn_resume = tk.Button(
+            btn_frame, text="Continuer",
+            bg=PAUSED_BTN_RESUME_BG, fg="white", font=FONT_QUICK_BTN,
+            padx=14, pady=4, bd=0, cursor="hand2",
+            activebackground=PAUSED_BTN_RESUME_HOVER, activeforeground="white",
+            command=lambda: self._on_paused_resume(replay_id),
+        )
+        btn_resume.pack(side=tk.LEFT, padx=(0, 8))
+
+        btn_abort = tk.Button(
+            btn_frame, text="Annuler",
+            bg=PAUSED_BTN_ABORT_BG, fg="white", font=FONT_QUICK_BTN,
+            padx=14, pady=4, bd=0, cursor="hand2",
+            activebackground=PAUSED_BTN_ABORT_HOVER, activeforeground="white",
+            command=lambda: self._on_paused_abort(replay_id),
+        )
+        btn_abort.pack(side=tk.LEFT)
+
+        self._active_paused_bubble = {
+            "container": container, "inner": inner,
+            "btn_resume": btn_resume, "btn_abort": btn_abort,
+            "replay_id": replay_id,
+        }
+
+    def _close_active_paused_bubble(self, reason: str) -> None:
+        if self._active_paused_bubble is None or self._root is None:
+            return
+        self._root.after(0, lambda: self._do_close_paused_bubble(reason))
+
+    def _do_close_paused_bubble(self, reason: str) -> None:
+        bubble = self._active_paused_bubble
+        if bubble is None:
+            return
+        try:
+            bubble["btn_resume"].config(state="disabled")
+            bubble["btn_abort"].config(state="disabled")
+            label_text = {
+                "lea:resumed": "→ Reprise",
+                "lea:done": "→ Terminé",
+            }.get(reason, f"→ {reason}")
+            self._tk.Label(
+                bubble["inner"], text=label_text,
+                bg=PAUSED_BG, fg=PAUSED_FG, font=FONT_TIMESTAMP, anchor="w",
+            ).pack(fill="x", anchor="w", pady=(6, 0))
+        except Exception:
+            logger.debug("close paused bubble silenced", exc_info=True)
+        self._active_paused_bubble = None
+
+    def _on_paused_resume(self, replay_id: str) -> None:
+        if not replay_id or self._bus is None or not self._bus.connected:
+            self._add_lea_message("⚠ Bus indisponible — impossible de relancer")
+            return
+        self._bus.resume_replay(replay_id)
+        if self._active_paused_bubble:
+            try:
+                self._active_paused_bubble["btn_resume"].config(state="disabled")
+                self._active_paused_bubble["btn_abort"].config(state="disabled")
+            except Exception:
+                pass
+
+    def _on_paused_abort(self, replay_id: str) -> None:
+        if self._bus is None or not self._bus.connected:
+            self._add_lea_message("⚠ Bus indisponible — impossible d'annuler")
+            return
+        self._bus.abort_replay(replay_id)
+        if self._active_paused_bubble:
+            try:
+                self._active_paused_bubble["btn_resume"].config(state="disabled")
+                self._active_paused_bubble["btn_abort"].config(state="disabled")
+            except Exception:
+                pass
+
     # ======================================================================
     # Ajout de messages dans la zone de chat
     # ======================================================================

agent_v0/agent_v1/vision/capturer.py

@@ -15,7 +15,7 @@ import time
 import logging
 import hashlib
 import platform
-from typing import Any, Dict, Optional
+from typing import Any, Dict, List, Optional
 from PIL import Image, ImageFilter, ImageStat
 import mss
 from ..config import TARGETED_CROP_SIZE, SCREENSHOT_QUALITY, BLUR_SENSITIVE
@@ -26,6 +26,66 @@ logger = logging.getLogger(__name__)
 # OS courant (détecté une seule fois)
 _SYSTEM = platform.system()
 
+# QW1 — détection multi-écrans (fallback gracieux si screeninfo absent)
+try:
+    from screeninfo import get_monitors as _screeninfo_get_monitors
+    _SCREENINFO_AVAILABLE = True
+except ImportError:
+    _SCREENINFO_AVAILABLE = False
+
+
+def _get_monitors_geometry() -> List[Dict[str, Any]]:
+    """Retourne la liste des monitors physiques avec leurs offsets.
+
+    Returns:
+        List[dict] : [{idx, x, y, w, h, primary}, ...]. Vide si screeninfo
+        indisponible (le serveur tombera sur fallback composite).
+    """
+    if not _SCREENINFO_AVAILABLE:
+        return []
+    try:
+        monitors = _screeninfo_get_monitors()
+        return [
+            {
+                "idx": i,
+                "x": int(m.x),
+                "y": int(m.y),
+                "w": int(m.width),
+                "h": int(m.height),
+                "primary": bool(getattr(m, "is_primary", False)),
+            }
+            for i, m in enumerate(monitors)
+        ]
+    except Exception:
+        return []
+
+
+def _get_active_monitor_index() -> Optional[int]:
+    """Retourne l'index logique du monitor où se trouve le curseur (focus actif).
+
+    Returns:
+        int ou None si indéterminable.
+    """
+    if not _SCREENINFO_AVAILABLE:
+        return None
+    try:
+        import pyautogui  # import paresseux : évite la dépendance dure
+        cx, cy = pyautogui.position()
+        for i, m in enumerate(_screeninfo_get_monitors()):
+            if m.x <= cx < m.x + m.width and m.y <= cy < m.y + m.height:
+                return i
+    except Exception:
+        return None
+    return None
+
+
+def _enrich_with_monitor_info(payload: dict) -> dict:
+    """Ajoute monitor_index et monitors_geometry au payload (in-place + return)."""
+    if isinstance(payload, dict):
+        payload["monitor_index"] = _get_active_monitor_index()
+        payload["monitors_geometry"] = _get_monitors_geometry()
+    return payload
+
 class VisionCapturer:
     def __init__(self, session_dir: str):
         self.session_dir = session_dir
@@ -121,6 +181,9 @@ class VisionCapturer:
                 if window_info:
                     result["window_capture"] = window_info
 
+                # QW1 — enrichissement multi-écrans (additif, fallback gracieux)
+                _enrich_with_monitor_info(result)
+
                 return result
         except Exception as e:
             logger.error(f"Erreur Dual Capture: {e}")
@@ -223,6 +286,9 @@ class VisionCapturer:
                 "click_inside_window": click_inside,
             }
 
+            # QW1 — enrichissement multi-écrans (additif)
+            _enrich_with_monitor_info(result)
+
             logger.debug(
                 f"Fenêtre capturée : {title} ({win_w}x{win_h}) — "
                 f"clic relatif ({click_rel_x}, {click_rel_y})"

agent_v0/deploy/windows_client/agent_v1/core/executor.py

@@ -512,6 +512,21 @@ class ActionExecutorV1:
             x_pct = action.get("x_pct", 0.0)
             y_pct = action.get("y_pct", 0.0)
 
+            # QW1 — Si le serveur a résolu un monitor cible (idx >= 0),
+            # appliquer son offset aux coords absolues. Pour idx == -1
+            # (composite_fallback), aucun offset (backward compat).
+            # Le calcul des coords reste percent * (width/height) du monitor[1]
+            # côté client (x_pct est exprimé sur l'écran physique principal).
+            mon_res = action.get("monitor_resolution") or {}
+            mon_idx = mon_res.get("idx", -1)
+            mon_offset_x = mon_res.get("offset_x", 0) if mon_idx >= 0 else 0
+            mon_offset_y = mon_res.get("offset_y", 0) if mon_idx >= 0 else 0
+            if mon_idx >= 0 and (mon_offset_x or mon_offset_y):
+                logger.info(
+                    f"[REPLAY] QW1 monitor cible idx={mon_idx} source={mon_res.get('source')} "
+                    f"offset=({mon_offset_x},{mon_offset_y}) — appliqué aux coords"
+                )
+
             # ── Diagnostic résolution ──
             logger.info(
                 f"[REPLAY] Action {action_id} ({action_type}) — "
@@ -578,8 +593,8 @@ class ActionExecutorV1:
                         print(f"    [OBSERVER] Popup détectée : '{popup_label}' — fermeture")
                         logger.info(f"Observer : popup '{popup_label}' détectée avant résolution")
                         if popup_coords:
-                            real_x = int(popup_coords["x_pct"] * width)
+                            real_x = int(popup_coords["x_pct"] * width) + mon_offset_x
-                            real_y = int(popup_coords["y_pct"] * height)
+                            real_y = int(popup_coords["y_pct"] * height) + mon_offset_y
                             self._click((real_x, real_y), "left")
                             time.sleep(1.0)
                             print(f"    [OBSERVER] Popup fermée — reprise du flow normal")
@@ -718,8 +733,8 @@ class ActionExecutorV1:
                         self.notifier.replay_target_not_found(target_desc)
                         return result
 
-                real_x = int(x_pct * width)
+                real_x = int(x_pct * width) + mon_offset_x
-                real_y = int(y_pct * height)
+                real_y = int(y_pct * height) + mon_offset_y
                 button = action.get("button", "left")
                 mode = "VISUAL" if result.get("visual_resolved") else "COORD"
                 print(
@@ -781,8 +796,8 @@ class ActionExecutorV1:
                     print(f"    [TYPE] raw_keys disponibles ({len(raw_keys)} events) — replay exact")
                 # Cliquer sur le champ avant de taper (si coordonnees disponibles)
                 if x_pct > 0 and y_pct > 0:
-                    real_x = int(x_pct * width)
+                    real_x = int(x_pct * width) + mon_offset_x
-                    real_y = int(y_pct * height)
+                    real_y = int(y_pct * height) + mon_offset_y
                     print(f"    [TYPE] Clic prealable sur ({real_x}, {real_y})")
                     self._click((real_x, real_y), "left")
                     time.sleep(0.3)
@@ -808,8 +823,8 @@ class ActionExecutorV1:
                 logger.info(f"Replay key_combo : {keys} (raw_keys={'oui' if raw_keys else 'non'})")
 
             elif action_type == "scroll":
-                real_x = int(x_pct * width) if x_pct > 0 else int(0.5 * width)
+                real_x = (int(x_pct * width) if x_pct > 0 else int(0.5 * width)) + mon_offset_x
-                real_y = int(y_pct * height) if y_pct > 0 else int(0.5 * height)
+                real_y = (int(y_pct * height) if y_pct > 0 else int(0.5 * height)) + mon_offset_y
                 delta = action.get("delta", -3)
                 print(f"    [SCROLL] delta={delta} a ({real_x}, {real_y})")
                 self.mouse.position = (real_x, real_y)
@@ -1386,6 +1401,16 @@ Example: x_pct=0.50, y_pct=0.30"""
             data = resp.json()
             action = data.get("action")
             if action is None:
+                # pause_for_human : afficher le message de décision à l'utilisateur
+                if data.get("replay_paused") and data.get("pause_message"):
+                    msg = data["pause_message"]
+                    print(f"[PAUSE] {msg}")
+                    logger.info(f"Replay en pause — message : {msg}")
+                    self.notifier.notify(
+                        title="Léa — Validation requise",
+                        message=msg[:250],
+                        timeout=30,
+                    )
                 return False
 
         except (requests.exceptions.ConnectionError, requests.exceptions.Timeout) as e:

agent_v0/deploy/windows_client/agent_v1/main.py

@@ -319,7 +319,22 @@ class AgentV1:
                     if img_hash != self._last_heartbeat_hash:
                         self._last_heartbeat_hash = img_hash
                         self.streamer.push_image(full_path, f"heartbeat_{int(time.time())}")
-                        self.streamer.push_event({"type": "heartbeat", "image": full_path, "timestamp": time.time(), "machine_id": self.machine_id})
+                        heartbeat_event = {
+                            "type": "heartbeat",
+                            "image": full_path,
+                            "timestamp": time.time(),
+                            "machine_id": self.machine_id,
+                        }
+                        # QW1 — enrichissement multi-écrans (monitor_index + monitors_geometry)
+                        # Additif, fallback gracieux : sans cet enrichissement, le serveur
+                        # ne reçoit l'info qu'au moment des clics, donc QW1 ne s'active
+                        # pas en continu sur poste Windows multi-écrans.
+                        try:
+                            from .vision.capturer import _enrich_with_monitor_info
+                            _enrich_with_monitor_info(heartbeat_event)
+                        except Exception as e:
+                            logger.debug("QW1 enrichissement heartbeat échoué: %s", e)
+                        self.streamer.push_event(heartbeat_event)
             except Exception as e:
                 logger.error(f"Heartbeat error: {e}")
             time.sleep(5)

agent_v0/deploy/windows_client/agent_v1/vision/capturer.py

@@ -8,12 +8,73 @@ import os
 import time
 import logging
 import hashlib
+from typing import Any, Dict, List, Optional
 from PIL import Image, ImageFilter, ImageStat
 import mss
 from ..config import TARGETED_CROP_SIZE, SCREENSHOT_QUALITY
 
 logger = logging.getLogger(__name__)
 
+# QW1 — détection multi-écrans (fallback gracieux si screeninfo absent)
+try:
+    from screeninfo import get_monitors as _screeninfo_get_monitors
+    _SCREENINFO_AVAILABLE = True
+except ImportError:
+    _SCREENINFO_AVAILABLE = False
+
+
+def _get_monitors_geometry() -> List[Dict[str, Any]]:
+    """Retourne la liste des monitors physiques avec leurs offsets.
+
+    Returns:
+        List[dict] : [{idx, x, y, w, h, primary}, ...]. Vide si screeninfo
+        indisponible (le serveur tombera sur fallback composite).
+    """
+    if not _SCREENINFO_AVAILABLE:
+        return []
+    try:
+        monitors = _screeninfo_get_monitors()
+        return [
+            {
+                "idx": i,
+                "x": int(m.x),
+                "y": int(m.y),
+                "w": int(m.width),
+                "h": int(m.height),
+                "primary": bool(getattr(m, "is_primary", False)),
+            }
+            for i, m in enumerate(monitors)
+        ]
+    except Exception:
+        return []
+
+
+def _get_active_monitor_index() -> Optional[int]:
+    """Retourne l'index logique du monitor où se trouve le curseur (focus actif).
+
+    Returns:
+        int ou None si indéterminable.
+    """
+    if not _SCREENINFO_AVAILABLE:
+        return None
+    try:
+        import pyautogui  # import paresseux : évite la dépendance dure
+        cx, cy = pyautogui.position()
+        for i, m in enumerate(_screeninfo_get_monitors()):
+            if m.x <= cx < m.x + m.width and m.y <= cy < m.y + m.height:
+                return i
+    except Exception:
+        return None
+    return None
+
+
+def _enrich_with_monitor_info(payload: dict) -> dict:
+    """Ajoute monitor_index et monitors_geometry au payload (in-place + return)."""
+    if isinstance(payload, dict):
+        payload["monitor_index"] = _get_active_monitor_index()
+        payload["monitors_geometry"] = _get_monitors_geometry()
+    return payload
+
 class VisionCapturer:
     def __init__(self, session_dir: str):
         self.session_dir = session_dir
@@ -72,7 +133,12 @@ class VisionCapturer:
                 # Mise à jour du hash pour le prochain heartbeat
                 self.last_img_hash = self._compute_quick_hash(img)
 
-                return {"full": full_path, "crop": crop_path}
+                result = {"full": full_path, "crop": crop_path}
+
+                # QW1 — enrichissement multi-écrans (additif, fallback gracieux)
+                _enrich_with_monitor_info(result)
+
+                return result
         except Exception as e:
             logger.error(f"Erreur Dual Capture: {e}")
             return {}

agent_v0/deploy/windows_client/requirements.txt

@@ -3,7 +3,9 @@ mss>=9.0.1              # Capture d'écran haute performance
 pynput>=1.7.7           # Clavier/Souris Cross-plateforme
 Pillow>=10.0.0          # Crops et processing image
 requests>=2.31.0        # Streaming réseau
+python-socketio[client]>=5.10,<6.0  # Bus feedback Léa 'lea:*' (compat Flask-SocketIO 5.3.x serveur)
 psutil>=5.9.0           # Monitoring CPU/RAM
+screeninfo>=0.8         # QW1 — détection des monitors physiques + offsets
 pystray>=0.19.5         # Icône Tray UI
 plyer>=2.1.0            # Notifications toast natives (remplace PyQt5)
 

agent_v0/server_v1/api_stream.py

@@ -33,6 +33,8 @@ from .audit_trail import AuditTrail, AuditEntry
 from .agent_registry import AgentRegistry, AgentAlreadyEnrolledError
 from .stream_processor import StreamProcessor, build_replay_from_raw_events, enrich_click_from_screenshot
 from .worker_stream import StreamWorker
+from .monitor_router import resolve_target_monitor  # QW1 — résolution écran cible
+from .loop_detector import LoopDetector  # QW2 — détection de boucle pendant replay
 from .execution_plan_runner import (
     execution_plan_to_actions,
     inject_plan_into_queue,
@@ -219,6 +221,11 @@ from .replay_engine import (
     _is_learned_workflow,
     _edge_to_normalized_actions,
     _substitute_variables,
+    _resolve_runtime_vars,
+    _SERVER_SIDE_ACTION_TYPES,
+    _handle_extract_text_action,
+    _handle_extract_table_action,
+    _handle_t2a_decision_action,
     _expand_compound_steps,
     _pre_check_screen_state as _pre_check_screen_state_impl,
     _detect_popup_hint as _detect_popup_hint_impl,
@@ -355,6 +362,18 @@ REPLAY_LOCK_FILE = _DATA_DIR / "_replay_active.lock"
 processor = StreamProcessor(data_dir=str(LIVE_SESSIONS_DIR))
 worker = StreamWorker(live_dir=str(LIVE_SESSIONS_DIR), processor=processor)
 
+# QW2 — LoopDetector singleton lazy (utilise le CLIP embedder du processor)
+_loop_detector: Optional["LoopDetector"] = None
+
+
+def _get_loop_detector() -> "LoopDetector":
+    """Singleton lazy — crée le LoopDetector avec le CLIP embedder du processor."""
+    global _loop_detector
+    if _loop_detector is None:
+        embedder = getattr(processor, "_clip_embedder", None)
+        _loop_detector = LoopDetector(clip_embedder=embedder)
+    return _loop_detector
+
 # Registre des postes Lea enroles (table enrolled_agents dans rpa_data.db)
 # Emplacement configurable via RPA_AGENTS_DB_PATH pour les tests.
 _AGENTS_DB_PATH = os.environ.get(
@@ -507,6 +526,7 @@ class ReplayRequest(BaseModel):
     session_id: str
     machine_id: Optional[str] = None  # Machine cible pour le replay (multi-machine)
     params: Optional[Dict[str, Any]] = None
+    variables: Optional[Dict[str, Any]] = None  # Variables runtime initiales (templating {{var}})
 
 
 class RawReplayRequest(BaseModel):
@@ -761,6 +781,21 @@ async def startup():
     _cleanup_thread = threading.Thread(target=_cleanup_loop, daemon=True, name="replay_cleanup")
     _cleanup_thread.start()
 
+    # Préchargement EasyOCR en arrière-plan : sans ça, le 1er extract_text /
+    # extract_table déclenche un cold start de ~3-5s qui bloque l'event loop
+    # FastAPI (constaté 2026-05-05 : streaming server inaccessible 2 min).
+    # Le thread tourne pendant que le boot continue ; le 1er appel OCR sera rapide.
+    def _preload_easyocr():
+        try:
+            t0 = time.time()
+            from core.llm.ocr_extractor import _get_reader
+            _get_reader()
+            logger.info("[OCR] EasyOCR préchargé (fr+en, CPU) en %.1fs", time.time() - t0)
+        except Exception as e:
+            logger.warning("[OCR] Échec préchargement EasyOCR : %s", e)
+
+    threading.Thread(target=_preload_easyocr, daemon=True, name="preload_easyocr").start()
+
     logger.info(
         "API Streaming démarrée — StreamProcessor, Worker et Cleanup prêts. "
         "VLM Worker dans un process séparé (run_worker.py)."
@@ -1958,6 +1993,11 @@ async def start_replay(request: ReplayRequest):
             machine_id=resolved_machine_id,
             actions=actions,
         )
+        # Pré-injection des variables runtime (templating {{var}} sur by_text,
+        # text, target_spec.* etc.). Permet à l'orchestrateur d'appeler ce
+        # workflow avec p.ex. variables={"patient_id": "25003284"} pour boucler.
+        if request.variables:
+            _replay_states[replay_id]["variables"].update(request.variables)
         # Enregistrer le mapping machine -> session pour le replay ciblé
         if resolved_machine_id and resolved_machine_id != "default":
             _machine_replay_target[resolved_machine_id] = session_id
@@ -2758,8 +2798,29 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
 
     Si la session de l'agent n'a pas d'actions en attente, cherche dans les
     autres queues de la MÊME machine (pas cross-machine).
+
+    Acquire timeout : si une action serveur lente (extract_text OCR,
+    t2a_decision LLM) tient le lock, on retourne immédiatement
+    {action: None, server_busy: True} avant que le client ne timeout à 5s.
+    Sans cela, des actions seraient popped serveur puis envoyées sur des
+    sockets clients déjà fermées par timeout — perdues silencieusement.
+
+    L'acquire et les actions serveur lentes sont exécutés via
+    run_in_executor : sinon l'appel synchrone bloque l'event loop FastAPI
+    (single-threaded) et même les polls qui devraient recevoir server_busy
+    sont bloqués jusqu'à libération — ce qui annule l'effet du timeout.
     """
-    with _replay_lock:
+    import asyncio
+    loop = asyncio.get_event_loop()
+    acquired = await loop.run_in_executor(None, _replay_lock.acquire, True, 4.5)
+    if not acquired:
+        return {
+            "action": None,
+            "session_id": session_id,
+            "machine_id": machine_id,
+            "server_busy": True,
+        }
+    try:
         # Verifier si le replay est en pause supervisee (target_not_found).
         # Dans ce cas, NE PAS envoyer d'action — attendre l'intervention utilisateur.
         for state in _replay_states.values():
@@ -2824,6 +2885,7 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
                             break
                     if target_state:
                         queue = target_queue
+                        owning_replay = target_state
                         _replay_queues[session_id] = target_queue
                         del _replay_queues[target_sid]
                         target_state["session_id"] = session_id
@@ -2840,6 +2902,7 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
                         other_queue = _replay_queues.get(other_sid, [])
                         if other_queue:
                             queue = other_queue
+                            owning_replay = state
                             _replay_queues[session_id] = other_queue
                             del _replay_queues[other_sid]
                             state["session_id"] = session_id
@@ -2850,8 +2913,132 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
         if not queue:
             return {"action": None, "session_id": session_id, "machine_id": machine_id}
 
-        # Peek à la prochaine action SANS la retirer (pour le pre-check)
+        # ── Boucle de traitement : actions serveur (extract_text, t2a_decision)
-        action = queue[0]
+        # exécutées entièrement côté serveur jusqu'à trouver une action visuelle
+        # à transmettre à l'Agent V1 ou un pause_for_human qui bloque le replay.
+        action = None
+        while queue:
+            action = queue[0]
+
+            # Résoudre les variables runtime ({{var}} et {{var.field}})
+            if owning_replay is not None:
+                runtime_vars = owning_replay.get("variables") or {}
+                if runtime_vars:
+                    action = _resolve_runtime_vars(action, runtime_vars)
+
+            type_ = action.get("type")
+
+            # pause_for_human : pause supervisée si safety_level/safety_checks ou mode supervised,
+            # sinon no-op en mode autonome (skip).
+            if type_ == "pause_for_human":
+                _params = action.get("parameters") or {}
+                _exec_mode = (
+                    (owning_replay or {}).get("params", {}).get("execution_mode", "autonomous")
+                    if owning_replay else "autonomous"
+                )
+                _has_safety_decl = bool(_params.get("safety_level") or _params.get("safety_checks"))
+                _is_supervised = _exec_mode != "autonomous"
+
+                if owning_replay is not None and (_has_safety_decl or _is_supervised):
+                    # QW4 — Construire le payload de pause enrichi (déclaratif + LLM contextuel)
+                    try:
+                        from agent_v0.server_v1.safety_checks_provider import build_pause_payload
+                        last_screenshot_path = owning_replay.get("last_screenshot")
+                        payload = build_pause_payload(action, owning_replay, last_screenshot_path)
+                        owning_replay["safety_checks"] = payload.checks
+                        owning_replay["pause_payload"] = {
+                            "checks": payload.checks,
+                            "pause_reason": payload.pause_reason,
+                            "message": payload.message,
+                        }
+                        if payload.message:
+                            owning_replay["pause_message"] = payload.message
+                        # Bus event d'observabilité (pattern QW1/QW2 = logger.info)
+                        logger.info(
+                            "[BUS] lea:safety_checks_generated replay=%s count=%d sources=%s",
+                            owning_replay.get("replay_id", "?"),
+                            len(payload.checks),
+                            [c["source"] for c in payload.checks],
+                        )
+                    except Exception as e:
+                        logger.warning("QW4 build_pause_payload échec (%s) — pause sans checks", e)
+                        owning_replay["safety_checks"] = []
+
+                    # Conserver le contexte de l'action (audit + reprise)
+                    owning_replay["failed_action"] = {
+                        "action_id": action.get("action_id"),
+                        "type": "pause_for_human",
+                        "reason": "user_request",
+                    }
+                    owning_replay["status"] = "paused_need_help"
+                    queue.pop(0)
+                    _replay_queues[session_id] = queue
+                    return {"action": None, "session_id": session_id, "machine_id": machine_id}
+
+                # Mode autonome sans safety_checks → skip (comportement legacy)
+                logger.info(
+                    "pause_for_human ignorée (mode autonome) — replay %s continue",
+                    owning_replay["replay_id"] if owning_replay else "?"
+                )
+                queue.pop(0)
+                _replay_queues[session_id] = queue
+                continue
+
+            # Actions serveur : exécuter HORS event loop pour ne pas bloquer
+            # les autres polls (extract_text OCR ~5s, t2a_decision LLM ~8-13s).
+            # Le lock reste tenu (queue cohérente) mais l'event loop est libre,
+            # donc les polls concurrents peuvent recevoir {server_busy: True}.
+            if type_ in _SERVER_SIDE_ACTION_TYPES and owning_replay is not None:
+                try:
+                    if type_ == "extract_text":
+                        await loop.run_in_executor(
+                            None,
+                            _handle_extract_text_action,
+                            action, owning_replay, session_id, _last_heartbeat,
+                        )
+                    elif type_ == "extract_table":
+                        await loop.run_in_executor(
+                            None,
+                            _handle_extract_table_action,
+                            action, owning_replay, session_id, _last_heartbeat,
+                        )
+                    elif type_ == "t2a_decision":
+                        await loop.run_in_executor(
+                            None,
+                            _handle_t2a_decision_action,
+                            action, owning_replay,
+                        )
+                except Exception as e:
+                    logger.warning(f"Action serveur {type_} a levé : {e}")
+                queue.pop(0)
+                _replay_queues[session_id] = queue
+                continue  # action suivante
+
+            # Clic conditionnel : si l'action a un paramètre "condition", évaluer la variable
+            # Format : "dec.critere1_valide" → runtime_vars["dec"]["critere1_valide"]
+            condition_key = (action.get("parameters") or {}).get("condition")
+            if condition_key and owning_replay is not None:
+                runtime_vars = owning_replay.get("variables") or {}
+                parts = condition_key.split(".", 1)
+                if len(parts) == 2:
+                    val = (runtime_vars.get(parts[0]) or {}).get(parts[1])
+                else:
+                    val = runtime_vars.get(parts[0])
+                if not val:
+                    logger.info("Clic conditionnel ignoré (%s=%s) — action %s",
+                                condition_key, val, action.get("action_id", "?"))
+                    queue.pop(0)
+                    _replay_queues[session_id] = queue
+                    continue
+
+            # Action visuelle : sortir de la boucle pour la transmettre à l'Agent V1
+            break
+
+        # Si la queue s'est vidée après les exécutions serveur, rien à transmettre
+        if not queue or action is None:
+            return {"action": None, "session_id": session_id, "machine_id": machine_id}
+    finally:
+        _replay_lock.release()
 
     # ---- Pre-check écran (optionnel, non bloquant) ----
     # Ne s'applique qu'aux actions qui ont un from_node (actions de workflow,
@@ -3018,6 +3205,51 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
         f"{_precheck_sim}"
     )
 
+    # QW1 — Résoudre l'écran cible et joindre l'info à l'action
+    # Cascade : action.monitor_index → session.last_focused_monitor → composite_fallback
+    try:
+        session_qw1 = processor.session_manager.get_session(session_id)
+        last_window_info_qw1 = (
+            session_qw1.last_window_info if session_qw1 is not None else {}
+        ) or {}
+        session_state_qw1 = {
+            "monitors_geometry": last_window_info_qw1.get("monitors_geometry", []),
+            "last_focused_monitor": last_window_info_qw1.get("monitor_index"),
+        }
+        target = resolve_target_monitor(action, session_state_qw1)
+        action["monitor_resolution"] = {
+            "idx": target.idx,
+            "offset_x": target.offset_x,
+            "offset_y": target.offset_y,
+            "w": target.w,
+            "h": target.h,
+            "source": target.source,
+        }
+        # QW1 — Émission bus lea:monitor_routed (no-op si bus indisponible)
+        # Le serveur streaming n'a pas de SocketIO local : on logge en INFO
+        # bien lisible. Un consommateur (agent_chat / dashboard) peut tailer
+        # `journalctl -u rpa-streaming | grep '\[BUS\] lea:monitor_routed'`.
+        try:
+            _replay_id_bus = (
+                owning_replay.get("replay_id") if owning_replay else None
+            )
+            logger.info(
+                "[BUS] lea:monitor_routed replay=%s action=%s idx=%d source=%s "
+                "offset=(%d,%d) wh=(%d,%d)",
+                _replay_id_bus,
+                action.get("action_id"),
+                target.idx,
+                target.source,
+                target.offset_x,
+                target.offset_y,
+                target.w,
+                target.h,
+            )
+        except Exception as _e_bus:
+            logger.debug("emit lea:monitor_routed échec (non bloquant): %s", _e_bus)
+    except Exception as e:
+        logger.debug("QW1 monitor_resolution skip (%s)", e)
+
     response: Dict[str, Any] = {
         "action": action,
         "session_id": session_id,
@@ -3756,6 +3988,82 @@ async def report_action_result(report: ReplayResultReport):
                 f"— worker VLM autorisé à reprendre"
             )
 
+        # ===================================================================
+        # QW2 — LoopDetector : alimentation des anneaux + évaluation
+        # ===================================================================
+        # On n'évalue que si le replay est encore "running" — inutile de
+        # pauser quelque chose de déjà completed/error/paused.
+        if replay_state["status"] == "running":
+            # Snapshot image (PIL) dans l'anneau
+            try:
+                from PIL import Image
+                ss_raw = screenshot_after or replay_state.get("last_screenshot")
+                img = None
+                if isinstance(ss_raw, str) and ss_raw:
+                    if os.path.isfile(ss_raw):
+                        img = Image.open(ss_raw).copy()  # détache du file handle
+                    else:
+                        # Possible base64 — décoder
+                        try:
+                            import base64
+                            import io as _io
+                            img_bytes = base64.b64decode(ss_raw, validate=False)
+                            img = Image.open(_io.BytesIO(img_bytes)).copy()
+                        except Exception:
+                            img = None
+                if img is not None:
+                    replay_state.setdefault("_screenshot_history", []).append(img)
+                    replay_state["_screenshot_history"] = replay_state["_screenshot_history"][-5:]
+            except Exception as e:
+                logger.debug("LoopDetector: snapshot historique échoué: %s", e)
+
+            # Snapshot signature de l'action courante
+            try:
+                _act_pos = report.actual_position or {}
+                action_sig = {
+                    "type": (original_action or {}).get("type")
+                            or replay_state.get("_last_action_type", ""),
+                    "x_pct": _act_pos.get("x_pct") if isinstance(_act_pos, dict)
+                             else (original_action or {}).get("x_pct"),
+                    "y_pct": _act_pos.get("y_pct") if isinstance(_act_pos, dict)
+                             else (original_action or {}).get("y_pct"),
+                }
+                replay_state.setdefault("_action_history", []).append(action_sig)
+                replay_state["_action_history"] = replay_state["_action_history"][-5:]
+            except Exception as e:
+                logger.debug("LoopDetector: snapshot action_sig échoué: %s", e)
+
+            # Évaluation (silencieux si rien)
+            try:
+                verdict = _get_loop_detector().evaluate(
+                    replay_state,
+                    screenshots=replay_state.get("_screenshot_history", []),
+                    actions=replay_state.get("_action_history", []),
+                )
+                if verdict.detected:
+                    replay_state["status"] = "paused_need_help"
+                    replay_state["pause_reason"] = "loop_detected"
+                    replay_state["pause_message"] = (
+                        f"Léa semble bloquée — {verdict.signal} "
+                        f"(détail: {verdict.evidence})"
+                    )
+                    logger.warning(
+                        "LoopDetector: replay %s mis en pause — signal=%s evidence=%s",
+                        replay_state["replay_id"], verdict.signal, verdict.evidence,
+                    )
+                    # Bus event d'observabilité (logger pattern QW1)
+                    try:
+                        logger.info(
+                            "[BUS] lea:loop_detected replay=%s signal=%s evidence=%s",
+                            replay_state["replay_id"],
+                            verdict.signal,
+                            verdict.evidence,
+                        )
+                    except Exception as _e_bus:
+                        logger.debug("emit lea:loop_detected échec: %s", _e_bus)
+            except Exception as e:
+                logger.warning("LoopDetector: évaluation échouée (non bloquant): %s", e)
+
     return {
         "status": "recorded",
         "action_id": action_id,
@@ -3842,8 +4150,16 @@ async def list_replays():
         }
 
 
+class ReplayResumeRequest(BaseModel):
+    """Body optionnel pour /replay/resume — QW4 acquittement de safety_checks."""
+    acknowledged_check_ids: List[str] = []
+
+
 @app.post("/api/v1/traces/stream/replay/{replay_id}/resume")
-async def resume_replay(replay_id: str):
+async def resume_replay(
+    replay_id: str,
+    payload: Optional[ReplayResumeRequest] = None,
+):
     """Reprendre un replay en pause supervisee (paused_need_help).
 
     L'utilisateur a intervenu manuellement (naviguer vers le bon ecran,
@@ -3851,6 +4167,10 @@ async def resume_replay(replay_id: str):
     est reinjectee en tete de queue pour etre re-tentee.
 
     Si le replay n'est pas en pause, retourne une erreur 409 (conflit).
+
+    QW4 — Si des safety_checks sont attachés à la pause, tous ceux marqués
+    `required` doivent figurer dans `acknowledged_check_ids`. Sinon → 400
+    avec `{"error": "required_checks_missing", "missing": [...]}`.
     """
     with _replay_lock:
         state = _replay_states.get(replay_id)
@@ -3869,6 +4189,25 @@ async def resume_replay(replay_id: str):
                 ),
             )
 
+        # QW4 — Vérification des safety_checks required avant reprise
+        safety_checks = state.get("safety_checks") or []
+        ack_ids = (payload.acknowledged_check_ids if payload else []) or []
+        if safety_checks:
+            required_ids = {c["id"] for c in safety_checks if c.get("required")}
+            ack_set = set(ack_ids)
+            missing = sorted(required_ids - ack_set)
+            if missing:
+                raise HTTPException(
+                    status_code=400,
+                    detail={"error": "required_checks_missing", "missing": missing},
+                )
+            # Audit trail
+            state["checks_acknowledged"] = sorted(ack_set)
+            logger.info(
+                "QW4 resume replay=%s acquittements=%d (%s)",
+                state.get("replay_id"), len(ack_set), sorted(ack_set),
+            )
+
         # Recuperer l'action echouee pour la reinjecter
         failed_action = state.get("failed_action")
         session_id = state["session_id"]
@@ -3877,9 +4216,15 @@ async def resume_replay(replay_id: str):
         state["status"] = "running"
         state["failed_action"] = None
         state["pause_message"] = None
+        # QW4 — vider safety_checks après acquittement (la pause est résolue)
+        state["safety_checks"] = []
+        state["pause_payload"] = None
+        state["pause_reason"] = ""
 
         # Reinjecter l'action echouee en tete de queue (sera re-tentee)
-        if failed_action and failed_action.get("action_id"):
+        # pause_for_human est une pause intentionnelle, pas une erreur — ne pas réinjecter
+        if (failed_action and failed_action.get("action_id")
+                and failed_action.get("reason") != "user_request"):
             # Reconstruire l'action a partir du retry_pending ou de l'original
             original_action_id = failed_action["action_id"]
             # Chercher l'action originale dans les retry_pending
@@ -3920,6 +4265,26 @@ async def resume_replay(replay_id: str):
     }
 
 
+@app.post("/api/v1/traces/stream/replay/{replay_id}/cancel")
+async def cancel_replay(replay_id: str):
+    """Annuler un replay (quel que soit son statut) et vider sa queue."""
+    with _replay_lock:
+        state = _replay_states.get(replay_id)
+        if not state:
+            raise HTTPException(status_code=404, detail=f"Replay '{replay_id}' non trouvé")
+        session_id = state["session_id"]
+        state["status"] = "cancelled"
+        state["failed_action"] = None
+        state["pause_message"] = None
+        _replay_queues[session_id] = []
+        keys_to_del = [k for k, v in _retry_pending.items() if v.get("replay_id") == replay_id]
+        for k in keys_to_del:
+            _retry_pending.pop(k, None)
+
+    logger.info("Replay %s annulé manuellement", replay_id)
+    return {"status": "cancelled", "replay_id": replay_id, "session_id": session_id}
+
+
 # =========================================================================
 # Visual Replay — Résolution visuelle des cibles (module resolve_engine)
 # =========================================================================

agent_v0/server_v1/live_session_manager.py

@@ -256,6 +256,20 @@ class LiveSessionManager:
                     session.last_window_info["title"] = wc_title
                 if wc_app:
                     session.last_window_info["app_name"] = wc_app
+                # QW1 — propager monitor_index et monitors_geometry depuis window_capture
+                if "monitor_index" in window_capture:
+                    session.last_window_info["monitor_index"] = window_capture["monitor_index"]
+                if "monitors_geometry" in window_capture:
+                    session.last_window_info["monitors_geometry"] = window_capture["monitors_geometry"]
+
+            # QW1 — propager monitor_index/monitors_geometry du payload event
+            # (cas heartbeat enrichi sans window/window_title). Toujours
+            # rafraîchir le focus actif (change souvent) et la géométrie
+            # (l'utilisateur peut brancher/débrancher un écran).
+            if "monitor_index" in event_data:
+                session.last_window_info["monitor_index"] = event_data["monitor_index"]
+            if "monitors_geometry" in event_data and event_data["monitors_geometry"]:
+                session.last_window_info["monitors_geometry"] = event_data["monitors_geometry"]
 
             # Accumuler les titres/apps pour le nommage automatique
             title = session.last_window_info.get("title", "").strip()

agent_v0/server_v1/loop_detector.py

@@ -0,0 +1,154 @@
+# agent_v0/server_v1/loop_detector.py
+"""LoopDetector composite — détection de stagnation de Léa pendant un replay (QW2).
+
+Trois signaux indépendants :
+- screen_static : N captures consécutives avec CLIP similarity > seuil
+- action_repeat : N actions consécutives identiques (type + coords)
+- retry_threshold : nombre de retries cumulés >= seuil
+
+Un seul signal positif → verdict.detected=True. Le serveur bascule alors le
+replay en paused_need_help avec pause_reason explicite.
+
+Désactivable via env var RPA_LOOP_DETECTOR_ENABLED=0.
+"""
+
+import logging
+import os
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class LoopVerdict:
+    detected: bool = False
+    reason: str = ""
+    signal: str = ""  # "screen_static" | "action_repeat" | "retry_threshold" | ""
+    evidence: Dict[str, Any] = field(default_factory=dict)
+
+
+def _env_int(name: str, default: int) -> int:
+    try:
+        return int(os.environ.get(name, default))
+    except (TypeError, ValueError):
+        return default
+
+
+def _env_float(name: str, default: float) -> float:
+    try:
+        return float(os.environ.get(name, default))
+    except (TypeError, ValueError):
+        return default
+
+
+def _env_bool_enabled(name: str) -> bool:
+    val = os.environ.get(name, "1").strip().lower()
+    return val not in ("0", "false", "no", "off", "")
+
+
+def _cosine_similarity(a, b) -> float:
+    """Similarité cosine entre deux vecteurs (listes ou np.array). Robuste vecteur nul."""
+    import numpy as np
+    av = np.asarray(a, dtype=np.float32).flatten()
+    bv = np.asarray(b, dtype=np.float32).flatten()
+    na, nb = float(np.linalg.norm(av)), float(np.linalg.norm(bv))
+    if na < 1e-8 or nb < 1e-8:
+        return 0.0
+    return float(np.dot(av, bv) / (na * nb))
+
+
+class LoopDetector:
+    def __init__(self, clip_embedder=None):
+        self.clip_embedder = clip_embedder
+
+    def evaluate(
+        self,
+        state: Dict[str, Any],
+        screenshots: List[Any],
+        actions: List[Dict[str, Any]],
+    ) -> LoopVerdict:
+        """Évalue les 3 signaux. Retourne le premier déclenché.
+
+        Args:
+            state: replay_state (utilisé pour retried_actions)
+            screenshots: anneau d'embeddings CLIP (les N derniers)
+            actions: anneau des N dernières actions exécutées
+        """
+        if not _env_bool_enabled("RPA_LOOP_DETECTOR_ENABLED"):
+            return LoopVerdict(detected=False)
+
+        # Signal A : screen_static
+        verdict = self._check_screen_static(screenshots)
+        if verdict.detected:
+            return verdict
+
+        # Signal B : action_repeat
+        verdict = self._check_action_repeat(actions)
+        if verdict.detected:
+            return verdict
+
+        # Signal C : retry_threshold
+        verdict = self._check_retry_threshold(state)
+        if verdict.detected:
+            return verdict
+
+        return LoopVerdict(detected=False)
+
+    def _check_screen_static(self, screenshots: List[Any]) -> LoopVerdict:
+        n_required = _env_int("RPA_LOOP_SCREEN_STATIC_N", 4)
+        threshold = _env_float("RPA_LOOP_SCREEN_STATIC_THRESHOLD", 0.99)
+
+        if self.clip_embedder is None or len(screenshots) < n_required:
+            return LoopVerdict()
+
+        try:
+            recent = screenshots[-n_required:]
+            # Embed chaque capture via le CLIP embedder (peut lever)
+            embeddings = [self.clip_embedder.embed_image(img) for img in recent]
+            sims = [_cosine_similarity(embeddings[i], embeddings[i + 1])
+                    for i in range(len(embeddings) - 1)]
+            min_sim = min(sims)
+            if min_sim > threshold:
+                return LoopVerdict(
+                    detected=True,
+                    reason="loop_detected",
+                    signal="screen_static",
+                    evidence={"min_similarity": round(min_sim, 4),
+                              "n_captures": n_required,
+                              "threshold": threshold},
+                )
+        except Exception as e:
+            logger.warning("LoopDetector signal_A erreur (%s) — signal inerte ce tick", e)
+        return LoopVerdict()
+
+    def _check_action_repeat(self, actions: List[Dict[str, Any]]) -> LoopVerdict:
+        n_required = _env_int("RPA_LOOP_ACTION_REPEAT_N", 3)
+        if len(actions) < n_required:
+            return LoopVerdict()
+        recent = actions[-n_required:]
+
+        def _signature(a: Dict[str, Any]) -> tuple:
+            return (a.get("type"), a.get("x_pct"), a.get("y_pct"))
+
+        sigs = [_signature(a) for a in recent]
+        if all(s == sigs[0] for s in sigs):
+            return LoopVerdict(
+                detected=True,
+                reason="loop_detected",
+                signal="action_repeat",
+                evidence={"signature": sigs[0], "count": n_required},
+            )
+        return LoopVerdict()
+
+    def _check_retry_threshold(self, state: Dict[str, Any]) -> LoopVerdict:
+        threshold = _env_int("RPA_LOOP_RETRY_THRESHOLD", 3)
+        retried = int(state.get("retried_actions", 0))
+        if retried >= threshold:
+            return LoopVerdict(
+                detected=True,
+                reason="loop_detected",
+                signal="retry_threshold",
+                evidence={"retried_actions": retried, "threshold": threshold},
+            )
+        return LoopVerdict()

agent_v0/server_v1/monitor_router.py

@@ -0,0 +1,99 @@
+# agent_v0/server_v1/monitor_router.py
+"""MonitorRouter — résolution de l'écran cible pour le replay (QW1).
+
+Stratégie en cascade :
+1. action.monitor_index (hérité de la session source) → cible cet écran
+2. session.last_focused_monitor (focus actif vu en dernier heartbeat) → fallback
+3. composite (offset 0, 0) → backward compat
+
+Émet sur le bus lea:* l'event monitor_routed avec la source de la décision.
+"""
+
+import logging
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MonitorTarget:
+    """Représente l'écran cible résolu pour une action de replay."""
+    idx: int
+    offset_x: int
+    offset_y: int
+    w: int
+    h: int
+    source: str  # "action" | "focus" | "composite_fallback"
+
+
+_COMPOSITE_FALLBACK = MonitorTarget(
+    idx=-1,
+    offset_x=0,
+    offset_y=0,
+    w=0,
+    h=0,
+    source="composite_fallback",
+)
+
+
+def _find_monitor(geometry: List[Dict[str, Any]], idx: int) -> Optional[Dict[str, Any]]:
+    """Retourne le monitor d'index donné, ou None si absent."""
+    for m in geometry:
+        if m.get("idx") == idx:
+            return m
+    return None
+
+
+def _to_target(monitor: Dict[str, Any], source: str) -> MonitorTarget:
+    return MonitorTarget(
+        idx=int(monitor["idx"]),
+        offset_x=int(monitor.get("x", 0)),
+        offset_y=int(monitor.get("y", 0)),
+        w=int(monitor.get("w", 0)),
+        h=int(monitor.get("h", 0)),
+        source=source,
+    )
+
+
+def resolve_target_monitor(
+    action: Dict[str, Any],
+    session_state: Dict[str, Any],
+) -> MonitorTarget:
+    """Résout l'écran cible d'une action de replay.
+
+    Args:
+        action: Dict de l'action (peut contenir `monitor_index`).
+        session_state: État de la session (doit contenir `monitors_geometry`
+            et `last_focused_monitor`).
+
+    Returns:
+        MonitorTarget avec l'offset à appliquer aux coordonnées de grounding.
+    """
+    geometry: List[Dict[str, Any]] = session_state.get("monitors_geometry") or []
+
+    # 1. Cible explicite via action
+    explicit_idx = action.get("monitor_index")
+    if explicit_idx is not None and geometry:
+        m = _find_monitor(geometry, int(explicit_idx))
+        if m is not None:
+            return _to_target(m, source="action")
+        # Index invalide → on tombe sur le fallback focus
+        logger.warning(
+            "[BUS] lea:monitor_invalid_index requested=%d available_idx=%s",
+            int(explicit_idx), [g.get("idx") for g in geometry],
+        )
+
+    # 2. Fallback focus actif
+    focused_idx = session_state.get("last_focused_monitor")
+    if focused_idx is not None and geometry:
+        m = _find_monitor(geometry, int(focused_idx))
+        if m is not None:
+            return _to_target(m, source="focus")
+        logger.warning(
+            "[BUS] lea:monitor_unavailable focused_idx=%d available_idx=%s",
+            int(focused_idx), [g.get("idx") for g in geometry],
+        )
+
+    # 3. Fallback composite (backward compat — comportement actuel mss.monitors[0])
+    return _COMPOSITE_FALLBACK

agent_v0/server_v1/replay_engine.py

@@ -32,8 +32,16 @@ _ALLOWED_ACTION_TYPES = {
     "click", "type", "key_combo", "scroll", "wait",
     "file_open", "file_save", "file_close", "file_new", "file_dialog",
     "double_click", "right_click", "drag",
-    "verify_screen",  # Replay hybride : vérification visuelle entre groupes
+    "verify_screen",     # Replay hybride : vérification visuelle entre groupes
+    "pause_for_human",   # Pause supervisée explicite (interceptée par /replay/next)
+    "extract_text",      # OCR serveur sur dernier heartbeat → variable workflow
+    "t2a_decision",      # Analyse LLM facturation T2A → variable workflow
 }
+
+# Types d'actions exécutées CÔTÉ SERVEUR (jamais transmises à l'Agent V1).
+# Le pipeline /replay/next les traite en boucle interne et passe à l'action
+# suivante jusqu'à trouver une action visuelle (à transmettre au client).
+_SERVER_SIDE_ACTION_TYPES = {"extract_text", "t2a_decision"}
 _MAX_ACTION_TEXT_LENGTH = 10000
 _MAX_KEYS_PER_COMBO = 10
 # Touches autorisées dans les key_combo (modificateurs + touches spéciales + caractères simples)
@@ -852,6 +860,30 @@ def _edge_to_normalized_actions(edge, params: Dict[str, Any]) -> List[Dict[str,
             keys = [action_params["key"]]
         normalized["keys"] = keys
 
+    elif action_type == "pause_for_human":
+        normalized["type"] = "pause_for_human"
+        normalized["parameters"] = {
+            "message": action_params.get("message", "Validation requise"),
+        }
+        return [normalized]  # pas de target/coords pour cette action logique
+
+    elif action_type == "extract_text":
+        normalized["type"] = "extract_text"
+        normalized["parameters"] = {
+            "output_var": action_params.get("output_var", "extracted_text"),
+            "paragraph": bool(action_params.get("paragraph", True)),
+        }
+        return [normalized]
+
+    elif action_type == "t2a_decision":
+        normalized["type"] = "t2a_decision"
+        normalized["parameters"] = {
+            "input_template": action_params.get("input_template", ""),
+            "output_var": action_params.get("output_var", "t2a_result"),
+            "model": action_params.get("model"),
+        }
+        return [normalized]
+
     else:
         logger.warning(f"Type d'action inconnu : {action_type}")
         return []
@@ -886,6 +918,143 @@ def _substitute_variables(text: str, params: Dict[str, Any], defaults: Dict[str,
     return re.sub(r'\$\{(\w+)\}', replacer, text)
 
 
+# Regex pour le templating runtime : {{var}} ou {{var.champ}} ou {{var.champ.sous}}
+_RUNTIME_VAR_PATTERN = re.compile(r'\{\{\s*(\w+)(?:\.([\w.]+))?\s*\}\}')
+
+
+def _resolve_runtime_vars_in_str(text: str, variables: Dict[str, Any]) -> str:
+    """Remplace {{var}} et {{var.field}} par leur valeur depuis le dict variables.
+
+    Variables/champs absents : laissés tels quels (ne casse pas le pipeline).
+    Pour les valeurs non-str (dict, list), str() est appelé.
+    """
+    def replacer(match):
+        var_name = match.group(1)
+        path = match.group(2)
+        if var_name not in variables:
+            return match.group(0)
+        value = variables[var_name]
+        if path:
+            for field in path.split('.'):
+                if isinstance(value, dict) and field in value:
+                    value = value[field]
+                else:
+                    return match.group(0)
+        return str(value)
+
+    return _RUNTIME_VAR_PATTERN.sub(replacer, text)
+
+
+def _resolve_runtime_vars(value: Any, variables: Dict[str, Any]) -> Any:
+    """Résout récursivement les {{var}} et {{var.field}} dans une valeur.
+
+    Supporte str, dict, list. Les autres types sont retournés tels quels.
+    Si variables est vide ou None, value est retournée inchangée.
+    """
+    if not variables:
+        return value
+    if isinstance(value, str):
+        return _resolve_runtime_vars_in_str(value, variables)
+    if isinstance(value, dict):
+        return {k: _resolve_runtime_vars(v, variables) for k, v in value.items()}
+    if isinstance(value, list):
+        return [_resolve_runtime_vars(item, variables) for item in value]
+    return value
+
+
+# =========================================================================
+# Handlers pour les actions exécutées côté serveur (extract_text, t2a_decision)
+# =========================================================================
+
+def _handle_extract_text_action(
+    action: Dict[str, Any],
+    replay_state: Dict[str, Any],
+    session_id: str,
+    last_heartbeat: Dict[str, Dict[str, Any]],
+) -> bool:
+    """Traite une action extract_text côté serveur. Stocke le texte OCRisé dans
+    replay_state["variables"][output_var]. Retourne True si succès.
+
+    Robuste aux échecs : si pas de heartbeat ou OCR raté, stocke "" et retourne
+    False (le pipeline continue, pas de blocage).
+    """
+    params = action.get("parameters") or {}
+    output_var = (params.get("output_var") or "extracted_text").strip()
+    paragraph = bool(params.get("paragraph", True))
+
+    heartbeat = last_heartbeat.get(session_id) or {}
+    path = heartbeat.get("path")
+    text = ""
+
+    if path:
+        try:
+            from core.llm import extract_text_from_image
+            text = extract_text_from_image(path, paragraph=paragraph)
+        except Exception as e:
+            logger.warning("extract_text OCR échoué (%s) — variable '%s' = ''", e, output_var)
+    else:
+        logger.warning(
+            "extract_text : pas de heartbeat pour session %s — variable '%s' = ''",
+            session_id, output_var,
+        )
+
+    replay_state.setdefault("variables", {})[output_var] = text
+    logger.info(
+        "extract_text → variable '%s' (%d chars) replay %s",
+        output_var, len(text), replay_state.get("replay_id", "?"),
+    )
+    return bool(text)
+
+
+def _handle_t2a_decision_action(
+    action: Dict[str, Any],
+    replay_state: Dict[str, Any],
+) -> bool:
+    """Traite une action t2a_decision côté serveur. Stocke le résultat JSON
+    dans replay_state["variables"][output_var]. Retourne True si succès.
+
+    Le DPI à analyser vient de action.parameters.input_template (déjà résolu
+    par _resolve_runtime_vars donc les {{var}} sont remplis).
+    """
+    params = action.get("parameters") or {}
+    output_var = (params.get("output_var") or "t2a_result").strip()
+    dpi_text = (params.get("input_template") or params.get("dpi") or "").strip()
+    model = params.get("model") or None  # None → DEFAULT_MODEL
+
+    if not dpi_text:
+        logger.warning(
+            "t2a_decision : input vide — variable '%s' = {decision: 'INDETERMINE'}", output_var,
+        )
+        replay_state.setdefault("variables", {})[output_var] = {
+            "decision": "INDETERMINE",
+            "justification": "DPI vide ou non extrait",
+            "confiance": "faible",
+            "_error": "empty_input",
+        }
+        return False
+
+    try:
+        from core.llm import analyze_dpi, DEFAULT_MODEL
+        result = analyze_dpi(dpi_text, model=model or DEFAULT_MODEL)
+    except Exception as e:
+        logger.warning("t2a_decision : analyze_dpi exception %s", e)
+        result = {
+            "decision": "INDETERMINE",
+            "justification": f"Erreur analyse : {e}",
+            "confiance": "faible",
+            "_error": str(e),
+        }
+
+    replay_state.setdefault("variables", {})[output_var] = result
+    decision = result.get("decision", "?")
+    elapsed = result.get("_elapsed_s", "?")
+    logger.info(
+        "t2a_decision → variable '%s' decision=%s (%ss) replay %s",
+        output_var, decision, elapsed, replay_state.get("replay_id", "?"),
+    )
+    return "_error" not in result
+
+
 def _expand_compound_steps(
     steps: List[Dict[str, Any]], base: Dict[str, Any], params: Dict[str, Any]
 ) -> List[Dict[str, Any]]:
@@ -1208,6 +1377,18 @@ def _create_replay_state(
         # Champs pour pause supervisée (target_not_found)
         "failed_action": None,   # Contexte de l'action en echec (quand paused_need_help)
         "pause_message": None,   # Message a afficher a l'utilisateur
+        # Variables d'exécution produites en cours de workflow (extract_text,
+        # t2a_decision, etc.). Résolues via templating {{var}} ou {{var.field}}
+        # dans les paramètres des actions suivantes.
+        "variables": {},
+        # QW2 — Anneaux d'historique pour LoopDetector (5 derniers max)
+        "_screenshot_history": [],   # images PIL des N derniers heartbeats (LoopDetector embed à chaque tick)
+        "_action_history": [],       # N dernières actions exécutées (signature)
+        # QW4 — Safety checks (hybride déclaratif + LLM contextuel) et audit acquittements
+        "safety_checks": [],            # liste produite par SafetyChecksProvider
+        "checks_acknowledged": [],      # ids acquittés via /replay/resume (audit trail)
+        "pause_reason": "",             # "loop_detected" | "" pour V1
+        "pause_payload": None,          # payload complet pour debug/audit
     }
 
 

agent_v0/server_v1/resolve_engine.py

@@ -2193,22 +2193,33 @@ def _validate_resolution_quality(
         dx = abs(resolved_x - fallback_x_pct)
         dy = abs(resolved_y - fallback_y_pct)
         if dx > _RESOLUTION_MAX_DRIFT or dy > _RESOLUTION_MAX_DRIFT:
+            # Exception : si le template matching trouve l'image avec une
+            # similarité quasi parfaite, on fait confiance à la position
+            # visuelle peu importe le drift. Une image retrouvée à >= 0.95
+            # de score est SUR l'écran à l'endroit indiqué — le drift par
+            # rapport à l'enregistrement ne reflète qu'un changement de
+            # layout (scroll, redimensionnement, F11, devtools), pas une
+            # erreur de résolution.
+            _HIGH_CONFIDENCE = 0.95
+            if score >= _HIGH_CONFIDENCE and method.startswith("template_matching"):
+                logger.info(
+                    "[REPLAY] Drift (%.3f, %.3f) > %.2f IGNORÉ : score=%.3f >= %.2f "
+                    "sur %s — résultat visuel fiable, on l'utilise",
+                    dx, dy, _RESOLUTION_MAX_DRIFT, score, _HIGH_CONFIDENCE, method,
+                )
+                return result
+
             logger.warning(
-                "[REPLAY] Resolution REJETÉE (drift trop grand) : "
+                "[REPLAY] Drift trop grand (%.3f, %.3f) > %.2f — fallback coords enregistrées (%.3f, %.3f)",
-                "method=%s resolved=(%.3f, %.3f) expected=(%.3f, %.3f) "
+                dx, dy, _RESOLUTION_MAX_DRIFT, fallback_x_pct, fallback_y_pct,
-                "drift=(%.3f, %.3f) max=%.2f",
-                method, resolved_x, resolved_y,
-                fallback_x_pct, fallback_y_pct,
-                dx, dy, _RESOLUTION_MAX_DRIFT,
             )
+            # Fallback : coordonnées enregistrées lors de la capture (écran identique = safe)
             return {
-                "resolved": False,
+                "resolved": True,
-                "method": f"rejected_drift_{method}",
+                "method": "fallback_recorded_coords",
-                "reason": f"drift_dx{dx:.3f}_dy{dy:.3f}_max{_RESOLUTION_MAX_DRIFT:.2f}",
+                "reason": f"drift_dx{dx:.3f}_dy{dy:.3f}_using_recorded",
                 "original_method": method,
                 "original_score": score,
-                "drift_dx": round(dx, 3),
-                "drift_dy": round(dy, 3),
                 "x_pct": fallback_x_pct,
                 "y_pct": fallback_y_pct,
             }

agent_v0/server_v1/safety_checks_provider.py

@@ -0,0 +1,190 @@
+# agent_v0/server_v1/safety_checks_provider.py
+"""SafetyChecksProvider — checks hybrides déclaratifs + LLM contextuels (QW4).
+
+Pour une action pause_for_human :
+- les checks déclaratifs (workflow) sont toujours inclus
+- si safety_level == "medical_critical" et RPA_SAFETY_CHECKS_LLM_ENABLED=1,
+  un appel LLM (medgemma:4b par défaut) ajoute jusqu'à N checks contextuels
+
+Tout échec côté LLM (timeout, exception, parse) → additional_checks=[] :
+le replay continue avec uniquement les déclaratifs (fallback safe).
+"""
+
+import base64
+import json
+import logging
+import os
+import uuid
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PausePayload:
+    checks: List[Dict[str, Any]] = field(default_factory=list)
+    pause_reason: str = ""
+    message: str = ""
+
+
+def _env(name: str, default: str) -> str:
+    return os.environ.get(name, default).strip()
+
+
+def _env_int(name: str, default: int) -> int:
+    try:
+        return int(os.environ.get(name, default))
+    except (TypeError, ValueError):
+        return default
+
+
+def _env_bool_enabled(name: str) -> bool:
+    val = os.environ.get(name, "1").strip().lower()
+    return val not in ("0", "false", "no", "off", "")
+
+
+def build_pause_payload(
+    action: Dict[str, Any],
+    replay_state: Dict[str, Any],
+    last_screenshot: Optional[str],
+) -> PausePayload:
+    """Construit le payload de pause enrichi pour une action pause_for_human."""
+    params = action.get("parameters") or {}
+    message = params.get("message", "Validation requise")
+    safety_level = params.get("safety_level")
+    declarative = params.get("safety_checks") or []
+
+    # Normalisation des checks déclaratifs
+    checks: List[Dict[str, Any]] = []
+    for d in declarative:
+        checks.append({
+            "id": d.get("id") or f"decl_{uuid.uuid4().hex[:6]}",
+            "label": d.get("label", "Validation"),
+            "required": bool(d.get("required", True)),
+            "source": "declarative",
+            "evidence": None,
+        })
+
+    # Ajout LLM contextual si applicable
+    if safety_level == "medical_critical" and _env_bool_enabled("RPA_SAFETY_CHECKS_LLM_ENABLED"):
+        try:
+            additional = _call_llm_for_contextual_checks(
+                action=action,
+                replay_state=replay_state,
+                last_screenshot=last_screenshot,
+                existing_labels=[c["label"] for c in checks],
+            )
+        except Exception as e:
+            logger.warning("[BUS] lea:safety_checks_llm_failed reason=exception detail=%s", e)
+            additional = []
+
+        for a in additional:
+            checks.append({
+                "id": f"llm_{uuid.uuid4().hex[:6]}",
+                "label": a.get("label", ""),
+                "required": False,  # checks LLM = informationnels, pas obligatoires V1
+                "source": "llm_contextual",
+                "evidence": a.get("evidence", ""),
+            })
+
+    return PausePayload(
+        checks=checks,
+        pause_reason="",
+        message=message,
+    )
+
+
+def _call_llm_for_contextual_checks(
+    action: Dict[str, Any],
+    replay_state: Dict[str, Any],
+    last_screenshot: Optional[str],
+    existing_labels: List[str],
+) -> List[Dict[str, str]]:
+    """Appelle Ollama en mode JSON strict pour générer 0-N checks contextuels.
+
+    Returns:
+        List[{label, evidence}] (max RPA_SAFETY_CHECKS_LLM_MAX_CHECKS).
+        [] sur tout échec (timeout, JSON invalide, exception).
+    """
+    import requests
+
+    model = _env("RPA_SAFETY_CHECKS_LLM_MODEL", "medgemma:4b")
+    timeout_s = _env_int("RPA_SAFETY_CHECKS_LLM_TIMEOUT_S", 5)
+    max_checks = _env_int("RPA_SAFETY_CHECKS_LLM_MAX_CHECKS", 3)
+    ollama_url = _env("OLLAMA_URL", "http://localhost:11434")
+
+    params = action.get("parameters") or {}
+    workflow_message = params.get("message", "")
+    existing = ", ".join(existing_labels) if existing_labels else "aucun"
+
+    prompt = f"""Tu es Léa, assistante médicale supervisée.
+Avant de continuer le workflow, tu dois lister 0 à {max_checks} vérifications supplémentaires
+que l'humain doit acquitter, en regardant l'écran actuel.
+
+Contexte workflow : {workflow_message}
+Checks déjà demandés : {existing}
+
+NE répète PAS un check déjà demandé.
+Si rien d'inhabituel à signaler, retourne {{"additional_checks": []}}.
+
+Réponds UNIQUEMENT en JSON :
+{{
+  "additional_checks": [
+    {{"label": "string court", "evidence": "ce que tu as vu d'inhabituel"}}
+  ]
+}}
+"""
+
+    payload = {
+        "model": model,
+        "prompt": prompt,
+        "stream": False,
+        "format": "json",
+        "options": {"temperature": 0.1, "num_predict": 200},
+    }
+
+    if last_screenshot and os.path.isfile(last_screenshot):
+        try:
+            with open(last_screenshot, "rb") as f:
+                payload["images"] = [base64.b64encode(f.read()).decode("ascii")]
+        except Exception as e:
+            logger.debug("safety_checks: lecture screenshot échouée (%s) — appel sans image", e)
+
+    try:
+        response = requests.post(
+            f"{ollama_url}/api/generate",
+            json=payload,
+            timeout=timeout_s,
+        )
+        if response.status_code != 200:
+            logger.warning("[BUS] lea:safety_checks_llm_failed reason=http_status detail=%s", response.status_code)
+            return []
+        text = response.json().get("response", "").strip()
+    except requests.Timeout:
+        logger.warning("[BUS] lea:safety_checks_llm_failed reason=timeout detail=%ss", timeout_s)
+        return []
+    except Exception as e:
+        logger.warning("[BUS] lea:safety_checks_llm_failed reason=network detail=%s", e)
+        return []
+
+    # format=json garantit normalement du JSON valide
+    try:
+        parsed = json.loads(text)
+    except json.JSONDecodeError as e:
+        logger.warning("[BUS] lea:safety_checks_llm_failed reason=json_decode detail=%s", e)
+        return []
+
+    additional = parsed.get("additional_checks") or []
+    if not isinstance(additional, list):
+        return []
+
+    # Filtre + tronc
+    valid = []
+    for item in additional[:max_checks]:
+        if isinstance(item, dict) and item.get("label"):
+            valid.append({
+                "label": str(item["label"])[:200],
+                "evidence": str(item.get("evidence", ""))[:300],
+            })
+    return valid

core/execution/input_handler.py

@@ -22,6 +22,18 @@ try:
 except ImportError:
     PYAUTOGUI_AVAILABLE = False
 
+try:
+    import mss
+    MSS_AVAILABLE = True
+except ImportError:
+    MSS_AVAILABLE = False
+
+try:
+    from PIL import Image as PILImage
+    PIL_AVAILABLE = True
+except ImportError:
+    PIL_AVAILABLE = False
+
 
 def safe_type_text(text: str):
     """Saisie de texte compatible VM/Citrix et claviers AZERTY/QWERTY.
@@ -116,13 +128,13 @@ def check_screen_for_patterns() -> Optional[Dict[str, Any]]:
 
         pattern = lib.find_pattern(ocr_text)
         if pattern and pattern['category'] in ('dialog', 'popup'):
-            logger.info(f"Pattern UI détecté: {pattern['pattern']} → {pattern['action']} '{pattern['target']}'")
+            print(f"🧠 [PatternCheck] Détecté: '{pattern['pattern']}' → {pattern['action']} '{pattern['target']}'")
             return pattern
 
         return None
 
     except Exception as e:
-        logger.debug(f"Pattern check échoué: {e}")
+        print(f"⚠️ [PatternCheck] Erreur: {e}")
         return None
 
 
@@ -145,26 +157,42 @@ def handle_detected_pattern(pattern: Dict[str, Any]) -> bool:
 
     if action == 'click':
         candidates_labels = [target] + alternatives
+        print(f"🔧 [Réflexe/handle] Recherche bouton parmi: {candidates_labels}")
 
         try:
             import mss
+            import numpy as np
             from PIL import Image
 
-            # Importer OCR (essayer les deux chemins)
-            try:
-                from services.ocr_service import ocr_extract_words
-            except ImportError:
-                from core.extraction.field_extractor import FieldExtractor
-                extractor = FieldExtractor()
-                def ocr_extract_words(img):
-                    return extractor.extract_words_from_image(img)
-
             with mss.mss() as sct:
                 monitor = sct.monitors[0]
                 screenshot = sct.grab(monitor)
                 screen = Image.frombytes('RGB', screenshot.size, screenshot.bgra, 'raw', 'BGRX')
 
-            words = ocr_extract_words(screen)
+            # EasyOCR (rapide, bonne qualité GUI) avec fallback docTR.
+            # gpu=True : harmonisé avec dialog_handler.py et title_verifier.py.
+            # Coût VRAM ~0.5 GB, sous le budget RTX 5070 (cf. deploy/VRAM_BUDGET.md).
+            words = []
+            try:
+                import easyocr
+                _reader = easyocr.Reader(['fr', 'en'], gpu=True, verbose=False)
+                results = _reader.readtext(np.array(screen))
+                for (bbox_pts, text, conf) in results:
+                    if not text or len(text.strip()) < 1:
+                        continue
+                    x1 = int(min(p[0] for p in bbox_pts))
+                    y1 = int(min(p[1] for p in bbox_pts))
+                    x2 = int(max(p[0] for p in bbox_pts))
+                    y2 = int(max(p[1] for p in bbox_pts))
+                    words.append({'text': text.strip(), 'bbox': [x1, y1, x2, y2]})
+            except ImportError:
+                try:
+                    from services.ocr_service import ocr_extract_words
+                    words = ocr_extract_words(screen) or []
+                except ImportError:
+                    pass
+
+            print(f"🔧 [Réflexe/handle] {len(words)} mots OCR détectés")
 
             # Collecter tous les matchs, prendre le plus bas (bouton = bas du dialogue)
             all_matches = []
@@ -175,58 +203,28 @@ def handle_detected_pattern(pattern: Dict[str, Any]) -> bool:
                     word_text = word['text'].lower()
                     if len(word_text) < 2 or len(candidate_lower) < 2:
                         continue
-                    if word_text == candidate_lower:
+                    # Match exact ou inclusion
+                    if word_text == candidate_lower or candidate_lower in word_text or word_text in candidate_lower:
                         x1, y1, x2, y2 = word['bbox']
                         all_matches.append({
                             'text': word['text'],
                             'x': int((x1 + x2) / 2),
                             'y': int((y1 + y2) / 2),
-                            'match_type': 'exact',
+                            'candidate': candidate,
                         })
 
-            # Recherche partielle (lettre soulignée manquante)
-            if not all_matches:
-                for candidate in candidates_labels:
-                    if len(candidate) > 3:
-                        partial = candidate[1:].lower()
-                        for word in words:
-                            if partial in word['text'].lower():
-                                x1, y1, x2, y2 = word['bbox']
-                                all_matches.append({
-                                    'text': word['text'],
-                                    'x': int((x1 + x2) / 2),
-                                    'y': int((y1 + y2) / 2),
-                                    'match_type': 'partial',
-                                })
-
             if all_matches:
                 best = max(all_matches, key=lambda m: m['y'])
-                logger.info(f"Clic sur '{best['text']}' à ({best['x']}, {best['y']})")
+                print(f"✅ [Réflexe/handle] Clic sur '{best['text']}' à ({best['x']}, {best['y']})")
                 pyautogui.click(best['x'], best['y'])
                 time.sleep(1.0)
                 return True
 
-            logger.info(f"Bouton '{target}' introuvable par OCR — appel VLM...")
+            print(f"⚠️ [Réflexe/handle] Bouton '{target}' introuvable parmi {[w['text'] for w in words[:15]]}")
-            vlm_result = vlm_reason_about_screen(
-                objective=f"Cliquer sur le bouton '{target}'",
-                context=f"Un dialogue '{pattern.get('pattern')}' est détecté"
-            )
-            if vlm_result and vlm_result.get('action') == 'click' and vlm_result.get('target'):
-                vlm_target = vlm_result['target']
-                for word in words:
-                    if vlm_target.lower() in word['text'].lower():
-                        x1, y1, x2, y2 = word['bbox']
-                        x = int((x1 + x2) / 2)
-                        y = int((y1 + y2) / 2)
-                        logger.info(f"VLM → clic sur '{word['text']}' à ({x}, {y})")
-                        pyautogui.click(x, y)
-                        time.sleep(1.0)
-                        return True
-
             return False
 
         except Exception as e:
-            logger.warning(f"OCR bouton échoué: {e}")
+            print(f"⚠️ [Réflexe/handle] Erreur: {e}")
             return False
 
     elif action == 'hotkey':
@@ -328,6 +326,7 @@ def find_element_on_screen(
     target_description: str = "",
     anchor_image_base64: Optional[str] = None,
     anchor_bbox: Optional[Dict] = None,
+    monitor_idx: Optional[int] = None,
 ) -> Optional[Dict[str, Any]]:
     """
     Cherche un élément sur l'écran en utilisant 3 méthodes en cascade.
@@ -341,6 +340,7 @@ def find_element_on_screen(
         target_description: Description plus longue (ex: "le dossier Demo sur le bureau")
         anchor_image_base64: Image de référence de l'ancre (pour CLIP matching, réservé futur)
         anchor_bbox: Position originale de l'ancre (pour désambiguïser les matchs multiples)
+        monitor_idx: Index logique 0..N-1 du monitor à scruter. None = composite legacy.
 
     Returns:
         {'x': int, 'y': int, 'method': str, 'confidence': float} ou None
@@ -363,6 +363,13 @@ def find_element_on_screen(
         logger.debug("find_element_on_screen: ni target_text ni target_description fournis")
         return None
 
+    # Propager monitor_idx au niveau OCR via anchor_bbox (sans muter l'argument original)
+    if monitor_idx is not None and anchor_bbox is not None:
+        anchor_bbox = dict(anchor_bbox)  # copie pour ne pas muter l'argument
+        anchor_bbox["monitor_idx"] = monitor_idx
+    elif monitor_idx is not None:
+        anchor_bbox = {"monitor_idx": monitor_idx}
+
     search_label = target_description or target_text
     logger.info(f"[Grounding] Recherche élément: '{search_label}' (cascade 3 niveaux)")
 
@@ -372,12 +379,12 @@ def find_element_on_screen(
         return result
 
     # ─── Niveau 2 — UI-TARS grounding (~3s) ───
-    result = _grounding_ui_tars(target_text, target_description)
+    result = _grounding_ui_tars(target_text, target_description, monitor_idx=monitor_idx)
     if result:
         return result
 
     # ─── Niveau 3 — VLM reasoning (~10s) ───
-    result = _grounding_vlm(target_text, target_description)
+    result = _grounding_vlm(target_text, target_description, monitor_idx=monitor_idx)
     if result:
         return result
 
@@ -427,20 +434,43 @@ def _describe_anchor_image(anchor_image_base64: str) -> Optional[str]:
         return None
 
 
-def _capture_screen():
+def _capture_screen(monitor_idx=None):
-    """Capture l'écran principal et retourne (PIL.Image, width, height)."""
+    """Capture l'écran et retourne (PIL.Image, width, height, offset_x, offset_y).
-    try:
-        import mss
-        from PIL import Image as PILImage
 
+    Args:
+        monitor_idx: Index logique 0..N-1 du monitor à capturer (cf. screeninfo).
+            Si None : capture composite (mss.monitors[0]) — comportement legacy.
+
+    Returns:
+        (image, w, h, offset_x, offset_y). offset = (0,0) en mode composite.
+    """
+    try:
         with mss.mss() as sct:
-            monitor = sct.monitors[0]
+            if monitor_idx is None:
+                # Comportement actuel : composite tous écrans
+                monitor = sct.monitors[0]
+                offset_x, offset_y = 0, 0
+            else:
+                # mss skip monitors[0] (composite). Index logique 0 → mss.monitors[1].
+                mss_idx = int(monitor_idx) + 1
+                if mss_idx >= len(sct.monitors):
+                    logger.warning(
+                        "mss.monitors[%d] hors limites (n=%d) — fallback composite",
+                        mss_idx, len(sct.monitors),
+                    )
+                    monitor = sct.monitors[0]
+                    offset_x, offset_y = 0, 0
+                else:
+                    monitor = sct.monitors[mss_idx]
+                    offset_x = int(monitor.get("left", 0))
+                    offset_y = int(monitor.get("top", 0))
+
             screenshot = sct.grab(monitor)
             screen = PILImage.frombytes('RGB', screenshot.size, screenshot.bgra, 'raw', 'BGRX')
-            return screen, monitor['width'], monitor['height']
+            return screen, monitor['width'], monitor['height'], offset_x, offset_y
     except Exception as e:
         logger.debug(f"Capture écran échouée: {e}")
-        return None, 0, 0
+        return None, 0, 0, 0, 0
 
 
 def _grounding_ocr(target_text: str, anchor_bbox: Optional[Dict] = None) -> Optional[Dict[str, Any]]:
@@ -455,7 +485,8 @@ def _grounding_ocr(target_text: str, anchor_bbox: Optional[Dict] = None) -> Opti
         return None
 
     try:
-        screen, screen_w, screen_h = _capture_screen()
+        monitor_idx_param = anchor_bbox.get("monitor_idx") if anchor_bbox else None
+        screen, screen_w, screen_h, ox, oy = _capture_screen(monitor_idx=monitor_idx_param)
         if screen is None:
             return None
 
@@ -519,14 +550,14 @@ def _grounding_ocr(target_text: str, anchor_bbox: Optional[Dict] = None) -> Opti
             sel = " ← CHOISI" if m is best else ""
             logger.info(f"  [OCR] Candidat: '{m['text']}' à ({m['x']}, {m['y']}) [{m['type']}]{sel}")
 
-        return {'x': best['x'], 'y': best['y'], 'method': 'ocr', 'confidence': best['conf']}
+        return {'x': best['x'] + ox, 'y': best['y'] + oy, 'method': 'ocr', 'confidence': best['conf']}
 
     except Exception as e:
         logger.debug(f"[Grounding/OCR] Erreur: {e}")
         return None
 
 
-def _grounding_ui_tars(target_text: str, target_description: str = "") -> Optional[Dict[str, Any]]:
+def _grounding_ui_tars(target_text: str, target_description: str = "", monitor_idx=None) -> Optional[Dict[str, Any]]:
     """Niveau 2 — UI-TARS grounding visuel (~3s)."""
     try:
         import requests
@@ -535,7 +566,7 @@ def _grounding_ui_tars(target_text: str, target_description: str = "") -> Option
         import re
         import os
 
-        screen, screen_w, screen_h = _capture_screen()
+        screen, screen_w, screen_h, ox, oy = _capture_screen(monitor_idx=monitor_idx)
         if screen is None:
             return None
 
@@ -580,7 +611,7 @@ def _grounding_ui_tars(target_text: str, target_description: str = "") -> Option
             # Valider que les coordonnées sont dans l'écran
             if 0 <= x <= screen_w and 0 <= y <= screen_h:
                 logger.info(f"[Grounding/UI-TARS] Grounding → ({x}, {y})")
-                return {'x': x, 'y': y, 'method': 'ui_tars', 'confidence': 0.85}
+                return {'x': x + ox, 'y': y + oy, 'method': 'ui_tars', 'confidence': 0.85}
             else:
                 logger.warning(f"[Grounding/UI-TARS] Coordonnées hors écran: ({x}, {y}) pour {screen_w}x{screen_h}")
                 return None
@@ -640,7 +671,7 @@ def _parse_ui_tars_coordinates(text: str, screen_w: int, screen_h: int) -> Optio
     return None
 
 
-def _grounding_vlm(target_text: str, target_description: str = "") -> Optional[Dict[str, Any]]:
+def _grounding_vlm(target_text: str, target_description: str = "", monitor_idx=None) -> Optional[Dict[str, Any]]:
     """Niveau 3 — VLM reasoning + confirmation OCR (~10s)."""
     try:
         search_label = target_description or target_text
@@ -662,7 +693,7 @@ def _grounding_vlm(target_text: str, target_description: str = "") -> Optional[D
         logger.info(f"[Grounding/VLM] VLM suggère de cliquer sur: '{vlm_target}'")
 
         # Confirmation par OCR : chercher le target VLM sur l'écran
-        screen, screen_w, screen_h = _capture_screen()
+        screen, screen_w, screen_h, ox, oy = _capture_screen(monitor_idx=monitor_idx)
         if screen is None:
             return None
 
@@ -684,7 +715,7 @@ def _grounding_vlm(target_text: str, target_description: str = "") -> Optional[D
                     x = int((x1 + x2) / 2)
                     y = int((y1 + y2) / 2)
                     logger.info(f"[Grounding/VLM] Confirmé par OCR: '{word['text']}' à ({x}, {y})")
-                    return {'x': x, 'y': y, 'method': 'vlm', 'confidence': 0.75}
+                    return {'x': x + ox, 'y': y + oy, 'method': 'vlm', 'confidence': 0.75}
 
             logger.debug(f"[Grounding/VLM] Target VLM '{vlm_target}' non trouvé par OCR")
             return None

core/execution/observe_reason_act.py

@@ -213,8 +213,40 @@ class ORALoop:
 
         # --- Mapper action_type vers action Decision ---
 
+        # Types d'action qui ne sont PAS des descriptions valides
+        _action_type_names = {'click_anchor', 'double_click_anchor', 'right_click_anchor',
+                              'hover_anchor', 'focus_anchor', 'scroll_to_anchor',
+                              'click', 'type_text', 'keyboard_shortcut', 'wait_for_anchor'}
+
         if action_type in ('click_anchor', 'click', 'double_click_anchor', 'right_click_anchor'):
-            target_text = anchor.get('target_text', '') or label
+            target_text = anchor.get('target_text', '') or anchor.get('description', '')
+
+            # Détecter les target_text absurdes : vide, nom d'action, ou bruit OCR
+            def _is_garbage(t):
+                if not t or t in _action_type_names:
+                    return True
+                # Bruit OCR : que des caractères spéciaux/chiffres/espaces
+                cleaned = t.replace('-', '').replace(' ', '').replace('.', '').replace('_', '')
+                if len(cleaned) < 3:
+                    return True
+                # Que des chiffres
+                if cleaned.isdigit():
+                    return True
+                return False
+
+            # Note: plus d'appel à _describe_anchor_image() (qwen2.5vl) ici.
+            # Le crop d'ancre (screenshot_b64) servira directement au template matching
+            # cv2 dans _act_click, puis fallback InfiGUI fusionné si nécessaire.
+            # Cela évite le conflit VRAM (qwen2.5vl 9.4GB + InfiGUI 2.4GB > 11.5GB GPU).
+
+            # Dernier fallback : label si pas un nom d'action
+            if _is_garbage(target_text):
+                target_text = label if label not in _action_type_names else ''
+                if target_text:
+                    print(f"🏷️ [ORA/reason] Label garbage, fallback texte: '{target_text}'")
+                else:
+                    print(f"🏷️ [ORA/reason] Pas de label texte — grounding via crop visuel uniquement")
+
             action = 'click'
             value = 'double' if action_type == 'double_click_anchor' else (
                 'right' if action_type == 'right_click_anchor' else 'left')
@@ -1222,6 +1254,7 @@ Règles:
             )
 
         print(f"🚀 [ORA] Démarrage workflow: {total} étapes, verify={self.verify_level}, retries={self.max_retries}")
+        print(f"🔧 [ORA] CODE VERSION: post-shortcut-dialog-handler ACTIF (26 avril 17h30)")
 
         for i, step in enumerate(steps):
             if not self._should_continue():
@@ -1234,6 +1267,28 @@ Règles:
             # --- 1. Observer l'état pré-action ---
             pre = self.observe()
 
+            # --- 1b. Réflexe : dialogue inattendu ? ---
+            # Déclenché si le pHash a changé de manière inattendue.
+            # Flux : titre fenêtre (50ms) → dialogue connu ? → InfiGUI clique (3s)
+            if i > 0 and hasattr(self, '_last_post_phash') and self._last_post_phash:
+                _phash_distance = self._phash_distance(pre.phash, self._last_post_phash)
+                if _phash_distance > 10:
+                    print(f"🧠 [ORA/réflexe] pHash changé (distance={_phash_distance}) → vérification dialogue")
+                    try:
+                        from core.grounding.dialog_handler import DialogHandler
+                        _dh = DialogHandler()
+                        _dh_result = _dh.handle_if_dialog(pre.screenshot)
+                        if _dh_result.get('handled'):
+                            print(f"✅ [ORA/réflexe] Dialogue '{_dh_result['title'][:30]}' géré → {_dh_result['action']}")
+                            time.sleep(0.5)
+                            pre = self.observe()
+                        elif _dh_result.get('dialog_type'):
+                            print(f"⚠️ [ORA/réflexe] Dialogue '{_dh_result.get('dialog_type')}' détecté mais non géré: {_dh_result.get('reason')}")
+                        else:
+                            print(f"🧠 [ORA/réflexe] Pas de dialogue détecté: {_dh_result.get('reason', '?')}")
+                    except Exception as _reflex_err:
+                        print(f"⚠️ [ORA/réflexe] Erreur: {_reflex_err}")
+
             # --- 2. Raisonner : construire la Decision ---
             decision = self.reason_workflow_step(step, pre)
 
@@ -1281,11 +1336,74 @@ Règles:
                         )
                     )
 
+            # --- 3b. Post-raccourci : attendre changement écran + gérer dialogue ---
+            # Après un keyboard_shortcut (pas scroll), on polle le pHash pour détecter
+            # si un dialogue est apparu (ex: "Enregistrer sous" après Ctrl+Shift+S).
+            # Si oui → InfiGUI localise et clique le bouton visuellement.
+            if act_success and decision.action == 'hotkey' and not decision.value.startswith('scroll_'):
+                print(f"🔍 [ORA/post-shortcut] ENTRÉ dans le bloc post-shortcut (action={decision.action}, value={decision.value})")
+                dialog_handled = self._handle_post_shortcut(pre)
+                if dialog_handled:
+                    time.sleep(0.5)
+                    post = self.observe()
+                    self._last_post_phash = post.phash
+                    if on_progress:
+                        on_progress(i + 1, total, VerificationResult(
+                            success=True, change_level='major',
+                            matches_expected=True,
+                            detail="Dialogue géré visuellement après raccourci"
+                        ))
+                    continue
+                else:
+                    # Invariant : aucune étape suivante ne doit s'exécuter tant que
+                    # la cascade déclenchée par le raccourci n'est pas pleinement résolue.
+                    # Cas typique : Ctrl+S → "Enregistrer sous" non géré → on ABORT plutôt
+                    # que de cliquer sur des coordonnées potentiellement obsolètes.
+                    msg = (
+                        f"Étape {i+1}: raccourci '{decision.value}' — cascade post-raccourci "
+                        f"non résolue (dialogue absent ou bloqué). Workflow stoppé pour éviter "
+                        f"un clic dans un contexte incohérent."
+                    )
+                    print(f"❌ [ORA/post-shortcut] {msg}")
+                    logger.warning(f"🆘 [ORA] {msg}")
+                    if on_progress:
+                        on_progress(i + 1, total, VerificationResult(
+                            success=False, change_level='none',
+                            matches_expected=False,
+                            detail="Cascade post-raccourci non résolue"
+                        ))
+                    return LoopResult(
+                        success=False, steps_completed=i, total_steps=total,
+                        reason=msg,
+                    )
+
             # Petit délai pour laisser l'écran se stabiliser
             time.sleep(0.3)
 
             # --- 4. Observer l'état post-action ---
             post = self.observe()
+            # Stocker le pHash post-action pour le réflexe check du step suivant
+            self._last_post_phash = post.phash
+
+            # --- 4b. Vérification titre OCR (non-bloquante, ~120ms) ---
+            _action_type = step.get('action_type', '')
+            if _action_type in ('double_click_anchor', 'click_anchor') and pre.screenshot and post.screenshot:
+                try:
+                    from core.grounding.title_verifier import TitleVerifier
+                    _tv = TitleVerifier()
+                    _tv_result = _tv.verify_action(pre.screenshot, post.screenshot, _action_type)
+                    if not _tv_result['success']:
+                        print(f"⚠️ [ORA/titre] {_tv_result['reason']} → retry")
+                        # Retry : recliquer
+                        time.sleep(0.5)
+                        self.act(decision, step)
+                        time.sleep(0.3)
+                        post = self.observe()
+                        self._last_post_phash = post.phash
+                    elif _tv_result['changed']:
+                        print(f"✅ [ORA/titre] '{_tv_result['title_after'][:40]}'")
+                except Exception as _tv_err:
+                    print(f"⚠️ [ORA/titre] Erreur: {_tv_err}")
 
             # --- 5. Vérifier ---
             verification = self.verify(pre, post, decision)
@@ -1345,10 +1463,112 @@ Règles:
     # Méthodes privées — actions
     # ═══════════════════════════════════════════════════════════
 
+    def _handle_post_shortcut(self, pre_obs: 'Observation') -> bool:
+        """Après un raccourci clavier, résoudre la cascade de dialogues réflexes.
+
+        Pilotage par DialogHandler (OCR direct), PAS par pHash. Raison :
+        un dialog modal qui s'ouvre dans une VM ne change quasiment pas le
+        pHash global de l'écran hôte (signature 8x8 sur 1920x1080 — un dialog
+        de 800x500 couvre ~3 pixels pHash, distance Hamming souvent < 3).
+        On poll donc directement DialogHandler.handle_if_dialog().
+
+        Returns:
+            True si au moins un dialog connu a été détecté + géré et qu'aucun
+            autre dialog n'apparaît dans la fenêtre de stabilité finale.
+            False si aucun dialog connu n'apparaît dans la fenêtre d'attente
+            initiale (le workflow doit ABORT — état incohérent).
+        """
+        from core.grounding.dialog_handler import DialogHandler
+
+        # Fenêtre d'attente du PREMIER dialog après le raccourci. Win11/QEMU :
+        # Ctrl+Shift+S → "Enregistrer sous" apparaît en <2s typiquement.
+        first_dialog_timeout = 8.0
+        # Budget total pour résoudre toute la cascade (InfiGUI ~15s/dialog).
+        total_timeout = 60.0
+        # Fenêtre de stabilité après le dernier dialog géré : si rien d'autre
+        # n'apparaît pendant cette durée, la cascade est considérée terminée.
+        # Doit couvrir l'apparition du popup modal suivant (post_click_wait + marge).
+        stable_window = 3.0
+        # Délai post-clic avant de tester le dialog suivant.
+        post_click_wait = 1.5
+        # Cadence de polling OCR (EasyOCR full-screen ~500ms/poll).
+        poll_interval = 0.5
+        # Garde-fou anti-boucle infinie.
+        max_dialog_iterations = 5
+
+        t_start = time.time()
+        dh = DialogHandler()
+        dialogs_handled = 0
+
+        def _elapsed() -> float:
+            return time.time() - t_start
+
+        def _poll_dialog(deadline: float) -> Optional[Dict[str, Any]]:
+            """Poll DialogHandler jusqu'à détection d'un dialog connu OU deadline.
+
+            Retourne le dict result si un dialog connu a été géré (cliqué),
+            None si la deadline est atteinte sans match. Si DialogHandler
+            détecte ET clique avec succès, le clic InfiGUI peut excéder la
+            deadline mais on retourne quand même le résultat (action déjà
+            engagée — on ne va pas l'annuler).
+            """
+            while time.time() < deadline:
+                obs = self.observe()
+                try:
+                    result = dh.handle_if_dialog(obs.screenshot)
+                except Exception as e:
+                    print(f"⚠️ [ORA/post-shortcut] Erreur dialog handler: {e}")
+                    return None
+                if result.get('handled'):
+                    return result
+                sleep_left = deadline - time.time()
+                if sleep_left > 0:
+                    time.sleep(min(poll_interval, sleep_left))
+            return None
+
+        # --- Étape 1 : attendre le PREMIER dialog ---
+        first_deadline = t_start + min(total_timeout, first_dialog_timeout)
+        result = _poll_dialog(first_deadline)
+        if result is None:
+            print(f"⏳ [ORA/post-shortcut] Aucun dialog connu détecté après "
+                  f"{_elapsed():.1f}s (fenêtre={first_dialog_timeout}s) — "
+                  f"raccourci sans effet attendu")
+            return False
+
+        dialogs_handled = 1
+        print(f"✅ [ORA/post-shortcut] Dialog #1 géré: {result.get('action')} "
+              f"({_elapsed():.1f}s)")
+        time.sleep(post_click_wait)
+
+        # --- Étape 2 : cascade — chaque dialog suivant doit apparaître dans stable_window ---
+        for iteration in range(1, max_dialog_iterations):
+            if _elapsed() >= total_timeout:
+                print(f"⏳ [ORA/post-shortcut] Timeout cascade ({total_timeout:.0f}s, "
+                      f"{dialogs_handled} dialog(s) géré(s))")
+                return True  # au moins un dialog traité → considéré OK
+
+            next_deadline = min(time.time() + stable_window, t_start + total_timeout)
+            result = _poll_dialog(next_deadline)
+            if result is None:
+                # Pas de nouveau dialog dans stable_window → cascade terminée
+                print(f"✅ [ORA/post-shortcut] Cascade résolue "
+                      f"({dialogs_handled} dialog(s), {_elapsed():.1f}s)")
+                return True
+
+            dialogs_handled += 1
+            print(f"✅ [ORA/post-shortcut] Dialog #{dialogs_handled} géré: "
+                  f"{result.get('action')} ({_elapsed():.1f}s)")
+            time.sleep(post_click_wait)
+
+        print(f"⚠️ [ORA/post-shortcut] Trop d'itérations cascade "
+              f"({max_dialog_iterations}) — cascade malformée, on s'arrête là")
+        return dialogs_handled > 0
+
     def _act_click(self, decision: Decision, step_params: dict) -> bool:
         """Exécute un clic (simple, double, droit, hover, focus).
 
-        Pipeline : template matching → find_element_on_screen (OCR → UI-TARS → VLM).
+        Pipeline FAST→SMART→THINK (si activé) ou ancien pipeline en fallback.
+        Activé par la variable d'environnement RPA_USE_FAST_PIPELINE=1.
         """
         if not PYAUTOGUI_AVAILABLE:
             logger.error("pyautogui non disponible")
@@ -1357,29 +1577,23 @@ Règles:
         anchor = step_params.get('visual_anchor', {})
         screenshot_b64 = anchor.get('screenshot')
         bbox = anchor.get('bounding_box', {})
-        target_text = anchor.get('target_text', '') or decision.target
+        # Utiliser le target nettoyé par reason_workflow_step (pas relire le garbage de l'ancre)
+        target_text = decision.target
         target_desc = anchor.get('description', '')
 
+        print(f"🎯 [ORA/_act_click] target='{target_text}', desc='{target_desc[:40]}', bbox={bbox.get('x','?')},{bbox.get('y','?')}")
+
         x, y = None, None
         method_used = ''
+        # Score et position du template-first (réutilisés en fallback intermédiaire)
+        template_score = 0.0
+        template_xy: Optional[tuple] = None
 
-        # --- Méthode 1 : UI-TARS grounding (~3s, 94% précision) ---
+        # --- AVANT-POSTE : template matching cv2 sur le crop d'ancre ---
-        # Le plus fiable : on dit "click on X" et UI-TARS trouve les coordonnées
+        # Si l'UI n'a pas changé (cas dominant en replay), un match pixel-perfect
-        if target_text or target_desc:
+        # nous donne le clic en ~50ms sans toucher au GPU. On ne déclenche le
-            try:
+        # pipeline VLM que si le score est insuffisant.
-                from core.execution.input_handler import _grounding_ui_tars
+        if screenshot_b64 and CV2_AVAILABLE and PIL_AVAILABLE and MSS_AVAILABLE:
-                click_label = target_desc or target_text
-                print(f"🎯 [ORA/UI-TARS] Recherche: '{click_label}'")
-                result = _grounding_ui_tars(target_text, target_desc)
-                if result:
-                    x, y = result['x'], result['y']
-                    method_used = 'ui_tars'
-                    print(f"✅ [ORA/UI-TARS] Trouvé à ({x}, {y})")
-            except Exception as e:
-                logger.debug(f"⚠️ [ORA/UI-TARS] Erreur: {e}")
-
-        # --- Méthode 2 : Template matching (~80ms) ---
-        if x is None and screenshot_b64 and CV2_AVAILABLE and PIL_AVAILABLE and MSS_AVAILABLE:
             try:
                 import io as _io
                 with mss_lib.mss() as sct:
@@ -1399,15 +1613,70 @@ Règles:
                     result_tm = cv2.matchTemplate(screen_cv, anchor_cv, cv2.TM_CCOEFF_NORMED)
                     _, max_val, _, max_loc = cv2.minMaxLoc(result_tm)
                     elapsed_ms = (time.time() - t0) * 1000
-                    print(f"⚡ [ORA/template] score={max_val:.3f} pos={max_loc} ({elapsed_ms:.0f}ms)")
+                    template_score = float(max_val)
-                    if max_val > 0.75:
+                    template_xy = (
-                        x = max_loc[0] + anchor_cv.shape[1] // 2
+                        max_loc[0] + anchor_cv.shape[1] // 2,
-                        y = max_loc[1] + anchor_cv.shape[0] // 2
+                        max_loc[1] + anchor_cv.shape[0] // 2,
-                        method_used = 'template'
+                    )
+                    print(f"⚡ [ORA/template-first] score={template_score:.3f} pos={max_loc} ({elapsed_ms:.0f}ms)")
+                    # Seuil élevé pour le mode "direct" : on veut être quasi-certain
+                    # que c'est le même élément, pixel-perfect, avant de zapper le VLM.
+                    if template_score >= 0.95:
+                        x, y = template_xy
+                        method_used = 'template_direct'
+                        print(f"✅ [ORA/template-first] Match direct → ({x}, {y}), skip pipeline")
             except Exception as e:
-                logger.debug(f"⚠️ [ORA/template] Erreur: {e}")
+                print(f"⚠️ [ORA/template-first] Erreur: {e}")
+
+        # --- Pipeline FAST→SMART→THINK (escalade si template-first n'a pas tranché) ---
+        _use_fast = os.environ.get('RPA_USE_FAST_PIPELINE', '1') == '1'
+
+        if x is None and _use_fast and (target_text or target_desc or screenshot_b64):
+            print(f"🎯 [ORA/_act_click] RPA_USE_FAST_PIPELINE={_use_fast}, has_target={bool(target_text or target_desc)}, template_score={template_score:.3f}")
+            try:
+                from core.grounding.fast_pipeline import FastSmartThinkPipeline
+                from core.grounding.target import GroundingTarget
+
+                _pipeline = FastSmartThinkPipeline.get_instance()
+
+                # Capture unique de l'écran
+                _screen_pil = None
+                if MSS_AVAILABLE and PIL_AVAILABLE:
+                    with mss_lib.mss() as _sct:
+                        _mon = _sct.monitors[0]
+                        _grab = _sct.grab(_mon)
+                        _screen_pil = Image.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+
+                _target = GroundingTarget(
+                    text=target_text,
+                    description=target_desc,
+                    template_b64=screenshot_b64 or "",
+                    original_bbox=bbox if bbox else None,
+                )
+
+                _result = _pipeline.locate(
+                    _target,
+                    screenshot_pil=_screen_pil,
+                    window_title=getattr(self, '_last_window_title', ''),
+                )
+
+                if _result:
+                    x, y = _result.x, _result.y
+                    method_used = _result.method
+                    print(f"🎯 [ORA/pipeline] ({x}, {y}) via {method_used} "
+                          f"conf={_result.confidence:.3f} ({_result.time_ms:.0f}ms)")
+
+            except Exception as e:
+                print(f"⚠️ [ORA/pipeline] Erreur: {e}")
+
+        # --- Fallback : on réutilise le score template-first si pertinent ---
+        # Si le pipeline VLM a échoué mais que le template-first avait un score
+        # intermédiaire (0.75-0.95), on accepte ce match comme secours.
+        if x is None and template_xy is not None and template_score >= 0.75:
+            x, y = template_xy
+            method_used = 'template_fallback'
+            print(f"⚡ [ORA/template-fallback] Réutilisation score={template_score:.3f} → ({x}, {y})")
 
-        # --- Méthode 3 : OCR texte (~1s) ---
         if x is None and target_text:
             try:
                 from core.execution.input_handler import _grounding_ocr
@@ -1417,22 +1686,21 @@ Règles:
                     method_used = 'ocr'
                     print(f"🔍 [ORA/OCR] Trouvé à ({x}, {y})")
             except Exception as e:
-                logger.debug(f"⚠️ [ORA/OCR] Erreur: {e}")
+                print(f"⚠️ [ORA/OCR] Erreur: {e}")
 
-        # --- Exécuter le clic ---
+        # --- Dernier recours : coordonnées statiques ---
         if x is None:
-            # Dernier recours : coordonnées statiques de l'ancre
             if bbox and bbox.get('width') and bbox.get('height'):
                 x = int(bbox.get('x', 0) + bbox.get('width', 0) / 2)
                 y = int(bbox.get('y', 0) + bbox.get('height', 0) / 2)
                 method_used = 'static_fallback'
-                logger.warning(f"⚠️ [ORA/click] Fallback coordonnées statiques: ({x}, {y})")
+                print(f"⚠️ [ORA/click] Fallback coordonnées statiques: ({x}, {y})")
             else:
-                logger.error(f"❌ [ORA/click] Impossible de localiser '{target_text}' — aucune méthode n'a fonctionné")
+                print(f"❌ [ORA/click] Impossible de localiser '{target_text}'")
                 return False
 
-        # --- Vérification pré-action : est-ce le bon élément ? ---
+        # --- Pas de pre-check VLM (le pipeline FAST→SMART→THINK a déjà validé) ---
-        if target_text and method_used not in ('template',) and MSS_AVAILABLE and PIL_AVAILABLE:
+        if False:
             try:
                 pre_check = self._verify_pre_click(x, y, target_text, target_desc)
                 if not pre_check:

core/grounding/__init__.py

@@ -0,0 +1,20 @@
+# core/grounding — Module de localisation d'éléments UI
+#
+# Centralise les méthodes de grounding visuel : template matching,
+# OCR, VLM, etc. Chaque méthode produit un GroundingResult uniforme.
+#
+# Le serveur de grounding (server.py) tourne dans un process séparé
+# sur le port 8200. Le client HTTP (UITarsGrounder) l'appelle via HTTP.
+# Le pipeline (GroundingPipeline) orchestre template → OCR → UI-TARS → static.
+
+from core.grounding.template_matcher import TemplateMatcher, MatchResult
+from core.grounding.target import GroundingTarget, GroundingResult
+from core.grounding.ui_tars_grounder import UITarsGrounder
+from core.grounding.pipeline import GroundingPipeline
+
+__all__ = [
+    'TemplateMatcher', 'MatchResult',
+    'GroundingTarget', 'GroundingResult',
+    'UITarsGrounder',
+    'GroundingPipeline',
+]

core/grounding/dialog_handler.py

@@ -0,0 +1,256 @@
+"""
+core/grounding/dialog_handler.py — Gestion intelligente des dialogues
+
+Quand un dialogue inattendu apparaît (pHash change après une action) :
+1. Lire le titre de la fenêtre (EasyOCR crop 45px, ~130ms)
+2. Si titre connu (Enregistrer sous, Confirmer, etc.) → action connue
+3. Demander à InfiGUI de cliquer sur le bon bouton (~3s)
+4. Vérifier que le dialogue a disparu (pHash)
+
+Pas de patterns prédéfinis pour les boutons. InfiGUI comprend
+visuellement le dialogue et clique au bon endroit.
+
+Utilisation :
+    from core.grounding.dialog_handler import DialogHandler
+
+    handler = DialogHandler()
+    result = handler.handle_if_dialog(screenshot_pil)
+    if result['handled']:
+        print(f"Dialogue '{result['title']}' géré → {result['action']}")
+"""
+
+from __future__ import annotations
+
+import time
+from typing import Any, Dict, Optional
+
+
+# Titres connus → quelle action demander à InfiGUI.
+#
+# IMPORTANT — ordre du dict = priorité de matching.
+# L'OCR est full-screen et capte souvent le texte du dialog parent ET du popup
+# modal qui apparaît par-dessus (ex: "Enregistrer sous" reste visible derrière
+# "Confirmer l'enregistrement"). Les popups modaux DOIVENT matcher avant les
+# fenêtres principales, sinon Léa clique sur le bouton du parent qui n'a pas
+# le focus.
+KNOWN_DIALOGS = {
+    # ── Popups modaux de confirmation (priorité HAUTE) ──────────────────
+    "voulez-vous le remplacer": {"target": "Oui", "description": "Clique sur Oui pour confirmer le remplacement du fichier"},
+    "do you want to replace": {"target": "Yes", "description": "Click Yes to confirm file replacement"},
+    "existe déjà": {"target": "Oui", "description": "Clique sur Oui, le fichier existe déjà et doit être remplacé"},
+    "already exists": {"target": "Yes", "description": "Click Yes, the file already exists"},
+    "remplacer": {"target": "Oui", "description": "Clique sur le bouton Oui pour confirmer le remplacement du fichier"},
+    "replace": {"target": "Yes", "description": "Click Yes to confirm file replacement"},
+    "écraser": {"target": "Oui", "description": "Clique sur Oui pour écraser le fichier"},
+    "overwrite": {"target": "Yes", "description": "Click Yes to overwrite"},
+    "confirmer l'enregistrement": {"target": "Oui", "description": "Clique sur Oui dans le popup de confirmation d'enregistrement"},
+    "confirmer": {"target": "Oui", "description": "Clique sur le bouton Oui dans le dialogue de confirmation"},
+    # ── Avertissements/erreurs (priorité haute, 1 seul bouton OK) ───────
+    "erreur": {"target": "OK", "description": "Clique sur OK pour fermer le message d'erreur"},
+    "error": {"target": "OK", "description": "Click OK to close the error message"},
+    "avertissement": {"target": "OK", "description": "Clique sur OK pour fermer l'avertissement"},
+    "warning": {"target": "OK", "description": "Click OK to close the warning"},
+    # ── Dialogs principaux de sauvegarde (priorité BASSE — fenêtres parents) ─
+    "voulez-vous enregistrer": {"target": "Enregistrer", "description": "Clique sur Enregistrer pour sauvegarder les modifications"},
+    "do you want to save": {"target": "Save", "description": "Click Save to save changes"},
+    "enregistrer sous": {"target": "Enregistrer", "description": "Clique sur le bouton Enregistrer dans le dialogue Enregistrer sous"},
+    "save as": {"target": "Save", "description": "Click the Save button in the Save As dialog"},
+}
+
+
+class DialogHandler:
+    """Gestion intelligente des dialogues via titre + InfiGUI."""
+
+    def __init__(self):
+        self._easyocr_reader = None
+
+    def handle_if_dialog(
+        self,
+        screenshot_pil,
+        previous_title: str = "",
+    ) -> Dict[str, Any]:
+        """Vérifie si l'écran montre un dialogue et le gère.
+
+        Args:
+            screenshot_pil: Screenshot PIL actuel.
+            previous_title: Titre de la fenêtre avant l'action (pour comparaison).
+
+        Returns:
+            Dict avec 'handled' (bool), 'title', 'action', 'position'.
+        """
+        t0 = time.time()
+
+        # 1. Lire le titre de la fenêtre
+        title = self._read_title(screenshot_pil)
+        if not title or len(title) < 3:
+            return {'handled': False, 'title': '', 'reason': 'Titre illisible'}
+
+        print(f"🔍 [Dialog] Titre lu: '{title}'")
+
+        # 2. Chercher si c'est un dialogue connu
+        matched_dialog = None
+        for key, action_info in KNOWN_DIALOGS.items():
+            if key in title.lower():
+                matched_dialog = (key, action_info)
+                break
+
+        if not matched_dialog:
+            # Pas un dialogue connu — le workflow continue normalement
+            return {'handled': False, 'title': title, 'reason': 'Pas un dialogue connu'}
+
+        dialog_key, action_info = matched_dialog
+        target = action_info['target']
+        description = action_info['description']
+
+        print(f"🧠 [Dialog] Dialogue détecté: '{dialog_key}' → clic '{target}'")
+
+        # 3. Demander à InfiGUI de cliquer sur le bouton
+        click_result = self._click_via_infigui(
+            target, description, screenshot_pil
+        )
+
+        dt = (time.time() - t0) * 1000
+
+        if click_result:
+            print(f"✅ [Dialog] Clic '{target}' à ({click_result['x']}, {click_result['y']}) ({dt:.0f}ms)")
+            return {
+                'handled': True,
+                'title': title,
+                'dialog_type': dialog_key,
+                'action': f"click '{target}'",
+                'position': (click_result['x'], click_result['y']),
+                'time_ms': dt,
+            }
+        else:
+            # InfiGUI n'a pas trouvé le bouton — essayer le clic direct via OCR
+            print(f"⚠️ [Dialog] InfiGUI n'a pas trouvé '{target}', essai OCR direct")
+            ocr_result = self._click_via_ocr(target, screenshot_pil)
+            dt = (time.time() - t0) * 1000
+
+            if ocr_result:
+                print(f"✅ [Dialog] OCR clic '{target}' à ({ocr_result[0]}, {ocr_result[1]}) ({dt:.0f}ms)")
+                return {
+                    'handled': True,
+                    'title': title,
+                    'dialog_type': dialog_key,
+                    'action': f"click '{target}' (OCR)",
+                    'position': ocr_result,
+                    'time_ms': dt,
+                }
+
+            print(f"❌ [Dialog] Impossible de cliquer '{target}' ({dt:.0f}ms)")
+            return {
+                'handled': False,
+                'title': title,
+                'dialog_type': dialog_key,
+                'reason': f"Bouton '{target}' introuvable",
+                'time_ms': dt,
+            }
+
+    # ------------------------------------------------------------------
+    # Lecture titre
+    # ------------------------------------------------------------------
+
+    def _read_title(self, screenshot_pil) -> str:
+        """Lit TOUT le texte visible via EasyOCR full-screen (~500ms).
+
+        En VM QEMU, la barre de titre Windows est à l'intérieur du framebuffer,
+        pas en haut absolu de l'écran. On fait l'OCR full-screen et on cherche
+        les mots-clés des dialogues connus dans le texte complet.
+        """
+        try:
+            import numpy as np
+
+            reader = self._get_easyocr()
+            if reader is None:
+                return ""
+
+            results = reader.readtext(np.array(screenshot_pil))
+            full_text = ' '.join(r[1] for r in results if r[1].strip())
+            return full_text
+
+        except Exception as e:
+            print(f"⚠️ [Dialog] Erreur lecture écran: {e}")
+            return ""
+
+    # ------------------------------------------------------------------
+    # Clic via InfiGUI (serveur grounding)
+    # ------------------------------------------------------------------
+
+    def _click_via_infigui(
+        self, target: str, description: str, screenshot_pil
+    ) -> Optional[Dict]:
+        """Demande à InfiGUI (subprocess one-shot) de localiser et cliquer sur le bouton."""
+        try:
+            from core.grounding.ui_tars_grounder import UITarsGrounder
+
+            grounder = UITarsGrounder.get_instance()
+            result = grounder.ground(
+                target_text=target,
+                target_description=description,
+                screen_pil=screenshot_pil,
+            )
+
+            if result and result.x is not None:
+                import pyautogui
+                pyautogui.click(result.x, result.y)
+                return {'x': result.x, 'y': result.y}
+
+            return None
+
+        except Exception as e:
+            print(f"⚠️ [Dialog/InfiGUI] Erreur: {e}")
+            return None
+
+    # ------------------------------------------------------------------
+    # Clic via OCR (fallback rapide)
+    # ------------------------------------------------------------------
+
+    def _click_via_ocr(self, target: str, screenshot_pil) -> Optional[tuple]:
+        """Cherche le bouton par OCR et clique dessus."""
+        try:
+            import numpy as np
+
+            reader = self._get_easyocr()
+            if reader is None:
+                return None
+
+            results = reader.readtext(np.array(screenshot_pil))
+
+            target_lower = target.lower()
+            matches = []
+            for (bbox_pts, text, conf) in results:
+                if target_lower in text.lower() or text.lower() in target_lower:
+                    x = int(sum(p[0] for p in bbox_pts) / 4)
+                    y = int(sum(p[1] for p in bbox_pts) / 4)
+                    matches.append((x, y, text))
+
+            if matches:
+                # Prendre le match le plus bas (boutons = bas du dialogue)
+                best = max(matches, key=lambda m: m[1])
+                import pyautogui
+                pyautogui.click(best[0], best[1])
+                return (best[0], best[1])
+
+            return None
+
+        except Exception as e:
+            print(f"⚠️ [Dialog/OCR] Erreur: {e}")
+            return None
+
+    # ------------------------------------------------------------------
+    # EasyOCR singleton
+    # ------------------------------------------------------------------
+
+    def _get_easyocr(self):
+        if self._easyocr_reader is not None:
+            return self._easyocr_reader
+
+        try:
+            import easyocr
+            self._easyocr_reader = easyocr.Reader(
+                ['fr', 'en'], gpu=True, verbose=False
+            )
+            return self._easyocr_reader
+        except ImportError:
+            return None

core/grounding/element_signature.py

@@ -0,0 +1,239 @@
+"""
+core/grounding/element_signature.py — Signatures d'éléments UI apprises
+
+Chaque élément cliqué avec succès enrichit sa signature :
+- texte OCR, type, position relative, voisins contextuels
+- nombre de succès/échecs, confiance moyenne
+- variantes observées (résolutions, positions)
+
+Les signatures sont stockées en SQLite pour un lookup rapide.
+Pattern identique à TargetMemoryStore (validé en prod).
+
+Utilisation :
+    from core.grounding.element_signature import SignatureStore
+
+    store = SignatureStore()
+
+    # Après un clic réussi
+    store.record_success("btn_valider", "notepad_1920x1080", element, confidence=0.92)
+
+    # Au replay
+    sig = store.lookup("btn_valider", "notepad_1920x1080")
+    if sig:
+        print(f"Signature connue : {sig['text']} position={sig['relative_position']}")
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import sqlite3
+import threading
+import time
+from typing import Any, Dict, List, Optional
+
+from core.grounding.fast_types import DetectedUIElement
+
+# Chemin par défaut de la DB
+_DEFAULT_DB = os.path.join(
+    os.path.dirname(os.path.dirname(os.path.dirname(__file__))),
+    "data", "learning", "element_signatures.db",
+)
+
+
+class SignatureStore:
+    """Stockage SQLite des signatures d'éléments UI appris."""
+
+    def __init__(self, db_path: str = _DEFAULT_DB):
+        self.db_path = db_path
+        self._lock = threading.Lock()
+        self._ensure_db()
+
+    def _ensure_db(self):
+        """Crée la DB et la table si nécessaire."""
+        os.makedirs(os.path.dirname(self.db_path), exist_ok=True)
+        with sqlite3.connect(self.db_path) as conn:
+            conn.execute("""
+                CREATE TABLE IF NOT EXISTS signatures (
+                    target_key TEXT NOT NULL,
+                    screen_context TEXT NOT NULL,
+                    text TEXT DEFAULT '',
+                    element_type TEXT DEFAULT 'element',
+                    relative_position TEXT DEFAULT '',
+                    neighbors TEXT DEFAULT '[]',
+                    success_count INTEGER DEFAULT 0,
+                    fail_count INTEGER DEFAULT 0,
+                    avg_confidence REAL DEFAULT 0.0,
+                    last_seen TEXT DEFAULT '',
+                    variants TEXT DEFAULT '[]',
+                    PRIMARY KEY (target_key, screen_context)
+                )
+            """)
+            conn.execute("""
+                CREATE INDEX IF NOT EXISTS idx_target_key
+                ON signatures(target_key)
+            """)
+
+    # ------------------------------------------------------------------
+    # Lookup
+    # ------------------------------------------------------------------
+
+    def lookup(self, target_key: str, screen_context: str = "") -> Optional[Dict[str, Any]]:
+        """Cherche une signature connue.
+
+        Args:
+            target_key: Clé unique de la cible (hash du texte + description).
+            screen_context: Contexte d'écran (hash titre fenêtre + résolution).
+
+        Returns:
+            Dict avec les champs de la signature, ou None.
+        """
+        with sqlite3.connect(self.db_path) as conn:
+            conn.row_factory = sqlite3.Row
+            # Chercher avec le contexte exact d'abord
+            row = conn.execute(
+                "SELECT * FROM signatures WHERE target_key = ? AND screen_context = ?",
+                (target_key, screen_context),
+            ).fetchone()
+
+            # Fallback : chercher sans contexte (toutes les variantes)
+            if row is None and screen_context:
+                row = conn.execute(
+                    "SELECT * FROM signatures WHERE target_key = ? ORDER BY success_count DESC LIMIT 1",
+                    (target_key,),
+                ).fetchone()
+
+            if row is None:
+                return None
+
+            return {
+                "target_key": row["target_key"],
+                "screen_context": row["screen_context"],
+                "text": row["text"],
+                "element_type": row["element_type"],
+                "relative_position": row["relative_position"],
+                "neighbors": json.loads(row["neighbors"]),
+                "success_count": row["success_count"],
+                "fail_count": row["fail_count"],
+                "avg_confidence": row["avg_confidence"],
+                "last_seen": row["last_seen"],
+                "variants": json.loads(row["variants"]),
+            }
+
+    # ------------------------------------------------------------------
+    # Enregistrement
+    # ------------------------------------------------------------------
+
+    def record_success(
+        self,
+        target_key: str,
+        screen_context: str,
+        element: DetectedUIElement,
+        confidence: float,
+    ):
+        """Enregistre un succès — crée ou enrichit la signature."""
+        with self._lock:
+            existing = self.lookup(target_key, screen_context)
+            now = time.strftime("%Y-%m-%dT%H:%M:%S")
+
+            if existing:
+                # Enrichir la signature existante
+                n = existing["success_count"]
+                new_avg = (existing["avg_confidence"] * n + confidence) / (n + 1)
+
+                # Ajouter la variante si position différente
+                variants = existing["variants"]
+                variant = {
+                    "position": element.relative_position,
+                    "center": list(element.center),
+                    "confidence": confidence,
+                    "timestamp": now,
+                }
+                variants.append(variant)
+                # Garder les 20 dernières variantes max
+                variants = variants[-20:]
+
+                # Mettre à jour les voisins (union)
+                neighbors = list(set(existing["neighbors"] + element.neighbors))[:10]
+
+                with sqlite3.connect(self.db_path) as conn:
+                    conn.execute("""
+                        UPDATE signatures SET
+                            success_count = success_count + 1,
+                            avg_confidence = ?,
+                            last_seen = ?,
+                            neighbors = ?,
+                            variants = ?,
+                            relative_position = ?
+                        WHERE target_key = ? AND screen_context = ?
+                    """, (
+                        new_avg, now,
+                        json.dumps(neighbors),
+                        json.dumps(variants),
+                        element.relative_position,
+                        target_key, screen_context,
+                    ))
+            else:
+                # Créer une nouvelle signature
+                with sqlite3.connect(self.db_path) as conn:
+                    conn.execute("""
+                        INSERT INTO signatures
+                        (target_key, screen_context, text, element_type, relative_position,
+                         neighbors, success_count, fail_count, avg_confidence, last_seen, variants)
+                        VALUES (?, ?, ?, ?, ?, ?, 1, 0, ?, ?, ?)
+                    """, (
+                        target_key, screen_context,
+                        element.ocr_text,
+                        element.element_type,
+                        element.relative_position,
+                        json.dumps(element.neighbors[:10]),
+                        confidence, now,
+                        json.dumps([{
+                            "position": element.relative_position,
+                            "center": list(element.center),
+                            "confidence": confidence,
+                            "timestamp": now,
+                        }]),
+                    ))
+
+            print(f"📝 [Signature] '{target_key}' {'enrichie' if existing else 'créée'} "
+                  f"(conf={confidence:.2f}, ctx='{screen_context[:30]}')")
+
+    def record_failure(self, target_key: str, screen_context: str):
+        """Enregistre un échec pour une signature."""
+        with self._lock:
+            with sqlite3.connect(self.db_path) as conn:
+                conn.execute("""
+                    UPDATE signatures SET fail_count = fail_count + 1, last_seen = ?
+                    WHERE target_key = ? AND screen_context = ?
+                """, (time.strftime("%Y-%m-%dT%H:%M:%S"), target_key, screen_context))
+
+    # ------------------------------------------------------------------
+    # Utilitaires
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def make_target_key(text: str, description: str = "") -> str:
+        """Génère une clé unique pour une cible."""
+        raw = f"{text.lower().strip()}|{description.lower().strip()}"
+        return hashlib.md5(raw.encode()).hexdigest()[:16]
+
+    @staticmethod
+    def make_screen_context(window_title: str, resolution: tuple = (0, 0)) -> str:
+        """Génère un contexte d'écran."""
+        raw = f"{window_title.lower().strip()}|{resolution[0]}x{resolution[1]}"
+        return hashlib.md5(raw.encode()).hexdigest()[:12]
+
+    def get_stats(self) -> Dict[str, Any]:
+        """Statistiques de la base de signatures."""
+        with sqlite3.connect(self.db_path) as conn:
+            total = conn.execute("SELECT COUNT(*) FROM signatures").fetchone()[0]
+            reliable = conn.execute(
+                "SELECT COUNT(*) FROM signatures WHERE success_count >= 3 AND fail_count = 0"
+            ).fetchone()[0]
+            return {
+                "total_signatures": total,
+                "reliable": reliable,
+                "db_path": self.db_path,
+            }

core/grounding/fast_detector.py

@@ -0,0 +1,326 @@
+"""
+core/grounding/fast_detector.py — Layer FAST : détection rapide des éléments UI
+
+Capture l'écran, détecte tous les éléments UI via RF-DETR (~120ms),
+enrichit chaque élément avec le texte OCR et le contexte spatial.
+
+Produit un ScreenSnapshot utilisable par le SmartMatcher.
+
+Utilisation :
+    from core.grounding.fast_detector import FastDetector
+
+    detector = FastDetector()
+    snapshot = detector.detect()
+    print(f"{len(snapshot.elements)} éléments en {snapshot.total_time_ms:.0f}ms")
+"""
+
+from __future__ import annotations
+
+import math
+import time
+from typing import Any, Dict, List, Optional, Tuple
+
+from core.grounding.fast_types import DetectedUIElement, ScreenSnapshot
+
+
+class FastDetector:
+    """Détection rapide de tous les éléments UI visibles sur l'écran.
+
+    Combine RF-DETR (détection bbox) + docTR (OCR) pour produire
+    un ScreenSnapshot enrichi.
+
+    Le modèle RF-DETR est un singleton chargé au premier appel (~1s),
+    puis les appels suivants sont rapides (~120ms).
+    """
+
+    def __init__(self, detection_threshold: float = 0.30):
+        self.detection_threshold = detection_threshold
+        self._last_snapshot: Optional[ScreenSnapshot] = None
+        self._last_phash: str = ""
+
+    def detect(
+        self,
+        screenshot_pil: Optional[Any] = None,
+        phash: str = "",
+        window_title: str = "",
+    ) -> ScreenSnapshot:
+        """Détecte et enrichit tous les éléments UI de l'écran.
+
+        Args:
+            screenshot_pil: Image PIL. Si None, capture via mss.
+            phash: Hash perceptuel pour le cache. Si identique au dernier, réutilise le cache.
+            window_title: Titre de la fenêtre active.
+
+        Returns:
+            ScreenSnapshot avec tous les éléments enrichis.
+        """
+        t0 = time.time()
+
+        # Cache : même écran → même résultat
+        if phash and phash == self._last_phash and self._last_snapshot is not None:
+            print(f"⚡ [FAST] Cache hit (pHash identique)")
+            return self._last_snapshot
+
+        # Capture si pas fourni
+        if screenshot_pil is None:
+            screenshot_pil = self._capture_screen()
+            if screenshot_pil is None:
+                return ScreenSnapshot(elements=[], ocr_words=[], resolution=(0, 0))
+
+        w, h = screenshot_pil.size
+
+        # --- Détection RF-DETR (~120ms) ---
+        t_det = time.time()
+        raw_elements = self._detect_rfdetr(screenshot_pil)
+        detection_ms = (time.time() - t_det) * 1000
+
+        # --- OCR sur les crops des éléments détectés (pas full screen) ---
+        t_ocr = time.time()
+        ocr_words = self._ocr_extract(screenshot_pil)
+        ocr_ms = (time.time() - t_ocr) * 1000
+
+        # --- Enrichissement : attribuer texte + voisins + position ---
+        enriched = self._enrich_elements(raw_elements, ocr_words, w, h)
+
+        total_ms = (time.time() - t0) * 1000
+
+        snapshot = ScreenSnapshot(
+            elements=enriched,
+            ocr_words=ocr_words,
+            resolution=(w, h),
+            window_title=window_title,
+            phash=phash,
+            detection_time_ms=detection_ms,
+            ocr_time_ms=ocr_ms,
+            total_time_ms=total_ms,
+        )
+
+        # Mettre en cache
+        if phash:
+            self._last_phash = phash
+            self._last_snapshot = snapshot
+
+        print(f"⚡ [FAST] {len(enriched)} éléments détectés en {total_ms:.0f}ms "
+              f"(det={detection_ms:.0f}ms, ocr={ocr_ms:.0f}ms)")
+
+        return snapshot
+
+    # ------------------------------------------------------------------
+    # Détection RF-DETR
+    # ------------------------------------------------------------------
+
+    def _detect_rfdetr(self, image) -> List[DetectedUIElement]:
+        """Détecte les éléments via RF-DETR (réutilise le singleton existant)."""
+        try:
+            import sys
+            sys.path.insert(0, 'visual_workflow_builder/backend')
+            from services.ui_detection_service import detect_ui_elements
+
+            result = detect_ui_elements(image, threshold=self.detection_threshold)
+
+            elements = []
+            for e in result.elements:
+                x1 = e.bbox["x1"]
+                y1 = e.bbox["y1"]
+                x2 = e.bbox["x2"]
+                y2 = e.bbox["y2"]
+                elements.append(DetectedUIElement(
+                    id=e.id,
+                    bbox=(x1, y1, x2, y2),
+                    center=(e.center["x"], e.center["y"]),
+                    confidence=e.confidence,
+                ))
+
+            return elements
+
+        except Exception as ex:
+            print(f"⚠️ [FAST/detect] RF-DETR erreur: {ex}")
+            return []
+
+    # ------------------------------------------------------------------
+    # OCR
+    # ------------------------------------------------------------------
+
+    _easyocr_reader = None  # Singleton EasyOCR (chargé une fois)
+
+    def _ocr_extract(self, image) -> List[Dict[str, Any]]:
+        """Extrait les mots visibles via EasyOCR (GPU, ~500ms).
+
+        Fallback sur docTR si EasyOCR non disponible.
+        """
+        try:
+            import numpy as np
+            import easyocr
+
+            # Singleton : charger le reader une seule fois
+            if FastDetector._easyocr_reader is None:
+                print(f"🔍 [FAST/ocr] Chargement EasyOCR (GPU)...")
+                FastDetector._easyocr_reader = easyocr.Reader(
+                    ['fr', 'en'], gpu=True, verbose=False
+                )
+
+            results = FastDetector._easyocr_reader.readtext(np.array(image))
+
+            words = []
+            for (bbox_pts, text, conf) in results:
+                if not text or len(text.strip()) < 1:
+                    continue
+                # bbox_pts = [[x1,y1],[x2,y1],[x2,y2],[x1,y2]]
+                x1 = int(min(p[0] for p in bbox_pts))
+                y1 = int(min(p[1] for p in bbox_pts))
+                x2 = int(max(p[0] for p in bbox_pts))
+                y2 = int(max(p[1] for p in bbox_pts))
+                words.append({
+                    'text': text.strip(),
+                    'bbox': [x1, y1, x2, y2],
+                    'confidence': float(conf),
+                })
+
+            return words
+
+        except ImportError:
+            # Fallback docTR
+            try:
+                import sys
+                sys.path.insert(0, 'visual_workflow_builder/backend')
+                from services.ocr_service import ocr_extract_words
+                return ocr_extract_words(image) or []
+            except Exception:
+                return []
+        except Exception as ex:
+            print(f"⚠️ [FAST/ocr] EasyOCR erreur: {ex}")
+            return []
+
+    # ------------------------------------------------------------------
+    # Enrichissement
+    # ------------------------------------------------------------------
+
+    def _enrich_elements(
+        self,
+        elements: List[DetectedUIElement],
+        ocr_words: List[Dict[str, Any]],
+        screen_w: int,
+        screen_h: int,
+    ) -> List[DetectedUIElement]:
+        """Enrichit chaque élément avec texte OCR, voisins et position relative."""
+
+        for elem in elements:
+            # 1. Attribuer le texte OCR par intersection bbox
+            elem.ocr_text = self._assign_ocr_text(elem, ocr_words)
+
+            # 2. Position relative dans l'écran (grille 3x3)
+            elem.relative_position = self._compute_relative_position(
+                elem.center, screen_w, screen_h
+            )
+
+            # 3. Classifier le type d'élément (heuristique taille + ratio)
+            elem.element_type = self._classify_element_type(elem)
+
+        # 4. Calculer les voisins (texte des éléments proches)
+        for elem in elements:
+            elem.neighbors = self._find_neighbors(elem, elements)
+
+        return elements
+
+    def _assign_ocr_text(
+        self,
+        elem: DetectedUIElement,
+        ocr_words: List[Dict[str, Any]],
+    ) -> str:
+        """Attribue le texte OCR à un élément par intersection géométrique."""
+        x1, y1, x2, y2 = elem.bbox
+        # Élargir la bbox de 20% pour capturer le texte autour
+        margin_x = int((x2 - x1) * 0.2)
+        margin_y = int((y2 - y1) * 0.2)
+        ex1, ey1 = x1 - margin_x, y1 - margin_y
+        ex2, ey2 = x2 + margin_x, y2 + margin_y
+
+        texts = []
+        for word in ocr_words:
+            wb = word.get('bbox', [0, 0, 0, 0])
+            if len(wb) < 4:
+                continue
+            wx1, wy1, wx2, wy2 = wb[0], wb[1], wb[2], wb[3]
+            # Intersection ?
+            if wx1 < ex2 and wx2 > ex1 and wy1 < ey2 and wy2 > ey1:
+                text = word.get('text', '').strip()
+                if text and len(text) > 1:
+                    texts.append(text)
+
+        return ' '.join(texts)
+
+    @staticmethod
+    def _compute_relative_position(
+        center: Tuple[int, int],
+        screen_w: int,
+        screen_h: int,
+    ) -> str:
+        """Calcule la position relative dans une grille 3x3."""
+        cx, cy = center
+        col = "left" if cx < screen_w / 3 else ("right" if cx > 2 * screen_w / 3 else "center")
+        row = "top" if cy < screen_h / 3 else ("bottom" if cy > 2 * screen_h / 3 else "middle")
+        return f"{row}_{col}"
+
+    @staticmethod
+    def _classify_element_type(elem: DetectedUIElement) -> str:
+        """Classifie le type d'élément par heuristique taille/ratio."""
+        w, h = elem.width, elem.height
+        if w == 0 or h == 0:
+            return "element"
+        ratio = w / h
+        area = w * h
+
+        # Petit carré → icône
+        if area < 5000 and 0.5 < ratio < 2.0:
+            return "icon"
+        # Large et fin → bouton ou champ
+        if ratio > 3.0 and h < 60:
+            return "input"
+        if ratio > 2.0 and h < 50:
+            return "button"
+        # Grand bloc → zone de contenu
+        if area > 50000:
+            return "container"
+
+        return "element"
+
+    @staticmethod
+    def _find_neighbors(
+        elem: DetectedUIElement,
+        all_elements: List[DetectedUIElement],
+        max_neighbors: int = 5,
+    ) -> List[str]:
+        """Trouve les textes OCR des éléments proches (rayon 1.5x diagonale)."""
+        diag = math.sqrt(elem.width**2 + elem.height**2)
+        radius = max(diag * 1.5, 100)  # minimum 100px
+
+        neighbors = []
+        for other in all_elements:
+            if other.id == elem.id or not other.ocr_text:
+                continue
+            dx = other.center[0] - elem.center[0]
+            dy = other.center[1] - elem.center[1]
+            dist = math.sqrt(dx**2 + dy**2)
+            if dist < radius:
+                neighbors.append(other.ocr_text)
+
+        return neighbors[:max_neighbors]
+
+    # ------------------------------------------------------------------
+    # Capture écran
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _capture_screen():
+        """Capture l'écran via mss."""
+        try:
+            import mss
+            from PIL import Image
+
+            with mss.mss() as sct:
+                mon = sct.monitors[0]
+                grab = sct.grab(mon)
+                return Image.frombytes('RGB', grab.size, grab.bgra, 'raw', 'BGRX')
+        except Exception as ex:
+            print(f"⚠️ [FAST/capture] Erreur: {ex}")
+            return None

core/grounding/fast_pipeline.py

@@ -0,0 +1,216 @@
+"""
+core/grounding/fast_pipeline.py — Pipeline FAST → SMART → THINK
+
+Orchestrateur central : détecte les éléments (FAST), matche avec la cible (SMART),
+et demande au VLM de trancher si le score est trop bas (THINK).
+
+Seuils de confiance :
+  ≥ 0.90  →  action directe (FAST/SMART)
+  0.60-0.90  →  VLM confirme (THINK)
+  < 0.60  →  VLM cherche seul (THINK)
+
+L'ancien GroundingPipeline est utilisé en fallback si tout échoue.
+
+Utilisation :
+    from core.grounding.fast_pipeline import FastSmartThinkPipeline
+    from core.grounding.target import GroundingTarget
+
+    pipeline = FastSmartThinkPipeline()
+    result = pipeline.locate(GroundingTarget(text="Valider"))
+    if result:
+        print(f"({result.x}, {result.y}) via {result.method} en {result.time_ms:.0f}ms")
+"""
+
+from __future__ import annotations
+
+import time
+import threading
+from typing import Optional
+
+from core.grounding.target import GroundingTarget, GroundingResult
+from core.grounding.fast_types import LocateResult
+from core.grounding.fast_detector import FastDetector
+from core.grounding.smart_matcher import SmartMatcher
+from core.grounding.think_arbiter import ThinkArbiter
+from core.grounding.element_signature import SignatureStore
+
+
+# Singleton
+_instance: Optional[FastSmartThinkPipeline] = None
+_instance_lock = threading.Lock()
+
+
+class FastSmartThinkPipeline:
+    """Pipeline FAST → SMART → THINK pour la localisation d'éléments UI.
+
+    Chaque appel à locate() suit la cascade :
+    1. FAST : détection RF-DETR + OCR enrichissement (~120ms+1s)
+    2. SMART : matching texte/type/position/voisins (< 1ms)
+    3. THINK : VLM arbitre si score insuffisant (~3-5s)
+    4. Fallback : ancien pipeline si tout échoue
+    """
+
+    def __init__(
+        self,
+        confidence_direct: float = 0.90,
+        confidence_think: float = 0.60,
+        enable_think: bool = True,
+        enable_learning: bool = True,
+    ):
+        self.confidence_direct = confidence_direct
+        self.confidence_think = confidence_think
+        self.enable_think = enable_think
+        self.enable_learning = enable_learning
+
+        self._detector = FastDetector()
+        self._matcher = SmartMatcher()
+        self._arbiter = ThinkArbiter()
+        self._signatures = SignatureStore()
+        self._fallback_pipeline = None
+
+    @classmethod
+    def get_instance(cls) -> FastSmartThinkPipeline:
+        """Retourne l'instance singleton."""
+        global _instance
+        if _instance is None:
+            with _instance_lock:
+                if _instance is None:
+                    _instance = cls()
+        return _instance
+
+    def set_fallback_pipeline(self, pipeline) -> None:
+        """Configure l'ancien pipeline comme safety net."""
+        self._fallback_pipeline = pipeline
+
+    # ------------------------------------------------------------------
+    # API principale
+    # ------------------------------------------------------------------
+
+    def locate(
+        self,
+        target: GroundingTarget,
+        screenshot_pil=None,
+        phash: str = "",
+        window_title: str = "",
+    ) -> Optional[GroundingResult]:
+        """Localise un élément UI via la cascade FAST → SMART → THINK.
+
+        Args:
+            target: Ce qu'on cherche (texte, description, bbox d'origine).
+            screenshot_pil: Image PIL. Si None, capture via mss.
+            phash: Hash perceptuel pour le cache.
+            window_title: Titre de la fenêtre active.
+
+        Returns:
+            GroundingResult compatible avec le pipeline existant, ou None.
+        """
+        t0 = time.time()
+
+        # --- FAST : détecter tous les éléments ---
+        snapshot = self._detector.detect(
+            screenshot_pil=screenshot_pil,
+            phash=phash,
+            window_title=window_title,
+        )
+
+        if not snapshot.elements:
+            print(f"⚡ [Pipeline] FAST : aucun élément détecté")
+            return self._try_fallback(target)
+
+        # --- Lookup signature apprise ---
+        target_key = SignatureStore.make_target_key(
+            target.text or "", target.description or ""
+        )
+        screen_ctx = SignatureStore.make_screen_context(
+            window_title, snapshot.resolution
+        )
+        signature = self._signatures.lookup(target_key, screen_ctx)
+
+        # --- SMART : matcher avec la cible ---
+        candidate = self._matcher.match(snapshot, target, signature)
+
+        if candidate:
+            dt = (time.time() - t0) * 1000
+
+            # Score suffisant → action directe
+            if candidate.score >= self.confidence_direct:
+                print(f"✅ [Pipeline] FAST→SMART direct : '{candidate.element.ocr_text}' "
+                      f"score={candidate.score:.3f} ({candidate.method}) "
+                      f"→ ({candidate.element.center[0]}, {candidate.element.center[1]}) "
+                      f"en {dt:.0f}ms")
+
+                # Apprentissage
+                if self.enable_learning:
+                    self._signatures.record_success(
+                        target_key, screen_ctx,
+                        candidate.element, candidate.score,
+                    )
+
+                return GroundingResult(
+                    x=candidate.element.center[0],
+                    y=candidate.element.center[1],
+                    method=f"fast_{candidate.method}",
+                    confidence=candidate.score,
+                    time_ms=dt,
+                )
+
+            # Score moyen → demander au VLM de confirmer
+            if candidate.score >= self.confidence_think and self.enable_think:
+                print(f"🤔 [Pipeline] SMART score={candidate.score:.3f} — THINK pour confirmer")
+                think_result = self._arbiter.arbitrate(
+                    target,
+                    candidates=[candidate],
+                    screenshot_pil=screenshot_pil or snapshot.elements[0] if False else screenshot_pil,
+                )
+                dt = (time.time() - t0) * 1000
+
+                if think_result:
+                    # VLM a confirmé
+                    if self.enable_learning:
+                        self._signatures.record_success(
+                            target_key, screen_ctx,
+                            candidate.element, think_result.confidence,
+                        )
+                    return GroundingResult(
+                        x=think_result.x, y=think_result.y,
+                        method="smart_think_confirmed",
+                        confidence=think_result.confidence,
+                        time_ms=dt,
+                    )
+
+        # --- THINK : score trop bas ou pas de candidat → VLM cherche seul ---
+        if self.enable_think:
+            score_info = f"score={candidate.score:.3f}" if candidate else "aucun candidat"
+            print(f"🤔 [Pipeline] {score_info} — THINK recherche complète")
+            think_result = self._arbiter.arbitrate(
+                target, candidates=[], screenshot_pil=screenshot_pil,
+            )
+            dt = (time.time() - t0) * 1000
+
+            if think_result:
+                return GroundingResult(
+                    x=think_result.x, y=think_result.y,
+                    method="think_vlm",
+                    confidence=think_result.confidence,
+                    time_ms=dt,
+                )
+
+        # --- Fallback : ancien pipeline ---
+        return self._try_fallback(target)
+
+    # ------------------------------------------------------------------
+    # Fallback
+    # ------------------------------------------------------------------
+
+    def _try_fallback(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """Tente l'ancien pipeline en dernier recours."""
+        if self._fallback_pipeline is None:
+            print(f"❌ [Pipeline] Aucune méthode n'a trouvé '{target.text}'")
+            return None
+
+        print(f"⚠️ [Pipeline] Fallback ancien pipeline pour '{target.text}'")
+        try:
+            return self._fallback_pipeline.locate(target)
+        except Exception as ex:
+            print(f"⚠️ [Pipeline] Fallback échoué: {ex}")
+            return None

core/grounding/fast_types.py

@@ -0,0 +1,81 @@
+"""
+core/grounding/fast_types.py — Structures de données pour le pipeline FAST→SMART→THINK
+
+Utilisées exclusivement par le pipeline de localisation rapide.
+Compatibles avec GroundingTarget/GroundingResult existants via conversion.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple
+
+
+@dataclass
+class DetectedUIElement:
+    """Élément UI détecté par le layer FAST (RF-DETR) puis enrichi par OCR."""
+    id: int
+    bbox: Tuple[int, int, int, int]      # (x1, y1, x2, y2) pixels absolus
+    center: Tuple[int, int]              # (cx, cy)
+    confidence: float                     # confidence détecteur (0-1)
+    element_type: str = "element"        # "button", "input", "icon", "text", "element"
+    ocr_text: str = ""                   # texte OCR extrait de la région
+    neighbors: List[str] = field(default_factory=list)  # textes des éléments proches
+    relative_position: str = ""          # "top_left", "center", "bottom_right", etc.
+
+    @property
+    def width(self) -> int:
+        return self.bbox[2] - self.bbox[0]
+
+    @property
+    def height(self) -> int:
+        return self.bbox[3] - self.bbox[1]
+
+    @property
+    def area(self) -> int:
+        return self.width * self.height
+
+
+@dataclass
+class ScreenSnapshot:
+    """État complet de l'écran à un instant t — sortie du layer FAST."""
+    elements: List[DetectedUIElement]
+    ocr_words: List[Dict[str, Any]]      # mots OCR bruts [{text, bbox}]
+    resolution: Tuple[int, int]          # (width, height)
+    window_title: str = ""
+    phash: str = ""
+    detection_time_ms: float = 0.0
+    ocr_time_ms: float = 0.0
+    total_time_ms: float = 0.0
+
+
+@dataclass
+class MatchCandidate:
+    """Résultat du matching SMART pour un élément candidat."""
+    element: DetectedUIElement
+    score: float                          # score combiné (0-1)
+    score_detail: Dict[str, float] = field(default_factory=dict)
+    method: str = ""                     # "exact_text", "fuzzy_text", "position", etc.
+
+
+@dataclass
+class LocateResult:
+    """Résultat final du pipeline FAST→SMART→THINK."""
+    x: int
+    y: int
+    confidence: float
+    method: str                          # "fast_exact", "fast_fuzzy", "smart_vote", "think_vlm"
+    time_ms: float
+    tier: str = "fast"                   # "fast", "smart", "think"
+    element: Optional[DetectedUIElement] = None
+    candidates_count: int = 0
+
+    def to_grounding_result(self):
+        """Conversion vers GroundingResult pour compatibilité."""
+        from core.grounding.target import GroundingResult
+        return GroundingResult(
+            x=self.x, y=self.y,
+            method=self.method,
+            confidence=self.confidence,
+            time_ms=self.time_ms,
+        )

core/grounding/infigui_worker.py

@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+"""
+Worker InfiGUI — process indépendant, communication par fichiers.
+
+Charge le modèle, surveille /tmp/infigui_request.json, infère, écrit /tmp/infigui_response.json.
+
+Lancement :
+    cd ~/ai/rpa_vision_v3
+    .venv/bin/python3 -m core.grounding.infigui_worker
+"""
+
+import json
+import math
+import os
+import re
+import sys
+import time
+import gc
+import warnings
+
+warnings.filterwarnings("ignore")
+
+import torch
+
+REQUEST_FILE = "/tmp/infigui_request.json"
+RESPONSE_FILE = "/tmp/infigui_response.json"
+READY_FILE = "/tmp/infigui_ready"
+
+
+def load_model():
+    """Charge InfiGUI-G1-3B en 4-bit NF4."""
+    torch.cuda.empty_cache()
+    gc.collect()
+
+    from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor, BitsAndBytesConfig
+
+    model_id = "InfiX-ai/InfiGUI-G1-3B"
+    print(f"[infigui-worker] Chargement {model_id}...")
+
+    bnb = BitsAndBytesConfig(
+        load_in_4bit=True, bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True,
+    )
+    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+        model_id, quantization_config=bnb, device_map={"": "cuda:0"},
+    )
+    model.eval()
+    processor = AutoProcessor.from_pretrained(
+        model_id, padding_side="left",
+        min_pixels=100 * 28 * 28, max_pixels=5600 * 28 * 28,
+    )
+
+    vram = torch.cuda.memory_allocated() / 1e9
+    print(f"[infigui-worker] Prêt — VRAM: {vram:.2f}GB")
+
+    # Signal "prêt"
+    with open(READY_FILE, "w") as f:
+        f.write(f"ready {vram:.2f}GB")
+
+    return model, processor
+
+
+def infer(model, processor, req):
+    """Fait une inférence.
+
+    Modes :
+        - texte seul (target/description) : grounding classique
+        - fusionné (anchor_image_path présent) : on passe en plus le crop d'ancre
+          comme image de référence et le modèle doit retrouver cet élément sur
+          le screenshot. Évite la double passe describe→ground.
+    """
+    from PIL import Image
+    from qwen_vl_utils import process_vision_info
+
+    target = req.get("target", "")
+    description = req.get("description", "")
+    label = f"{target} — {description}" if description else target
+
+    # Image principale (screenshot complet)
+    image_path = req.get("image_path", "")
+    if image_path and os.path.exists(image_path):
+        img = Image.open(image_path).convert("RGB")
+    else:
+        import mss
+        with mss.mss() as sct:
+            grab = sct.grab(sct.monitors[0])
+            img = Image.frombytes("RGB", grab.size, grab.bgra, "raw", "BGRX")
+
+    # Image d'ancre (optionnelle) — mode fusionné describe+ground
+    anchor_image_path = req.get("anchor_image_path", "")
+    anchor_img = None
+    if anchor_image_path and os.path.exists(anchor_image_path):
+        anchor_img = Image.open(anchor_image_path).convert("RGB")
+
+    if not label.strip() and anchor_img is None:
+        return {"x": None, "y": None, "error": "target ou anchor_image requis"}
+
+    W, H = img.size
+    factor = 28
+    rH = max(factor, round(H / factor) * factor)
+    rW = max(factor, round(W / factor) * factor)
+
+    system = (
+        "You FIRST think about the reasoning process as an internal monologue "
+        "and then provide the final answer.\n"
+        "The reasoning process MUST BE enclosed within <think> </think> tags."
+    )
+
+    # Construction du prompt selon le mode
+    if anchor_img is not None:
+        # Mode fusionné : Image1 = crop d'ancre, Image2 = screenshot
+        hint = f' Hint: this element looks like "{label}".' if label.strip() else ""
+        user_text = (
+            f"The first image is a small crop of a UI element captured previously. "
+            f"The second image is the current screen ({rW}x{rH}).{hint}\n"
+            f"Locate on the second image the UI element that visually matches the first image. "
+            f"Output the coordinates using JSON format: "
+            f'[{{"point_2d": [x, y]}}, ...]'
+        )
+        messages = [
+            {"role": "system", "content": system},
+            {"role": "user", "content": [
+                {"type": "image", "image": anchor_img},
+                {"type": "image", "image": img},
+                {"type": "text", "text": user_text},
+            ]},
+        ]
+    else:
+        # Mode classique : texte seul
+        user_text = (
+            f'The screen\'s resolution is {rW}x{rH}.\n'
+            f'Locate the UI element(s) for "{label}", '
+            f'output the coordinates using JSON format: '
+            f'[{{"point_2d": [x, y]}}, ...]'
+        )
+        messages = [
+            {"role": "system", "content": system},
+            {"role": "user", "content": [
+                {"type": "image", "image": img},
+                {"type": "text", "text": user_text},
+            ]},
+        ]
+
+    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    image_inputs, video_inputs = process_vision_info(messages)
+    inputs = processor(
+        text=[text], images=image_inputs, videos=video_inputs,
+        padding=True, return_tensors="pt",
+    ).to(model.device)
+
+    t0 = time.time()
+    with torch.no_grad():
+        gen = model.generate(**inputs, max_new_tokens=512)
+    infer_ms = (time.time() - t0) * 1000
+
+    trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, gen)]
+    raw = processor.batch_decode(
+        trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False,
+    )[0].strip()
+
+    mode_str = "fused" if anchor_img is not None else "text"
+    print(f"[infigui-worker] [{mode_str}] '{label[:40]}' ({infer_ms:.0f}ms)")
+
+    # Parser JSON point_2d
+    json_part = raw.split("</think>")[-1] if "</think>" in raw else raw
+    json_part = json_part.replace("```json", "").replace("```", "").strip()
+
+    px, py = None, None
+    try:
+        parsed = json.loads(json_part)
+        if isinstance(parsed, list) and len(parsed) > 0:
+            pt = parsed[0].get("point_2d", [])
+            if len(pt) >= 2:
+                px = int(pt[0] * W / rW)
+                py = int(pt[1] * H / rH)
+    except json.JSONDecodeError:
+        m = re.search(r'"point_2d"\s*:\s*\[(\d+),\s*(\d+)\]', raw)
+        if m:
+            px = int(int(m.group(1)) * W / rW)
+            py = int(int(m.group(2)) * H / rH)
+
+    return {
+        "x": px, "y": py,
+        "method": "infigui",
+        "confidence": 0.90 if px else 0.0,
+        "time_ms": round(infer_ms, 1),
+    }
+
+
+def main():
+    """Mode one-shot : lit une requête sur stdin, infère, écrit le résultat sur stdout."""
+    # Lire la requête
+    input_data = sys.stdin.read().strip()
+    if not input_data:
+        print(json.dumps({"x": None, "y": None, "error": "pas de requête"}))
+        return
+
+    try:
+        req = json.loads(input_data)
+    except json.JSONDecodeError:
+        print(json.dumps({"x": None, "y": None, "error": "JSON invalide"}))
+        return
+
+    model, processor = load_model()
+    result = infer(model, processor, req)
+    print(json.dumps(result))
+
+
+if __name__ == "__main__":
+    main()

core/grounding/pipeline.py

@@ -0,0 +1,190 @@
+"""
+core/grounding/pipeline.py — Pipeline de grounding en cascade
+
+Orchestre les methodes de localisation dans l'ordre :
+1. Template matching (TemplateMatcher, local, ~80ms)
+2. OCR (docTR via input_handler, local, ~1s)
+3. UI-TARS (HTTP vers serveur grounding, ~3s)
+4. Static fallback (coordonnees d'origine du workflow)
+
+Chaque methode est essayee dans l'ordre. Des qu'une reussit, on retourne
+le resultat. Cela permet un equilibre entre vitesse (template) et robustesse
+(UI-TARS pour les elements qui ont change de position/apparence).
+
+Utilisation :
+    from core.grounding.pipeline import GroundingPipeline
+    from core.grounding.target import GroundingTarget
+
+    pipeline = GroundingPipeline()
+    result = pipeline.locate(GroundingTarget(
+        text="Valider",
+        description="bouton vert en bas",
+        template_b64=screenshot_b64,
+        original_bbox={"x": 100, "y": 200, "width": 80, "height": 30},
+    ))
+    if result:
+        print(f"Trouve a ({result.x}, {result.y}) via {result.method}")
+"""
+
+from __future__ import annotations
+
+import time
+from typing import Optional
+
+from core.grounding.target import GroundingTarget, GroundingResult
+
+
+class GroundingPipeline:
+    """Pipeline de localisation en cascade : template -> OCR -> UI-TARS -> static."""
+
+    def __init__(self, template_threshold: float = 0.75, enable_uitars: bool = True):
+        self.template_threshold = template_threshold
+        self.enable_uitars = enable_uitars
+
+    def locate(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """Localise un element UI en essayant les methodes en cascade.
+
+        Args:
+            target: description de l'element a localiser
+
+        Returns:
+            GroundingResult ou None si aucune methode ne trouve l'element
+        """
+        t0 = time.time()
+
+        # --- Methode 1 : Template matching (~80ms) ---
+        result = self._try_template(target)
+        if result:
+            print(f"[GroundingPipeline] Localise via {result.method} en "
+                  f"{(time.time() - t0) * 1000:.0f}ms")
+            return result
+
+        # --- Methode 2 : OCR texte (~1s) ---
+        result = self._try_ocr(target)
+        if result:
+            print(f"[GroundingPipeline] Localise via {result.method} en "
+                  f"{(time.time() - t0) * 1000:.0f}ms")
+            return result
+
+        # --- Methode 3 : UI-TARS via serveur HTTP (~3s) ---
+        if self.enable_uitars:
+            result = self._try_uitars(target)
+            if result:
+                print(f"[GroundingPipeline] Localise via {result.method} en "
+                      f"{(time.time() - t0) * 1000:.0f}ms")
+                return result
+
+        # --- Methode 4 : Fallback statique ---
+        result = self._try_static(target)
+        if result:
+            print(f"[GroundingPipeline] Localise via {result.method} en "
+                  f"{(time.time() - t0) * 1000:.0f}ms")
+            return result
+
+        print(f"[GroundingPipeline] ECHEC: '{target.text}' introuvable "
+              f"(toutes methodes epuisees, {(time.time() - t0) * 1000:.0f}ms)")
+        return None
+
+    # ------------------------------------------------------------------
+    # Methodes individuelles
+    # ------------------------------------------------------------------
+
+    def _try_template(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """Template matching — rapide, exact, mais sensible aux changements visuels."""
+        if not target.template_b64:
+            return None
+
+        try:
+            from core.grounding.template_matcher import TemplateMatcher
+            matcher = TemplateMatcher(threshold=self.template_threshold)
+            match = matcher.match_screen(anchor_b64=target.template_b64)
+            if match:
+                print(f"[GroundingPipeline/template] score={match.score:.3f} "
+                      f"pos=({match.x},{match.y}) ({match.time_ms:.0f}ms)")
+                return GroundingResult(
+                    x=match.x,
+                    y=match.y,
+                    method='template',
+                    confidence=match.score,
+                    time_ms=match.time_ms,
+                )
+            else:
+                diag = matcher.match_screen_diagnostic(anchor_b64=target.template_b64)
+                print(f"[GroundingPipeline/template] pas de match — best={diag}")
+        except Exception as e:
+            print(f"[GroundingPipeline/template] ERREUR: {e}")
+
+        return None
+
+    def _try_ocr(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """OCR : cherche le texte cible sur l'ecran via docTR."""
+        if not target.text:
+            return None
+
+        try:
+            from core.execution.input_handler import _grounding_ocr
+            bbox = target.original_bbox if target.original_bbox else None
+            result = _grounding_ocr(target.text, anchor_bbox=bbox)
+            if result:
+                print(f"[GroundingPipeline/OCR] '{target.text}' -> ({result['x']}, {result['y']})")
+                return GroundingResult(
+                    x=result['x'],
+                    y=result['y'],
+                    method='ocr',
+                    confidence=result.get('confidence', 0.80),
+                    time_ms=result.get('time_ms', 0),
+                )
+            else:
+                print(f"[GroundingPipeline/OCR] '{target.text}' non trouve")
+        except Exception as e:
+            print(f"[GroundingPipeline/OCR] ERREUR: {e}")
+
+        return None
+
+    def _try_uitars(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """UI-TARS via serveur HTTP — robust, gere les changements de layout."""
+        if not target.text and not target.description:
+            return None
+
+        try:
+            from core.grounding.ui_tars_grounder import UITarsGrounder
+            grounder = UITarsGrounder.get_instance()
+            result = grounder.ground(
+                target_text=target.text,
+                target_description=target.description,
+            )
+            if result:
+                print(f"[GroundingPipeline/UI-TARS] ({result.x}, {result.y}) "
+                      f"conf={result.confidence:.2f} ({result.time_ms:.0f}ms)")
+                return result
+            else:
+                print(f"[GroundingPipeline/UI-TARS] pas de resultat")
+        except Exception as e:
+            print(f"[GroundingPipeline/UI-TARS] ERREUR: {e}")
+
+        return None
+
+    def _try_static(self, target: GroundingTarget) -> Optional[GroundingResult]:
+        """Fallback : coordonnees d'origine du workflow (centre du bounding box)."""
+        bbox = target.original_bbox
+        if not bbox:
+            return None
+
+        w = bbox.get('width', 0)
+        h = bbox.get('height', 0)
+        if not w or not h:
+            return None
+
+        x = int(bbox.get('x', 0) + w / 2)
+        y = int(bbox.get('y', 0) + h / 2)
+
+        print(f"[GroundingPipeline/static] fallback ({x}, {y}) "
+              f"depuis bbox {bbox}")
+
+        return GroundingResult(
+            x=x,
+            y=y,
+            method='static_fallback',
+            confidence=0.30,
+            time_ms=0.0,
+        )

core/grounding/server.py

@@ -0,0 +1,113 @@
+"""Serveur grounding minimaliste — Flask single-thread, même contexte CUDA."""
+import base64, io, json, math, os, re, time, gc
+import torch
+from flask import Flask, request, jsonify
+from PIL import Image
+
+app = Flask(__name__)
+
+MODEL_ID = os.environ.get("GROUNDING_MODEL", "InfiX-ai/InfiGUI-G1-3B")
+MIN_PIXELS = 100 * 28 * 28
+MAX_PIXELS = 5600 * 28 * 28
+_model = None
+_processor = None
+
+def _smart_resize(h, w, factor=28):
+    h_bar = max(factor, round(h/factor)*factor)
+    w_bar = max(factor, round(w/factor)*factor)
+    if h_bar*w_bar > MAX_PIXELS:
+        beta = math.sqrt((h*w)/MAX_PIXELS)
+        h_bar = math.floor(h/beta/factor)*factor
+        w_bar = math.floor(w/beta/factor)*factor
+    elif h_bar*w_bar < MIN_PIXELS:
+        beta = math.sqrt(MIN_PIXELS/(h*w))
+        h_bar = math.ceil(h*beta/factor)*factor
+        w_bar = math.ceil(w*beta/factor)*factor
+    return h_bar, w_bar
+
+def load_model():
+    global _model, _processor
+    if _model is not None:
+        return
+    from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor, BitsAndBytesConfig
+    torch.cuda.empty_cache(); gc.collect()
+    print(f"[grounding] Chargement {MODEL_ID}...")
+    bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
+    _model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+        MODEL_ID, quantization_config=bnb, device_map="auto")
+    _model.eval()
+    _processor = AutoProcessor.from_pretrained(MODEL_ID, min_pixels=MIN_PIXELS, max_pixels=MAX_PIXELS, padding_side="left")
+    print(f"[grounding] Prêt — VRAM: {torch.cuda.memory_allocated()/1e9:.2f}GB")
+
+@app.route('/health')
+def health():
+    return jsonify({"status": "ok", "model": MODEL_ID, "model_loaded": _model is not None,
+                    "cuda_available": torch.cuda.is_available(),
+                    "vram_allocated_gb": round(torch.cuda.memory_allocated()/1e9, 2)})
+
+@app.route('/ground', methods=['POST'])
+def ground():
+    if _model is None:
+        return jsonify({"error": "Modèle pas chargé"}), 503
+    from qwen_vl_utils import process_vision_info
+    data = request.json
+    target = data.get('target_text', '')
+    desc = data.get('target_description', '')
+    label = f"{target} — {desc}" if desc else target
+    if not label.strip():
+        return jsonify({"error": "target_text requis"}), 400
+
+    # Image
+    if data.get('image_b64'):
+        raw = data['image_b64'].split(',')[1] if ',' in data['image_b64'] else data['image_b64']
+        img = Image.open(io.BytesIO(base64.b64decode(raw))).convert('RGB')
+    else:
+        import mss
+        with mss.mss() as sct:
+            grab = sct.grab(sct.monitors[0])
+            img = Image.frombytes('RGB', grab.size, grab.bgra, 'raw', 'BGRX')
+
+    W, H = img.size
+    rH, rW = _smart_resize(H, W)
+
+    user_text = f'The screen\'s resolution is {rW}x{rH}.\nLocate the UI element(s) for "{label}", output the coordinates using JSON format: [{{"point_2d": [x, y]}}, ...]'
+    system = "You FIRST think about the reasoning process as an internal monologue and then provide the final answer.\nThe reasoning process MUST BE enclosed within <think> </think> tags."
+
+    messages = [{"role": "system", "content": system},
+                {"role": "user", "content": [{"type": "image", "image": img}, {"type": "text", "text": user_text}]}]
+
+    text = _processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    image_inputs, video_inputs = process_vision_info(messages)
+    inputs = _processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt").to(_model.device)
+
+    t0 = time.time()
+    with torch.no_grad():
+        gen = _model.generate(**inputs, max_new_tokens=512)
+    infer_ms = (time.time()-t0)*1000
+
+    trimmed = [o[len(i):] for i,o in zip(inputs.input_ids, gen)]
+    raw = _processor.batch_decode(trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0].strip()
+    print(f"[grounding] '{label[:40]}' → {raw[:100]} ({infer_ms:.0f}ms)")
+
+    # Parser JSON point_2d
+    json_part = raw.split("</think>")[-1] if "</think>" in raw else raw
+    json_part = json_part.replace("```json","").replace("```","").strip()
+    px, py = None, None
+    try:
+        parsed = json.loads(json_part)
+        if isinstance(parsed, list) and len(parsed) > 0:
+            pt = parsed[0].get("point_2d", [])
+            if len(pt) >= 2:
+                px, py = int(pt[0]*W/rW), int(pt[1]*H/rH)
+    except json.JSONDecodeError:
+        m = re.search(r'"point_2d"\s*:\s*\[(\d+),\s*(\d+)\]', raw)
+        if m:
+            px, py = int(int(m.group(1))*W/rW), int(int(m.group(2))*H/rH)
+
+    return jsonify({"x": px, "y": py, "method": "infigui", "confidence": 0.90 if px else 0.0,
+                    "time_ms": round(infer_ms, 1), "raw_output": raw[:300]})
+
+if __name__ == '__main__':
+    load_model()
+    app.run(host='0.0.0.0', port=8200, threaded=False)

core/grounding/shadow_learning_hook.py

@@ -0,0 +1,156 @@
+"""
+core/grounding/shadow_learning_hook.py — Hook d'apprentissage Shadow
+
+Connecte le ShadowObserver au SignatureStore : chaque clic observé pendant
+une session Shadow enrichit la base de signatures d'éléments.
+
+L'humain clique quelque part → on détecte quel élément UI est sous le clic →
+on stocke sa signature (texte, type, position, voisins) pour le replay.
+
+Ce module est un HOOK optionnel — il ne modifie pas le ShadowObserver,
+il s'y branche via callback.
+
+Utilisation :
+    from core.grounding.shadow_learning_hook import ShadowLearningHook
+
+    hook = ShadowLearningHook()
+
+    # Dans le ShadowObserver ou l'API de capture :
+    hook.on_click_observed(
+        click_x=542, click_y=318,
+        screenshot_pil=screen,
+        window_title="Bloc-notes",
+        target_label="Bouton Valider",
+    )
+"""
+
+from __future__ import annotations
+
+import threading
+import time
+from typing import Any, Dict, Optional
+
+from core.grounding.element_signature import SignatureStore
+from core.grounding.fast_types import DetectedUIElement
+
+
+class ShadowLearningHook:
+    """Hook d'apprentissage pour le mode Shadow.
+
+    À chaque clic humain observé, détecte l'élément sous le clic
+    et enrichit le SignatureStore.
+    """
+
+    def __init__(self, signature_store: Optional[SignatureStore] = None):
+        self._store = signature_store or SignatureStore()
+        self._detector = None  # Lazy load pour ne pas charger RF-DETR au startup
+        self._lock = threading.Lock()
+
+    def on_click_observed(
+        self,
+        click_x: int,
+        click_y: int,
+        screenshot_pil: Optional[Any] = None,
+        window_title: str = "",
+        target_label: str = "",
+        target_description: str = "",
+    ) -> Optional[Dict[str, Any]]:
+        """Appelé quand un clic humain est observé pendant le Shadow.
+
+        Args:
+            click_x, click_y: Position du clic (pixels écran).
+            screenshot_pil: Capture d'écran PIL au moment du clic.
+            window_title: Titre de la fenêtre active.
+            target_label: Label de l'étape (si connu).
+            target_description: Description de l'élément (si connue).
+
+        Returns:
+            Dict avec la signature créée/enrichie, ou None si échec.
+        """
+        t0 = time.time()
+
+        try:
+            # Lazy load du détecteur
+            if self._detector is None:
+                from core.grounding.fast_detector import FastDetector
+                self._detector = FastDetector()
+
+            # Détecter les éléments sur l'écran
+            snapshot = self._detector.detect(screenshot_pil=screenshot_pil)
+
+            if not snapshot.elements:
+                print(f"📝 [Shadow/learn] Aucun élément détecté à ({click_x}, {click_y})")
+                return None
+
+            # Trouver l'élément sous le clic
+            clicked_element = self._find_element_at(click_x, click_y, snapshot.elements)
+
+            if clicked_element is None:
+                print(f"📝 [Shadow/learn] Aucun élément sous ({click_x}, {click_y})")
+                return None
+
+            # Construire la clé de la cible
+            target_key = SignatureStore.make_target_key(
+                target_label or clicked_element.ocr_text,
+                target_description,
+            )
+            screen_ctx = SignatureStore.make_screen_context(
+                window_title, snapshot.resolution,
+            )
+
+            # Enregistrer la signature
+            self._store.record_success(
+                target_key=target_key,
+                screen_context=screen_ctx,
+                element=clicked_element,
+                confidence=1.0,  # L'humain a cliqué → confiance maximale
+            )
+
+            dt = (time.time() - t0) * 1000
+            print(f"📝 [Shadow/learn] Signature '{clicked_element.ocr_text}' "
+                  f"type={clicked_element.element_type} "
+                  f"pos={clicked_element.relative_position} "
+                  f"voisins={clicked_element.neighbors[:3]} ({dt:.0f}ms)")
+
+            return {
+                "target_key": target_key,
+                "text": clicked_element.ocr_text,
+                "element_type": clicked_element.element_type,
+                "relative_position": clicked_element.relative_position,
+                "neighbors": clicked_element.neighbors,
+                "center": clicked_element.center,
+            }
+
+        except Exception as e:
+            print(f"⚠️ [Shadow/learn] Erreur: {e}")
+            return None
+
+    @staticmethod
+    def _find_element_at(
+        x: int, y: int,
+        elements: list,
+        margin: int = 20,
+    ) -> Optional[DetectedUIElement]:
+        """Trouve l'élément dont la bbox contient le point (x, y).
+
+        Si aucun match exact, prend le plus proche dans un rayon de `margin` pixels.
+        """
+        # Match exact : le clic est dans la bbox
+        for elem in elements:
+            x1, y1, x2, y2 = elem.bbox
+            if x1 <= x <= x2 and y1 <= y <= y2:
+                return elem
+
+        # Match par proximité : le clic est proche du centre
+        best_elem = None
+        best_dist = float('inf')
+
+        for elem in elements:
+            dx = abs(elem.center[0] - x)
+            dy = abs(elem.center[1] - y)
+            dist = (dx**2 + dy**2) ** 0.5
+            if dist < margin and dist < best_dist:
+                best_dist = dist
+                best_elem = elem
+
+        return best_elem

core/grounding/smart_matcher.py

@@ -0,0 +1,263 @@
+"""
+core/grounding/smart_matcher.py — Layer SMART : matching déterministe/probabiliste
+
+Étant donné un ScreenSnapshot (tous les éléments détectés) et un GroundingTarget
+(ce qu'on cherche), trouve l'élément correspondant avec un score de confiance.
+
+Pipeline de matching (court-circuit au premier match haute confiance) :
+  1. Texte exact (2ms)        → score 0.95
+  2. Texte fuzzy ratio (5ms)  → score 0.70-0.90
+  3. Type + position (2ms)    → bonus/malus
+  4. Voisins contextuels (5ms) → bonus
+  5. Score combiné → MatchCandidate
+
+Utilisation :
+    from core.grounding.smart_matcher import SmartMatcher
+    from core.grounding.fast_types import ScreenSnapshot
+    from core.grounding.target import GroundingTarget
+
+    matcher = SmartMatcher()
+    candidate = matcher.match(snapshot, GroundingTarget(text="Valider"))
+    if candidate and candidate.score >= 0.90:
+        print(f"Match direct : ({candidate.element.center}) score={candidate.score}")
+"""
+
+from __future__ import annotations
+
+import re
+from difflib import SequenceMatcher
+from typing import Dict, List, Optional
+
+from core.grounding.fast_types import DetectedUIElement, MatchCandidate, ScreenSnapshot
+from core.grounding.target import GroundingTarget
+
+
+class SmartMatcher:
+    """Matching intelligent entre une cible et les éléments détectés.
+
+    Combine plusieurs signaux (texte, type, position, voisins) en un score
+    de confiance unique pour chaque candidat.
+    """
+
+    def __init__(
+        self,
+        weight_text: float = 0.50,
+        weight_type: float = 0.10,
+        weight_position: float = 0.15,
+        weight_neighbors: float = 0.25,
+    ):
+        self.w_text = weight_text
+        self.w_type = weight_type
+        self.w_position = weight_position
+        self.w_neighbors = weight_neighbors
+
+    def match(
+        self,
+        snapshot: ScreenSnapshot,
+        target: GroundingTarget,
+        signature: Optional[Dict] = None,
+    ) -> Optional[MatchCandidate]:
+        """Trouve le MEILLEUR élément correspondant à la cible.
+
+        Returns:
+            Le MatchCandidate avec le score le plus élevé, ou None si aucun match.
+        """
+        candidates = self.match_all(snapshot, target, signature)
+        if not candidates:
+            return None
+        return candidates[0]
+
+    def match_all(
+        self,
+        snapshot: ScreenSnapshot,
+        target: GroundingTarget,
+        signature: Optional[Dict] = None,
+    ) -> List[MatchCandidate]:
+        """Trouve TOUS les candidats triés par score décroissant.
+
+        Args:
+            snapshot: État de l'écran (éléments détectés + OCR).
+            target: Ce qu'on cherche (texte, description, bbox d'origine).
+            signature: Signature apprise (optionnel, enrichit le matching).
+
+        Returns:
+            Liste de MatchCandidate triée par score décroissant.
+        """
+        if not snapshot.elements:
+            return []
+
+        target_text = (target.text or "").strip()
+        target_desc = (target.description or "").strip()
+        search_text = target_text or target_desc
+
+        if not search_text:
+            return []
+
+        candidates = []
+        search_lower = self._normalize(search_text)
+
+        for elem in snapshot.elements:
+            score_detail: Dict[str, float] = {}
+            method = ""
+
+            # --- 1. Score texte ---
+            text_score = self._score_text(search_lower, elem.ocr_text)
+            score_detail["text"] = text_score
+
+            if text_score >= 0.95:
+                method = "exact_text"
+            elif text_score >= 0.70:
+                method = "fuzzy_text"
+
+            # --- 2. Score type (si signature connue) ---
+            type_score = 0.5  # neutre par défaut
+            if signature and signature.get("element_type"):
+                if elem.element_type == signature["element_type"]:
+                    type_score = 1.0
+                elif elem.element_type == "element":
+                    type_score = 0.5  # non classifié, neutre
+                else:
+                    type_score = 0.2
+            score_detail["type"] = type_score
+
+            # --- 3. Score position (si bbox d'origine connue) ---
+            position_score = 0.5  # neutre
+            if target.original_bbox:
+                position_score = self._score_position(
+                    elem.center, target.original_bbox,
+                    snapshot.resolution[0], snapshot.resolution[1],
+                )
+            elif signature and signature.get("relative_position"):
+                if elem.relative_position == signature["relative_position"]:
+                    position_score = 0.9
+                else:
+                    position_score = 0.3
+            score_detail["position"] = position_score
+
+            # --- 4. Score voisins (si signature connue) ---
+            neighbor_score = 0.5  # neutre
+            if signature and signature.get("neighbors"):
+                neighbor_score = self._score_neighbors(
+                    elem.neighbors, signature["neighbors"]
+                )
+            score_detail["neighbors"] = neighbor_score
+
+            # --- Score combiné ---
+            combined = (
+                self.w_text * text_score
+                + self.w_type * type_score
+                + self.w_position * position_score
+                + self.w_neighbors * neighbor_score
+            )
+
+            # Seuil minimum : pas de candidat si le texte ne matche pas du tout
+            if text_score < 0.30:
+                continue
+
+            if not method:
+                method = "combined"
+
+            candidates.append(MatchCandidate(
+                element=elem,
+                score=combined,
+                score_detail=score_detail,
+                method=method,
+            ))
+
+        # Trier par score décroissant
+        candidates.sort(key=lambda c: c.score, reverse=True)
+
+        return candidates
+
+    # ------------------------------------------------------------------
+    # Scoring texte
+    # ------------------------------------------------------------------
+
+    def _score_text(self, search: str, ocr_text: str) -> float:
+        """Score de similarité textuelle (0-1)."""
+        if not ocr_text:
+            return 0.0
+
+        ocr_lower = self._normalize(ocr_text)
+
+        # Match exact
+        if search == ocr_lower:
+            return 1.0
+
+        # Inclusion (l'un contient l'autre)
+        if search in ocr_lower or ocr_lower in search:
+            overlap = min(len(search), len(ocr_lower))
+            total = max(len(search), len(ocr_lower))
+            if total > 0:
+                return 0.70 + 0.25 * (overlap / total)
+
+        # Fuzzy matching (SequenceMatcher, standard library)
+        ratio = SequenceMatcher(None, search, ocr_lower).ratio()
+        if ratio >= 0.60:
+            return 0.50 + 0.40 * ratio
+
+        return ratio * 0.3
+
+    # ------------------------------------------------------------------
+    # Scoring position
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _score_position(
+        center: tuple,
+        original_bbox: dict,
+        screen_w: int,
+        screen_h: int,
+    ) -> float:
+        """Score de proximité par rapport à la position d'origine (0-1)."""
+        if not original_bbox:
+            return 0.5
+
+        orig_x = original_bbox.get("x", 0) + original_bbox.get("width", 0) / 2
+        orig_y = original_bbox.get("y", 0) + original_bbox.get("height", 0) / 2
+
+        dx = abs(center[0] - orig_x) / max(screen_w, 1)
+        dy = abs(center[1] - orig_y) / max(screen_h, 1)
+        distance_norm = (dx**2 + dy**2) ** 0.5
+
+        # distance 0 = score 1.0, distance 0.5 (demi-écran) = score ~0.2
+        return max(0.0, 1.0 - distance_norm * 2.0)
+
+    # ------------------------------------------------------------------
+    # Scoring voisins
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _score_neighbors(
+        current_neighbors: List[str],
+        expected_neighbors: List[str],
+    ) -> float:
+        """Score Jaccard sur les ensembles de mots voisins (0-1)."""
+        if not expected_neighbors:
+            return 0.5
+
+        current_set = {n.lower().strip() for n in current_neighbors if n}
+        expected_set = {n.lower().strip() for n in expected_neighbors if n}
+
+        if not current_set and not expected_set:
+            return 0.5
+
+        intersection = current_set & expected_set
+        union = current_set | expected_set
+
+        if not union:
+            return 0.5
+
+        return len(intersection) / len(union)
+
+    # ------------------------------------------------------------------
+    # Utilitaires
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def _normalize(text: str) -> str:
+        """Normalise un texte pour la comparaison."""
+        text = text.lower().strip()
+        text = re.sub(r'[_\-\./\\]', ' ', text)
+        text = re.sub(r'\s+', ' ', text)
+        return text

core/grounding/target.py

@@ -0,0 +1,48 @@
+"""
+core/grounding/target.py — Types partagés pour le grounding visuel
+
+Dataclasses décrivant une cible à localiser (GroundingTarget) et
+le résultat d'une localisation (GroundingResult).
+
+Ces types sont la brique commune pour tous les modules de grounding :
+template matching, OCR, VLM, CLIP, etc.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+
+@dataclass
+class GroundingTarget:
+    """Description d'un élément UI à localiser sur l'écran.
+
+    Attributs :
+        text : texte visible de l'élément (bouton, label, etc.)
+        description : description sémantique libre (ex: "le bouton Valider en bas à droite")
+        template_b64 : capture visuelle de l'élément, encodée en base64 PNG/JPEG
+        original_bbox : position d'origine lors de la capture {x, y, width, height}
+    """
+    text: str = ""
+    description: str = ""
+    template_b64: str = ""
+    original_bbox: Optional[Dict[str, int]] = field(default=None)
+
+
+@dataclass
+class GroundingResult:
+    """Résultat d'une localisation d'élément UI.
+
+    Attributs :
+        x : coordonnée X du centre de l'élément trouvé (pixels écran)
+        y : coordonnée Y du centre de l'élément trouvé (pixels écran)
+        method : méthode ayant produit le résultat ('template', 'ocr', 'vlm', 'clip', etc.)
+        confidence : score de confiance [0.0 – 1.0]
+        time_ms : temps de recherche en millisecondes
+    """
+    x: int
+    y: int
+    method: str
+    confidence: float
+    time_ms: float

core/grounding/template_matcher.py

@@ -0,0 +1,350 @@
+"""
+core/grounding/template_matcher.py — Template matching centralisé
+
+Fournit une classe TemplateMatcher qui localise une ancre visuelle (image template)
+dans un screenshot via cv2.matchTemplate. Supporte single-scale et multi-scale.
+
+Remplace les implémentations dupliquées dans :
+  - core/execution/observe_reason_act.py (~1348-1375)
+  - visual_workflow_builder/backend/api_v3/execute.py (~930-963)
+  - visual_workflow_builder/backend/catalog_routes_v2_vlm.py (~339-381)
+  - visual_workflow_builder/backend/services/intelligent_executor.py (~131-210)
+  - core/detection/omniparser_adapter.py (~330)
+
+Utilisation :
+    from core.grounding import TemplateMatcher, MatchResult
+
+    matcher = TemplateMatcher(threshold=0.75)
+    result = matcher.match_screen(anchor_b64="...")
+    if result:
+        print(f"Trouvé à ({result.x}, {result.y}) score={result.score:.3f}")
+"""
+
+from __future__ import annotations
+
+import base64
+import io
+import logging
+import time
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+# Imports optionnels — le module se charge même sans cv2/PIL/mss
+try:
+    import cv2
+    _CV2 = True
+except ImportError:
+    _CV2 = False
+
+try:
+    import numpy as np
+    _NP = True
+except ImportError:
+    _NP = False
+
+try:
+    from PIL import Image
+    _PIL = True
+except ImportError:
+    _PIL = False
+
+try:
+    import mss as mss_lib
+    _MSS = True
+except ImportError:
+    _MSS = False
+
+
+# ---------------------------------------------------------------------------
+# Résultat d'un match
+# ---------------------------------------------------------------------------
+
+@dataclass
+class MatchResult:
+    """Résultat d'un template matching."""
+    x: int
+    y: int
+    score: float
+    method: str           # 'template' | 'template_multiscale'
+    time_ms: float
+    scale: float = 1.0    # Échelle à laquelle le meilleur match a été trouvé
+
+
+# ---------------------------------------------------------------------------
+# TemplateMatcher
+# ---------------------------------------------------------------------------
+
+class TemplateMatcher:
+    """Localise une ancre visuelle dans un screenshot via template matching.
+
+    Paramètres :
+        threshold : score minimum pour accepter un match (défaut 0.75)
+        multiscale : active le matching multi-échelle (défaut False)
+        scales : liste d'échelles à tester en mode multi-scale
+        method : méthode cv2 (défaut cv2.TM_CCOEFF_NORMED)
+        grayscale : convertir en niveaux de gris avant matching (défaut False)
+    """
+
+    # Échelles par défaut pour le mode multi-scale, ordonnées par
+    # probabilité décroissante (1.0 en premier = rapide si ça matche)
+    DEFAULT_SCALES: List[float] = [1.0, 0.95, 1.05, 0.9, 1.1, 0.85, 1.15, 0.8, 1.2]
+
+    def __init__(
+        self,
+        threshold: float = 0.75,
+        multiscale: bool = False,
+        scales: Optional[List[float]] = None,
+        grayscale: bool = False,
+    ):
+        self.threshold = threshold
+        self.multiscale = multiscale
+        self.scales = scales or self.DEFAULT_SCALES
+        self.grayscale = grayscale
+        # cv2.TM_CCOEFF_NORMED est la méthode utilisée partout dans le projet
+        self._cv2_method = cv2.TM_CCOEFF_NORMED if _CV2 else None
+
+    # ------------------------------------------------------------------
+    # API publique
+    # ------------------------------------------------------------------
+
+    def match_screen(
+        self,
+        anchor_b64: Optional[str] = None,
+        anchor_pil: Optional["Image.Image"] = None,
+        screen_pil: Optional["Image.Image"] = None,
+    ) -> Optional[MatchResult]:
+        """Cherche l'ancre dans le screenshot courant (ou fourni).
+
+        L'ancre peut être passée en base64 ou en PIL Image.
+        Le screenshot est capturé via mss si non fourni.
+
+        Retourne un MatchResult ou None si aucun match >= seuil.
+        """
+        if not (_CV2 and _NP and _PIL):
+            logger.debug("[TemplateMatcher] cv2/numpy/PIL non disponible")
+            return None
+
+        # --- Préparer l'ancre ---
+        anchor_img = self._decode_anchor(anchor_b64, anchor_pil)
+        if anchor_img is None:
+            return None
+
+        # --- Préparer le screenshot ---
+        if screen_pil is None:
+            screen_pil = self._capture_screen()
+        if screen_pil is None:
+            return None
+
+        # --- Convertir en arrays cv2 ---
+        screen_cv = cv2.cvtColor(np.array(screen_pil), cv2.COLOR_RGB2BGR)
+        anchor_cv = cv2.cvtColor(np.array(anchor_img), cv2.COLOR_RGB2BGR)
+
+        # --- Matching ---
+        if self.multiscale:
+            return self._match_multiscale(screen_cv, anchor_cv)
+        else:
+            return self._match_single(screen_cv, anchor_cv)
+
+    def match_in_region(
+        self,
+        region_cv: "np.ndarray",
+        anchor_cv: "np.ndarray",
+        threshold: Optional[float] = None,
+    ) -> Optional[MatchResult]:
+        """Match dans une région déjà découpée (arrays BGR).
+
+        Utilisé par les pipelines qui font leur propre capture/découpe.
+        """
+        if not (_CV2 and _NP):
+            return None
+
+        thr = threshold if threshold is not None else self.threshold
+
+        if self.multiscale:
+            return self._match_multiscale(region_cv, anchor_cv, threshold_override=thr)
+        else:
+            return self._match_single(region_cv, anchor_cv, threshold_override=thr)
+
+    def match_screen_diagnostic(
+        self,
+        anchor_b64: Optional[str] = None,
+        anchor_pil: Optional["Image.Image"] = None,
+        screen_pil: Optional["Image.Image"] = None,
+    ) -> str:
+        """Retourne un diagnostic textuel (score + position) même sans match."""
+        if not (_CV2 and _NP and _PIL):
+            return "cv2/numpy/PIL non dispo"
+
+        anchor_img = self._decode_anchor(anchor_b64, anchor_pil)
+        if anchor_img is None:
+            return "ancre non décodable"
+
+        if screen_pil is None:
+            screen_pil = self._capture_screen()
+        if screen_pil is None:
+            return "capture écran échouée"
+
+        screen_cv = cv2.cvtColor(np.array(screen_pil), cv2.COLOR_RGB2BGR)
+        anchor_cv = cv2.cvtColor(np.array(anchor_img), cv2.COLOR_RGB2BGR)
+
+        if anchor_cv.shape[0] >= screen_cv.shape[0] or anchor_cv.shape[1] >= screen_cv.shape[1]:
+            return f"ancre {anchor_cv.shape[:2]} >= écran {screen_cv.shape[:2]}"
+
+        s_img, a_img = self._maybe_grayscale(screen_cv, anchor_cv)
+        result_tm = cv2.matchTemplate(s_img, a_img, self._cv2_method)
+        _, max_val, _, max_loc = cv2.minMaxLoc(result_tm)
+        return f"{max_val:.3f} pos={max_loc}"
+
+    # ------------------------------------------------------------------
+    # Méthodes internes
+    # ------------------------------------------------------------------
+
+    def _match_single(
+        self,
+        screen_cv: "np.ndarray",
+        anchor_cv: "np.ndarray",
+        threshold_override: Optional[float] = None,
+    ) -> Optional[MatchResult]:
+        """Template matching single-scale."""
+        threshold = threshold_override if threshold_override is not None else self.threshold
+
+        if anchor_cv.shape[0] >= screen_cv.shape[0] or anchor_cv.shape[1] >= screen_cv.shape[1]:
+            logger.debug("[TemplateMatcher] Ancre plus grande que le screen")
+            return None
+
+        s_img, a_img = self._maybe_grayscale(screen_cv, anchor_cv)
+
+        t0 = time.time()
+        result_tm = cv2.matchTemplate(s_img, a_img, self._cv2_method)
+        _, max_val, _, max_loc = cv2.minMaxLoc(result_tm)
+        elapsed_ms = (time.time() - t0) * 1000
+
+        logger.debug(
+            "[TemplateMatcher] score=%.3f pos=%s (%.0fms)",
+            max_val, max_loc, elapsed_ms,
+        )
+
+        if max_val >= threshold:
+            cx = max_loc[0] + anchor_cv.shape[1] // 2
+            cy = max_loc[1] + anchor_cv.shape[0] // 2
+            return MatchResult(
+                x=cx,
+                y=cy,
+                score=float(max_val),
+                method='template',
+                time_ms=elapsed_ms,
+                scale=1.0,
+            )
+        return None
+
+    def _match_multiscale(
+        self,
+        screen_cv: "np.ndarray",
+        anchor_cv: "np.ndarray",
+        threshold_override: Optional[float] = None,
+    ) -> Optional[MatchResult]:
+        """Template matching multi-scale."""
+        threshold = threshold_override if threshold_override is not None else self.threshold
+
+        best_score = -1.0
+        best_loc = None
+        best_scale = 1.0
+        best_anchor_shape = anchor_cv.shape
+
+        t0 = time.time()
+
+        for scale in self.scales:
+            if scale == 1.0:
+                scaled = anchor_cv
+            else:
+                new_w = int(anchor_cv.shape[1] * scale)
+                new_h = int(anchor_cv.shape[0] * scale)
+                if new_w < 8 or new_h < 8:
+                    continue
+                if new_h >= screen_cv.shape[0] or new_w >= screen_cv.shape[1]:
+                    continue
+                scaled = cv2.resize(anchor_cv, (new_w, new_h), interpolation=cv2.INTER_AREA)
+
+            if scaled.shape[0] >= screen_cv.shape[0] or scaled.shape[1] >= screen_cv.shape[1]:
+                continue
+
+            s_img, a_img = self._maybe_grayscale(screen_cv, scaled)
+            result_tm = cv2.matchTemplate(s_img, a_img, self._cv2_method)
+            _, max_val, _, max_loc = cv2.minMaxLoc(result_tm)
+
+            if max_val > best_score:
+                best_score = max_val
+                best_loc = max_loc
+                best_scale = scale
+                best_anchor_shape = scaled.shape
+
+        elapsed_ms = (time.time() - t0) * 1000
+
+        logger.debug(
+            "[TemplateMatcher/multiscale] best_score=%.3f scale=%.2f (%.0fms)",
+            best_score, best_scale, elapsed_ms,
+        )
+
+        if best_score >= threshold and best_loc is not None:
+            cx = best_loc[0] + best_anchor_shape[1] // 2
+            cy = best_loc[1] + best_anchor_shape[0] // 2
+            return MatchResult(
+                x=cx,
+                y=cy,
+                score=float(best_score),
+                method='template_multiscale',
+                time_ms=elapsed_ms,
+                scale=best_scale,
+            )
+        return None
+
+    def _maybe_grayscale(
+        self,
+        screen: "np.ndarray",
+        anchor: "np.ndarray",
+    ) -> Tuple["np.ndarray", "np.ndarray"]:
+        """Convertit en niveaux de gris si self.grayscale est True."""
+        if not self.grayscale:
+            return screen, anchor
+        s = cv2.cvtColor(screen, cv2.COLOR_BGR2GRAY) if len(screen.shape) == 3 else screen
+        a = cv2.cvtColor(anchor, cv2.COLOR_BGR2GRAY) if len(anchor.shape) == 3 else anchor
+        return s, a
+
+    @staticmethod
+    def _decode_anchor(
+        anchor_b64: Optional[str],
+        anchor_pil: Optional["Image.Image"],
+    ) -> Optional["Image.Image"]:
+        """Décode l'ancre depuis base64 ou retourne le PIL directement."""
+        if anchor_pil is not None:
+            return anchor_pil
+
+        if anchor_b64 is None:
+            logger.debug("[TemplateMatcher] Ni anchor_b64 ni anchor_pil fourni")
+            return None
+
+        try:
+            raw = anchor_b64.split(',')[1] if ',' in anchor_b64 else anchor_b64
+            data = base64.b64decode(raw)
+            return Image.open(io.BytesIO(data))
+        except Exception as e:
+            logger.debug("[TemplateMatcher] Erreur décodage ancre: %s", e)
+            return None
+
+    @staticmethod
+    def _capture_screen() -> Optional["Image.Image"]:
+        """Capture l'écran complet via mss (moniteur 0 = tous les écrans)."""
+        if not _MSS:
+            logger.debug("[TemplateMatcher] mss non disponible")
+            return None
+
+        try:
+            with mss_lib.mss() as sct:
+                mon = sct.monitors[0]
+                grab = sct.grab(mon)
+                return Image.frombytes('RGB', grab.size, grab.bgra, 'raw', 'BGRX')
+        except Exception as e:
+            logger.debug("[TemplateMatcher] Erreur capture écran: %s", e)
+            return None

core/grounding/think_arbiter.py

@@ -0,0 +1,103 @@
+"""
+core/grounding/think_arbiter.py — Layer THINK : VLM arbitre (InfiGUI via subprocess)
+
+Appelé UNIQUEMENT quand le SmartMatcher n'a pas assez confiance.
+Utilise le subprocess worker InfiGUI (pas de serveur HTTP).
+
+Utilisation :
+    from core.grounding.think_arbiter import ThinkArbiter
+
+    arbiter = ThinkArbiter()
+    result = arbiter.arbitrate(target, candidates, screenshot)
+"""
+
+from __future__ import annotations
+
+import time
+from typing import Any, Dict, List, Optional
+
+from core.grounding.fast_types import LocateResult, MatchCandidate
+from core.grounding.target import GroundingTarget
+
+
+class ThinkArbiter:
+    """Arbitre VLM — appelle InfiGUI via subprocess worker."""
+
+    def __init__(self):
+        self._grounder = None
+
+    def _get_grounder(self):
+        if self._grounder is None:
+            from core.grounding.ui_tars_grounder import UITarsGrounder
+            self._grounder = UITarsGrounder.get_instance()
+        return self._grounder
+
+    @property
+    def available(self) -> bool:
+        """Toujours disponible — le worker se lance à la demande."""
+        return True
+
+    def arbitrate(
+        self,
+        target: GroundingTarget,
+        candidates: List[MatchCandidate],
+        screenshot_pil: Optional[Any] = None,
+    ) -> Optional[LocateResult]:
+        """Demande au VLM de trancher.
+
+        Si target.template_b64 est fourni, on bascule en mode fusionné :
+        le crop est passé comme image de référence à InfiGUI, ce qui évite
+        une description Ollama qwen2.5vl coûteuse en VRAM.
+        """
+        t0 = time.time()
+
+        # Décodage du crop d'ancre si disponible (mode fusionné)
+        anchor_pil = None
+        if target.template_b64:
+            try:
+                import base64
+                import io
+                from PIL import Image
+
+                raw_b64 = target.template_b64
+                if ',' in raw_b64:
+                    raw_b64 = raw_b64.split(',', 1)[1]
+                anchor_pil = Image.open(io.BytesIO(base64.b64decode(raw_b64))).convert("RGB")
+            except Exception as ex:
+                print(f"⚠️ [THINK] Décodage anchor échoué: {ex}")
+                anchor_pil = None
+
+        try:
+            grounder = self._get_grounder()
+            result = grounder.ground(
+                target_text=target.text or "",
+                target_description=target.description or "",
+                screen_pil=screenshot_pil,
+                anchor_pil=anchor_pil,
+            )
+
+            dt = (time.time() - t0) * 1000
+
+            if result is None:
+                label = target.text or "<crop>"
+                print(f"🤔 [THINK] VLM n'a pas trouvé '{label}' ({dt:.0f}ms)")
+                return None
+
+            method = "think_vlm_fused" if anchor_pil is not None else "think_vlm"
+            locate = LocateResult(
+                x=result.x,
+                y=result.y,
+                confidence=result.confidence,
+                method=method,
+                time_ms=dt,
+                tier="think",
+                candidates_count=len(candidates),
+            )
+
+            print(f"🤔 [THINK/{method}] ({result.x}, {result.y}) conf={result.confidence:.2f} ({dt:.0f}ms)")
+            return locate
+
+        except Exception as ex:
+            dt = (time.time() - t0) * 1000
+            print(f"⚠️ [THINK] Erreur: {ex} ({dt:.0f}ms)")
+            return None

core/grounding/title_verifier.py

@@ -0,0 +1,174 @@
+"""
+core/grounding/title_verifier.py — Vérification post-action par titre de fenêtre
+
+Après chaque action (clic, double-clic), vérifie que la fenêtre active
+a changé de manière attendue en lisant le titre via OCR sur un crop
+de 45px en haut de l'écran.
+
+Léger (~120ms), non-bloquant (échec = warning + retry, pas stop).
+
+Utilisation :
+    from core.grounding.title_verifier import TitleVerifier
+
+    verifier = TitleVerifier()
+    title = verifier.read_title(screenshot_pil)
+    changed = verifier.has_title_changed(title_before, title_after)
+"""
+
+from __future__ import annotations
+
+import time
+from difflib import SequenceMatcher
+from typing import Optional
+
+
+class TitleVerifier:
+    """Vérifie le titre de la fenêtre active via OCR sur crop."""
+
+    # Hauteur du crop pour la barre de titre Windows
+    TITLE_BAR_HEIGHT = 45
+
+    def __init__(self):
+        self._ocr_fn = None  # Lazy load
+
+    def read_title(self, screenshot_pil) -> str:
+        """Lit le titre de la fenêtre active via OCR sur le crop supérieur.
+
+        Args:
+            screenshot_pil: Image PIL du screenshot complet.
+
+        Returns:
+            Texte du titre (peut être vide si OCR échoue).
+        """
+        t0 = time.time()
+
+        try:
+            w, h = screenshot_pil.size
+            # Crop la barre de titre (45px du haut)
+            title_crop = screenshot_pil.crop((0, 0, w, min(self.TITLE_BAR_HEIGHT, h)))
+
+            # OCR sur le petit crop
+            ocr_fn = self._get_ocr()
+            if ocr_fn is None:
+                return ""
+
+            text = ocr_fn(title_crop)
+            dt = (time.time() - t0) * 1000
+
+            # Nettoyer le texte
+            title = text.strip() if text else ""
+            if title:
+                print(f"📋 [TitleVerify] Titre lu: '{title[:60]}' ({dt:.0f}ms)")
+
+            return title
+
+        except Exception as e:
+            print(f"⚠️ [TitleVerify] Erreur lecture titre: {e}")
+            return ""
+
+    def has_title_changed(self, title_before: str, title_after: str) -> bool:
+        """Vérifie si le titre a changé de manière significative."""
+        if not title_before and not title_after:
+            return False
+        if not title_before or not title_after:
+            return True  # Un des deux est vide = changement
+
+        # Comparaison fuzzy — les titres peuvent avoir des variations mineures
+        ratio = SequenceMatcher(None, title_before.lower(), title_after.lower()).ratio()
+        return ratio < 0.85  # Changement si < 85% similaire
+
+    def verify_action(
+        self,
+        screenshot_before,
+        screenshot_after,
+        action_type: str,
+    ) -> dict:
+        """Vérifie qu'une action a produit l'effet attendu sur le titre.
+
+        Args:
+            screenshot_before: Screenshot PIL avant l'action.
+            screenshot_after: Screenshot PIL après l'action.
+            action_type: Type d'action ("double_click", "click", "type", "hotkey").
+
+        Returns:
+            Dict avec success, title_before, title_after, changed.
+        """
+        # Les actions qui ne changent pas le titre
+        if action_type in ('type_text', 'keyboard_shortcut', 'wait_for_anchor', 'hover'):
+            return {
+                'success': True,
+                'title_before': '',
+                'title_after': '',
+                'changed': False,
+                'reason': f"Action '{action_type}' — vérification titre non requise",
+            }
+
+        title_before = self.read_title(screenshot_before)
+        title_after = self.read_title(screenshot_after)
+        changed = self.has_title_changed(title_before, title_after)
+
+        # Pour un double-clic (ouverture fichier/dossier), le titre DOIT changer
+        # Mais seulement si les titres lus sont significatifs (> 3 chars)
+        # docTR sur un crop 45px dans une VM peut donner du bruit ('o', 'a', etc.)
+        if action_type in ('double_click_anchor',) and not changed:
+            if len(title_before) > 3 and len(title_after) > 3:
+                return {
+                    'success': False,
+                    'title_before': title_before,
+                    'title_after': title_after,
+                    'changed': False,
+                    'reason': f"Double-clic sans changement de titre ('{title_after[:40]}')",
+                }
+            # Titres trop courts = bruit OCR, on ne peut pas conclure
+            return {
+                'success': True,
+                'title_before': title_before,
+                'title_after': title_after,
+                'changed': False,
+                'reason': f"Titre trop court pour vérifier ('{title_after}')",
+            }
+
+        # Pour un clic simple, le changement est optionnel
+        return {
+            'success': True,
+            'title_before': title_before,
+            'title_after': title_after,
+            'changed': changed,
+            'reason': 'Titre changé' if changed else 'Titre identique (acceptable)',
+        }
+
+    _easyocr_reader = None  # Singleton partagé
+
+    def _get_ocr(self):
+        """Lazy load de la fonction OCR (EasyOCR prioritaire, fallback docTR)."""
+        if self._ocr_fn is not None:
+            return self._ocr_fn
+
+        # EasyOCR (rapide, bonne qualité GUI)
+        try:
+            import easyocr
+            import numpy as np
+
+            if TitleVerifier._easyocr_reader is None:
+                TitleVerifier._easyocr_reader = easyocr.Reader(
+                    ['fr', 'en'], gpu=True, verbose=False
+                )
+
+            def _easyocr_extract_text(img):
+                results = TitleVerifier._easyocr_reader.readtext(np.array(img))
+                return ' '.join(r[1] for r in results if r[1].strip())
+
+            self._ocr_fn = _easyocr_extract_text
+            return self._ocr_fn
+        except ImportError:
+            pass
+
+        # Fallback docTR
+        try:
+            import sys
+            sys.path.insert(0, 'visual_workflow_builder/backend')
+            from services.ocr_service import ocr_extract_text
+            self._ocr_fn = ocr_extract_text
+            return self._ocr_fn
+        except ImportError:
+            return None

core/grounding/ui_tars_grounder.py

@@ -0,0 +1,161 @@
+"""
+core/grounding/ui_tars_grounder.py — Grounding via script one-shot InfiGUI
+
+Chaque appel lance un subprocess Python qui charge le modèle, infère, et quitte.
+Lent (~15s) mais fiable — pas de crash CUDA en process persistant.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+import sys
+import threading
+import time
+from typing import Optional
+
+from core.grounding.target import GroundingResult
+
+_instance: Optional[UITarsGrounder] = None
+_instance_lock = threading.Lock()
+
+
+class UITarsGrounder:
+    """Grounding via script one-shot InfiGUI."""
+
+    def __init__(self):
+        self._lock = threading.Lock()
+        self._project_root = os.path.abspath(
+            os.path.join(os.path.dirname(__file__), "..", "..")
+        )
+
+    @classmethod
+    def get_instance(cls) -> UITarsGrounder:
+        global _instance
+        if _instance is None:
+            with _instance_lock:
+                if _instance is None:
+                    _instance = cls()
+        return _instance
+
+    @property
+    def available(self) -> bool:
+        return True  # Toujours disponible — le script se lance à la demande
+
+    def ground(
+        self,
+        target_text: str = "",
+        target_description: str = "",
+        screen_pil=None,
+        anchor_pil=None,
+    ) -> Optional[GroundingResult]:
+        """Localise un élément UI via un script one-shot InfiGUI.
+
+        Args:
+            target_text: nom textuel de la cible (peut être vide si anchor_pil fourni).
+            target_description: description sémantique libre.
+            screen_pil: screenshot complet (PIL.Image).
+            anchor_pil: crop visuel de l'ancre capturée précédemment (PIL.Image).
+                Si fourni, le worker passe en mode fusionné : Image1=crop, Image2=screen,
+                "trouve sur l'image 2 l'élément visuel de l'image 1".
+        """
+        t0 = time.time()
+
+        try:
+            with self._lock:
+                # Sauver l'image principale
+                image_path = "/tmp/infigui_screen.png"
+                if screen_pil is not None:
+                    screen_pil.save(image_path)
+
+                # Sauver l'image d'ancre (mode fusionné)
+                anchor_image_path = ""
+                if anchor_pil is not None:
+                    anchor_image_path = "/tmp/infigui_anchor.png"
+                    anchor_pil.save(anchor_image_path)
+
+                # Construire la requête JSON
+                req = json.dumps({
+                    "target": target_text,
+                    "description": target_description,
+                    "image_path": image_path,
+                    "anchor_image_path": anchor_image_path,
+                })
+
+                mode_str = "fused" if anchor_pil is not None else "text"
+                label_short = target_text[:30] if target_text else "<crop only>"
+                print(f"🎯 [InfiGUI] Lancement one-shot [{mode_str}]: '{label_short}'")
+
+                # Lancer le script one-shot
+                # IMPORTANT: depuis un service systemd où le parent a déjà chargé CUDA,
+                # le subprocess hérite d'un état GPU cassé (No CUDA GPUs available).
+                # Solutions : start_new_session=True (nouveau cgroup) + forcer
+                # CUDA_VISIBLE_DEVICES=0 explicitement pour bypass l'héritage parent.
+                _child_env = {**os.environ}
+                _child_env["PYTHONDONTWRITEBYTECODE"] = "1"
+                _child_env["CUDA_VISIBLE_DEVICES"] = "0"
+                _child_env["NVIDIA_VISIBLE_DEVICES"] = "all"
+                # Supprimer les variables Python qui pourraient pointer sur l'état parent
+                _child_env.pop("PYTORCH_NVML_BASED_CUDA_CHECK", None)
+
+                result = subprocess.run(
+                    [sys.executable, "-m", "core.grounding.infigui_worker"],
+                    input=req + "\n",
+                    capture_output=True,
+                    text=True,
+                    timeout=60,
+                    cwd=self._project_root,
+                    env=_child_env,
+                    start_new_session=True,  # nouveau session group, isole du parent
+                    close_fds=True,
+                )
+
+                if result.returncode != 0:
+                    stderr_lines = (result.stderr or '').strip().split('\n')
+                    # Afficher les dernières lignes significatives du stderr
+                    last_err = [l for l in stderr_lines[-5:] if l.strip()]
+                    print(f"⚠️ [InfiGUI] Script échoué (code {result.returncode})")
+                    for l in last_err:
+                        print(f"   ❌ {l}")
+                    return None
+
+                # Parser la sortie — chercher la ligne JSON de résultat
+                data = None
+                for line in result.stdout.strip().split("\n"):
+                    line = line.strip()
+                    if not line:
+                        continue
+                    try:
+                        parsed = json.loads(line)
+                        if "x" in parsed:
+                            data = parsed
+                    except json.JSONDecodeError:
+                        continue
+
+                if data is None:
+                    print(f"⚠️ [InfiGUI] Pas de réponse JSON dans la sortie")
+                    return None
+
+            dt = (time.time() - t0) * 1000
+
+            if data.get("x") is not None:
+                method_name = "infigui_fused" if anchor_pil is not None else "infigui"
+                print(f"🎯 [InfiGUI/{method_name}] ({data['x']}, {data['y']}) "
+                      f"conf={data.get('confidence', 0):.2f} ({dt:.0f}ms)")
+                return GroundingResult(
+                    x=data["x"], y=data["y"],
+                    method=method_name,
+                    confidence=data.get("confidence", 0.90),
+                    time_ms=dt,
+                )
+            else:
+                print(f"⚠️ [InfiGUI] Pas trouvé ({dt:.0f}ms)")
+                return None
+
+        except subprocess.TimeoutExpired:
+            print(f"⚠️ [InfiGUI] Timeout 60s")
+            return None
+        except Exception as e:
+            print(f"⚠️ [InfiGUI] Erreur: {e}")
+            return None

core/knowledge/ui_patterns.py

@@ -101,6 +101,35 @@ BUILTIN_PATTERNS: List[Dict[str, Any]] = [
         "typical_bbox": [0.35, 0.60, 0.45, 0.68],
         "os": "any",
     },
+    {
+        "name": "dialog_overwrite",
+        "category": "dialog",
+        "triggers": [
+            "voulez-vous remplacer", "voulez-vous écraser",
+            "remplacer le fichier", "replace existing",
+            "fichier existe déjà", "already exists",
+            "overwrite", "écraser",
+        ],
+        "action": "click",
+        "target": "Oui",
+        "alternatives": ["Yes", "Remplacer", "Replace", "Confirmer"],
+        "typical_zone": "dialog_center",
+        "os": "any",
+    },
+    {
+        "name": "dialog_dont_save",
+        "category": "dialog",
+        "triggers": [
+            "ne pas enregistrer", "don't save",
+            "ne pas sauvegarder", "quitter sans enregistrer",
+            "discard changes",
+        ],
+        "action": "click",
+        "target": "Ne pas enregistrer",
+        "alternatives": ["Don't Save", "Ne pas sauvegarder", "Non"],
+        "typical_zone": "dialog_center",
+        "os": "any",
+    },
 
     # === NAVIGATION FENÊTRE ===
     {

core/llm/__init__.py

@@ -0,0 +1,15 @@
+"""Modules LLM (clients Ollama et décisionnels métier) + extracteur OCR."""
+
+from .t2a_decision import (
+    PROMPT_TEMPLATE,
+    DEFAULT_MODEL,
+    analyze_dpi,
+)
+from .ocr_extractor import extract_text_from_image
+
+__all__ = [
+    "PROMPT_TEMPLATE",
+    "DEFAULT_MODEL",
+    "analyze_dpi",
+    "extract_text_from_image",
+]

core/llm/ocr_extractor.py

@@ -0,0 +1,71 @@
+"""Extracteur OCR — texte depuis une image (screenshot d'écran).
+
+Utilise EasyOCR fr+en. Singleton (chargement modèle ~3s au premier appel).
+
+Conçu pour le pipeline streaming serveur (action `extract_text`) : récupère
+un screenshot fresh (dernier heartbeat ou capture forcée), applique l'OCR,
+retourne le texte concaténé pour analyse downstream (ex: t2a_decision).
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+_easyocr_reader = None
+
+
+def _get_reader():
+    """Initialise EasyOCR fr+en au premier appel (singleton)."""
+    global _easyocr_reader
+    if _easyocr_reader is None:
+        import easyocr
+        try:
+            _easyocr_reader = easyocr.Reader(['fr', 'en'], gpu=True, verbose=False)
+            logger.info("EasyOCR initialisé (fr+en, GPU)")
+        except Exception as e:
+            logger.warning("EasyOCR GPU indisponible (%s), fallback CPU", e)
+            _easyocr_reader = easyocr.Reader(['fr', 'en'], gpu=False, verbose=False)
+    return _easyocr_reader
+
+
+def extract_text_from_image(
+    image_path: str,
+    region: Optional[Tuple[int, int, int, int]] = None,
+    paragraph: bool = True,
+) -> str:
+    """Extrait le texte d'une image via EasyOCR.
+
+    Args:
+        image_path: chemin du PNG sur disque.
+        region: (x, y, w, h) pour cropper avant OCR. None = image entière.
+        paragraph: True pour regrouper les lignes en paragraphes (lisible),
+                   False pour blocs séparés (granulaire).
+
+    Returns:
+        Texte concaténé. Chaque ligne / paragraphe est séparé par un saut de ligne.
+        En cas d'erreur, retourne une chaîne vide et log un warning.
+    """
+    path = Path(image_path)
+    if not path.exists():
+        logger.warning("extract_text: fichier introuvable %s", image_path)
+        return ""
+
+    try:
+        from PIL import Image
+        import numpy as np
+
+        img = Image.open(path)
+        if region:
+            x, y, w, h = region
+            img = img.crop((x, y, x + w, y + h))
+
+        reader = _get_reader()
+        results = reader.readtext(np.array(img), detail=0, paragraph=paragraph)
+        return "\n".join(str(r).strip() for r in results if r)
+    except Exception as e:
+        logger.warning("extract_text échoué sur %s : %s", image_path, e)
+        return ""

core/llm/t2a_decision.py

@@ -0,0 +1,168 @@
+"""Aide à la décision de facturation urgences T2A/PMSI via LLM local.
+
+Décide si un passage aux urgences relève :
+- du FORFAIT_URGENCE (passage simple, retour à domicile)
+- de la REQUALIFICATION_HOSPITALISATION (séjour MCO, valorisation 1k-5k€+)
+
+Le prompt impose une extraction littérale des faits du DPI (pas d'invention)
+et une modulation honnête de la confiance. Validé sur 15 DPI synthétiques :
+qwen2.5:7b atteint 100 % d'accuracy en ~5 s/cas avec 4,7 Go VRAM.
+
+Voir docs/clients/ght_sud_95/ et demo/facturation_urgences/RESULTATS.md pour le
+bench comparatif des 11 LLMs évalués.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import time
+import urllib.error
+import urllib.request
+from typing import Any, Dict
+
+logger = logging.getLogger(__name__)
+
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434/api/generate")
+DEFAULT_MODEL = os.environ.get("T2A_MODEL", "qwen2.5:7b")
+DEFAULT_TIMEOUT = 60  # secondes
+
+PROMPT_TEMPLATE = """Tu es médecin DIM (Département d'Information Médicale), expert en facturation T2A/PMSI aux urgences hospitalières en France.
+
+Analyse le dossier patient ci-dessous pour déterminer si le passage relève :
+- FORFAIT_URGENCE : passage simple, retour à domicile, sans surveillance prolongée ni soins continus
+- REQUALIFICATION_HOSPITALISATION : séjour MCO requis selon les 3 critères PMSI/ATIH
+
+LES 3 CRITÈRES UHCD (au moins 2 sur 3 validés ⇒ REQUALIFICATION) :
+1. Pathologie potentiellement évolutive (instabilité hémodynamique, terrain à risque, traitement nécessitant adaptation)
+2. Surveillance médicale et paramédicale prolongée (constantes itératives, observations IDE/médecin, durée > 6 h)
+3. Examens complémentaires ou actes thérapeutiques (biologie, imagerie, sutures, gestes techniques)
+
+INSTRUCTIONS STRICTES :
+1. N'utilise QUE des éléments littéralement présents dans le dossier patient. N'invente AUCUN critère.
+2. Pour CHAQUE critère (1, 2, 3), tu DOIS produire un texte de preuve qui contient AU MOINS UNE CITATION LITTÉRALE du dossier entre guillemets français « ... ». Exemple : « FC à 110 bpm, TA 92/60 ».
+3. Si le critère est NON validé, ne renvoie JAMAIS un fallback creux : explique factuellement ce qui manque, en citant le dossier (ex: « Sortie à H+2 », « Aucun acte technique au compte-rendu »).
+4. Le texte de chaque preuve fait 2-3 phrases : (i) la citation littérale, (ii) l'analyse PMSI, (iii) la conclusion validé/non validé.
+5. Calcule la durée totale du passage en heures (admission → sortie/transfert) à partir des horaires du dossier.
+6. Module ta confiance honnêtement :
+   - "elevee" uniquement si tous les indices convergent
+   - "moyenne" si éléments ambivalents
+   - "faible" si information manquante ou très atypique
+
+Réponds STRICTEMENT en JSON valide, sans texte avant ni après :
+{{
+  "duree_passage_heures": <nombre>,
+  "elements_pour_hospitalisation": [<phrases littéralement extraites du dossier>],
+  "elements_pour_forfait": [<phrases littéralement extraites du dossier>],
+  "decision": "FORFAIT_URGENCE" | "REQUALIFICATION_HOSPITALISATION",
+  "decision_court": "UHCD" | "Forfait Urgences",
+  "preuve_critere1": "<2-3 phrases incluant AU MOINS UNE citation littérale entre « » (motif, symptôme, terrain à risque, traitement). Si non validé : factualise ce qui manque en citant le dossier.>",
+  "critere1_valide": true | false,
+  "preuve_critere2": "<2-3 phrases incluant AU MOINS UNE citation littérale entre « » (constantes, observations IDE, durée surveillance). Si non validé : factualise.>",
+  "critere2_valide": true | false,
+  "preuve_critere3": "<2-3 phrases incluant AU MOINS UNE citation littérale entre « » (actes/examens : biologie, imagerie, suture, etc.). Si non validé : factualise.>",
+  "critere3_valide": true | false,
+  "justification": "<2-3 phrases synthétiques s'appuyant explicitement sur les preuves ci-dessus, avec au moins une citation>",
+  "confiance": "elevee" | "moyenne" | "faible"
+}}
+
+DOSSIER PATIENT :
+{dpi}
+"""
+
+
+def analyze_dpi(
+    dpi_text: str,
+    model: str = DEFAULT_MODEL,
+    timeout: int = DEFAULT_TIMEOUT,
+    ollama_url: str = OLLAMA_URL,
+) -> Dict[str, Any]:
+    """Soumet un DPI urgences à un LLM Ollama et retourne la décision JSON.
+
+    Args:
+        dpi_text: Texte du dossier patient (concaténation des onglets ou DPI brut).
+        model: Modèle Ollama à utiliser (default qwen2.5:7b — 100% accuracy bench).
+        timeout: Timeout HTTP en secondes.
+        ollama_url: Endpoint Ollama (default localhost:11434/api/generate).
+
+    Returns:
+        Dict avec :
+            decision: "FORFAIT_URGENCE" | "REQUALIFICATION_HOSPITALISATION"
+            elements_pour_hospitalisation: List[str]
+            elements_pour_forfait: List[str]
+            duree_passage_heures: float
+            justification: str
+            confiance: "elevee" | "moyenne" | "faible"
+            _elapsed_s: float (latence)
+            _model: str
+        En cas d'erreur :
+            {"_error": str, "_elapsed_s": float} (réseau / Ollama indisponible)
+            {"_parse_error": True, "_raw": str, "_elapsed_s": float} (JSON invalide)
+    """
+    payload = {
+        "model": model,
+        "prompt": PROMPT_TEMPLATE.format(dpi=dpi_text),
+        "stream": False,
+        "format": "json",
+        "keep_alive": "5m",
+        "options": {
+            "temperature": 0.1,
+            "num_predict": 1500,
+            "num_ctx": 16384,
+        },
+    }
+    data = json.dumps(payload).encode("utf-8")
+    req = urllib.request.Request(
+        ollama_url,
+        data=data,
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    t0 = time.time()
+    try:
+        with urllib.request.urlopen(req, timeout=timeout) as resp:
+            body = json.loads(resp.read().decode("utf-8"))
+    except (urllib.error.URLError, TimeoutError, ConnectionError) as e:
+        elapsed = round(time.time() - t0, 1)
+        logger.warning("analyze_dpi: Ollama indisponible (%s) après %.1fs", e, elapsed)
+        return {"_error": str(e), "_elapsed_s": elapsed, "_model": model}
+
+    elapsed = time.time() - t0
+
+    raw_response = body.get("response", "").strip()
+    raw_thinking = body.get("thinking", "").strip()
+
+    candidates = [raw_response]
+    if not raw_response and raw_thinking:
+        last_close = raw_thinking.rfind("}")
+        last_open = raw_thinking.rfind("{", 0, last_close)
+        if last_open != -1 and last_close != -1:
+            candidates.append(raw_thinking[last_open:last_close + 1])
+
+    parsed = None
+    for cand in candidates:
+        cleaned = cand
+        if cleaned.startswith("```"):
+            cleaned = cleaned.split("\n", 1)[-1]
+            if cleaned.endswith("```"):
+                cleaned = cleaned.rsplit("```", 1)[0]
+            cleaned = cleaned.strip()
+        try:
+            parsed = json.loads(cleaned)
+            break
+        except json.JSONDecodeError:
+            continue
+
+    if parsed is None:
+        return {
+            "_parse_error": True,
+            "_raw": (raw_response or raw_thinking)[:500],
+            "_elapsed_s": round(elapsed, 1),
+            "_model": model,
+        }
+
+    parsed["_elapsed_s"] = round(elapsed, 1)
+    parsed["_model"] = model
+    parsed["_eval_count"] = body.get("eval_count")
+    return parsed

deploy/systemd/rpa-mockup-easily.service

@@ -0,0 +1,28 @@
+[Unit]
+Description=Maquette Easily Assure (démo GHT Sud 95) - serveur statique HTTP
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=dom
+Group=dom
+WorkingDirectory=/home/dom/ai/rpa_vision_v3/docs/clients/ght_sud_95/mockup_easily_assure
+ExecStart=/usr/bin/python3 -m http.server 8765 --bind 0.0.0.0
+
+Restart=on-failure
+RestartSec=3
+TimeoutStopSec=10
+
+NoNewPrivileges=true
+PrivateTmp=true
+ProtectSystem=strict
+ProtectHome=read-only
+ReadOnlyPaths=/home/dom/ai/rpa_vision_v3/docs/clients/ght_sud_95/mockup_easily_assure
+
+StandardOutput=journal
+StandardError=journal
+SyslogIdentifier=rpa-mockup-easily
+
+[Install]
+WantedBy=multi-user.target

docs/CARTOGRAPHY.md

@@ -0,0 +1,233 @@
+# Cartographie d'exécution — RPA Vision V3 (Léa)
+
+> **Date** : 26 avril 2026  
+> **Objectif** : carte complète de ce qui est branché, ce qui ne l'est pas, et comment les données transitent.  
+> **Règle** : LIRE CE DOCUMENT AVANT TOUTE MODIFICATION DE CODE.
+
+---
+
+## 1. Point d'entrée : deux chemins disjoints
+
+```
+POST /api/v3/execute/start  (execute.py:1528)
+  ├── execution_mode = "verified"  → run_workflow_verified()  ← CHEMIN ORA
+  └── execution_mode = "basic"|"intelligent"|"debug"  → execute_workflow_thread()  ← CHEMIN LEGACY
+```
+
+**Il existe DEUX exécuteurs distincts** qui dupliquent le chargement des ancres, la boucle d'étapes, le grounding, la gestion d'erreurs. Ils ne partagent que `input_handler.py`.
+
+---
+
+## 2. Chemin LEGACY (modes basic/intelligent/debug)
+
+```
+[API] POST /execute/start (mode=intelligent)
+  → [execute.py:145] execute_workflow_thread()
+    → [execute.py:160] Charge steps depuis DB
+    → BOUCLE sur chaque step:
+      │
+      ├─ RÉFLEXE PRÉ-ÉTAPE (modes intelligent/debug)
+      │   → [input_handler.py:79] check_screen_for_patterns()
+      │       → UIPatternLibrary.find_pattern(ocr_text)               ← BRANCHÉ
+      │   → [input_handler.py:129] handle_detected_pattern()
+      │       → EasyOCR full screen + clic bouton                     ← BRANCHÉ
+      │
+      ├─ CHARGEMENT ANCRE [execute.py:222-256]
+      │   params['visual_anchor'] = {
+      │     screenshot: base64 du crop,
+      │     bounding_box: {x, y, width, height},
+      │     target_text: anchor.target_text,      ← PEUT ÊTRE VIDE ("")
+      │     description: anchor.ocr_description   ← PEUT ÊTRE VIDE ("")
+      │   }
+      │
+      ├─ execute_action(action_type, params) [execute.py:278]
+      │   │
+      │   ├─ ACTION = click_anchor [execute.py:862-1096]
+      │   │  │
+      │   │  ├─ MODE basic: coordonnées statiques (bbox centre)
+      │   │  │
+      │   │  └─ MODE intelligent/debug:
+      │   │     ├─ target_text = anchor.target_text || step.label
+      │   │     │   Si target_text == "click_anchor" et screenshot_base64:
+      │   │     │     → _describe_anchor_image() (VLM qwen2.5vl:3b)  ← BRANCHÉ
+      │   │     │
+      │   │     ├─ MÉTHODE 1: Template matching (cv2)                ← BRANCHÉ
+      │   │     ├─ MÉTHODE 2: CLIP matching (RF-DETR + CLIP)         ← BRANCHÉ
+      │   │     ├─ MÉTHODE 3: OCR → UI-TARS → VLM                   ← BRANCHÉ
+      │   │     └─ ÉCHEC: self-healing interactif                    ← BRANCHÉ
+      │   │
+      │   ├─ ACTION = type_text → safe_type_text()                   ← BRANCHÉ
+      │   ├─ ACTION = wait → sleep + pattern check                   ← BRANCHÉ
+      │   ├─ ACTION = keyboard_shortcut → pyautogui.hotkey()         ← BRANCHÉ
+      │   ├─ ACTION = ai_analyze_text → Ollama                       ← BRANCHÉ
+      │   ├─ ACTION = extract_text → docTR OCR                       ← BRANCHÉ
+      │   └─ ACTION = hover/scroll/focus → coords statiques          ← PAS DE GROUNDING
+```
+
+---
+
+## 3. Chemin ORA (mode "verified")
+
+```
+[API] POST /execute/start (mode=verified)
+  → [execute.py:1349] run_workflow_verified()
+    → [execute.py:1380-1428] Charge steps + ancres (MÊME logique que legacy)
+    → [execute.py:1433] ORALoop(verify_level='none', max_retries=2)
+    │                            ^^^^^^^^^^^^^^^^^^^
+    │                   VÉRIFICATION DÉSACTIVÉE EN DUR
+    │
+    → [ORA:1478] ora.run_workflow(steps=ora_steps)
+      │
+      BOUCLE sur chaque step:
+        │
+        ├─ [ORA:1258] OBSERVE: capture écran + pHash + titre fenêtre
+        │
+        ├─ [ORA:1263] RÉFLEXE DIALOGUE (si pHash changé > 10)
+        │   → DialogHandler.handle_if_dialog(screenshot)               ← BRANCHÉ
+        │     → EasyOCR full screen → mots-clés dialogues connus
+        │     → InfiGUI worker (/tmp/infigui_*)
+        │     → Fallback OCR clic
+        │
+        ├─ [ORA:196] REASON: reason_workflow_step()
+        │   target_text = anchor.target_text || anchor.description
+        │   Si vide ou nom d'action → _describe_anchor_image()         ← CORRIGÉ 26/04
+        │   Si encore vide → label (si pas un nom d'action)
+        │
+        ├─ [ORA:1306] ACT → _act_click()
+        │   │
+        │   ├─ RPA_USE_FAST_PIPELINE=1 (défaut)
+        │   │   → FastSmartThinkPipeline
+        │   │     → FastDetector (RF-DETR 120ms + EasyOCR 192ms)       ← BRANCHÉ
+        │   │     → SmartMatcher (texte+type+position+voisins <1ms)    ← BRANCHÉ
+        │   │     → SignatureStore.lookup() (apprentissage)             ← BRANCHÉ
+        │   │     → Score ≥ 0.90 → action directe                     ← BRANCHÉ
+        │   │     → Score 0.60-0.90 → ThinkArbiter
+        │   │       → UITarsGrounder → InfiGUI worker (/tmp)           ← BRANCHÉ
+        │   │     → Score < 0.60 → ThinkArbiter seul                  ← BRANCHÉ
+        │   │     → ÉCHEC → _try_fallback()
+        │   │       → GroundingPipeline                    ← NON BRANCHÉ (jamais connecté)
+        │   │
+        │   ├─ FALLBACK template matching (cv2, >0.75)                 ← BRANCHÉ
+        │   ├─ FALLBACK OCR (_grounding_ocr)                           ← BRANCHÉ
+        │   └─ DERNIER RECOURS: coords statiques                       ← BRANCHÉ
+        │
+        ├─ [ORA:1337] VÉRIFICATION TITRE (post-action)
+        │   → TitleVerifier → EasyOCR crop 45px                       ← BRANCHÉ
+        │   *** NE LIT RIEN EN VM (titre Windows dans le framebuffer) ← PROBLÈME
+        │
+        ├─ [ORA:1358] VERIFY: verify(pre, post, decision)
+        │   *** DÉSACTIVÉ (verify_level='none') ***                    ← NON BRANCHÉ
+        │
+        └─ [ORA:1362] RECOVERY (5 stratégies)
+            *** JAMAIS ATTEINT ***                                     ← NON BRANCHÉ
+            - _recover_element_not_found (wait+scroll+UI-TARS)
+            - _recover_overlay_blocking (pattern+Win+D)
+            - _recover_wrong_screen (Alt+Tab)
+            - _recover_no_effect (double-clic+décalage)
+            - _classify_error (4 types)
+```
+
+---
+
+## 4. Trace du champ `target_text`
+
+```
+CAPTURE (VWB CapturePanel → capture.py:201-263)
+  → OCR sur crop élargi (docTR)
+  → VLM qwen2.5vl:3b décrit le crop
+  → Si les deux échouent → target_text = ""
+  → Aucune erreur remontée au frontend
+
+STOCKAGE (DB)
+  → VisualAnchor.target_text (nullable) = "" si non renseigné
+
+CHARGEMENT (execute.py:1400-1428)
+  → SI anchor.target_text existe et non vide → injecté dans visual_anchor
+  → SINON → la clé 'target_text' N'EXISTE PAS dans le dict
+
+LEGACY (execute.py:893-907)
+  → target_text = anchor.get('target_text', '')
+  → SI vide ET c'est un nom d'action → _describe_anchor_image()       ← COMPENSE
+  → SINON → fallback sur step_label
+
+ORA (observe_reason_act.py:217) — CORRIGÉ LE 26 AVRIL
+  → target_text = anchor.target_text || anchor.description
+  → SI vide ou nom d'action → _describe_anchor_image()                ← AJOUTÉ
+  → SINON → label (si pas un nom d'action)
+```
+
+---
+
+## 5. Fonctions existantes NON BRANCHÉES
+
+| Fonction | Fichier | Raison |
+|----------|---------|--------|
+| `verify()` + `_classify_error()` + 5 `_recover_*()` | observe_reason_act.py | verify_level='none' en dur |
+| `GroundingPipeline` (ancien) | pipeline.py | set_fallback_pipeline() jamais appelé |
+| `TemplateMatcher` (classe centralisée) | template_matcher.py | Utilisé seulement par GroundingPipeline mort |
+| `ShadowLearningHook` | shadow_learning_hook.py | Jamais importé dans aucun flux |
+| `CognitiveContext` | working_memory.py | Mode instruction seulement |
+| `VLM pre-check` | observe_reason_act.py | `if False:` en dur |
+| hover/focus grounding | execute.py | Coords statiques uniquement |
+| `grounding/server.py` (FastAPI :8200) | server.py | Crash CUDA, remplacé par worker fichiers |
+
+---
+
+## 6. Les 12 systèmes de grounding
+
+| # | Système | Fichier | Branché ? |
+|---|---------|---------|-----------|
+| 1 | Template matching inline (legacy) | execute.py:914 | ✅ Legacy |
+| 2 | Template matching inline (ORA) | ORA:1475 | ✅ ORA fallback |
+| 3 | CLIP matching (IntelligentExecutor) | intelligent_executor.py | ✅ Legacy |
+| 4 | OCR docTR (_grounding_ocr) | input_handler.py:430 | ✅ Legacy + ORA |
+| 5 | UI-TARS Ollama (_grounding_ui_tars) | input_handler.py:513 | ✅ Legacy |
+| 6 | VLM reasoning (_grounding_vlm) | input_handler.py:627 | ✅ Legacy seulement |
+| 7 | FastDetector (RF-DETR + EasyOCR) | fast_detector.py | ✅ ORA |
+| 8 | SmartMatcher | smart_matcher.py | ✅ ORA |
+| 9 | ThinkArbiter → InfiGUI worker | think_arbiter.py + ui_tars_grounder.py | ✅ ORA |
+| 10 | DialogHandler → InfiGUI | dialog_handler.py | ✅ ORA réflexe |
+| 11 | GroundingPipeline (ancien) | pipeline.py | ❌ Jamais connecté |
+| 12 | TemplateMatcher classe | template_matcher.py | ❌ Via GroundingPipeline mort |
+
+---
+
+## 7. Gestion des dialogues (2 systèmes parallèles)
+
+| # | Système | Base de patterns | OCR | Clic | Utilisé par |
+|---|---------|-----------------|-----|------|-------------|
+| 1 | UIPatternLibrary + handle_detected_pattern | 28 patterns builtin | docTR/EasyOCR | OCR find bouton | Legacy |
+| 2 | DialogHandler + KNOWN_DIALOGS | 15 titres connus | EasyOCR full screen | InfiGUI | ORA |
+
+---
+
+## 8. Budget VRAM (configuration actuelle)
+
+| Composant | VRAM | Process |
+|-----------|------|---------|
+| InfiGUI-G1-3B (NF4) | 2.41 GB | Worker indépendant (/tmp) |
+| RF-DETR Medium | 0.8 GB | Process Flask |
+| EasyOCR | ~1 GB (GPU) | Process Flask |
+| Ollama qwen2.5vl:3b (si appelé) | ~3.2 GB | Process Ollama |
+| Chrome + système | ~1.3 GB | — |
+| **Total max** | **~8.7 GB / 12 GB** | |
+
+---
+
+## 9. Fichiers critiques par ordre d'importance
+
+1. `core/execution/observe_reason_act.py` — boucle ORA, _act_click, reason, verify
+2. `visual_workflow_builder/backend/api_v3/execute.py` — API, chargement ancres, legacy executor
+3. `core/grounding/fast_pipeline.py` — pipeline FAST→SMART→THINK
+4. `core/grounding/ui_tars_grounder.py` — client InfiGUI worker
+5. `core/grounding/infigui_worker.py` — worker InfiGUI (process indépendant)
+6. `core/execution/input_handler.py` — OCR, UI-TARS Ollama, safe_type_text, patterns
+7. `core/grounding/dialog_handler.py` — gestion dialogues ORA
+8. `core/grounding/fast_detector.py` — RF-DETR + EasyOCR
+9. `core/grounding/smart_matcher.py` — matching contextuel
+10. `core/knowledge/ui_patterns.py` — patterns réflexes
+
+---
+
+> **Dernière mise à jour** : 26 avril 2026  
+> **Prochaine action** : rebrancher verify + recovery, converger les 2 exécuteurs, nettoyer le code mort.

docs/QW_SMOKE_TESTS_2026-05-06.md

@@ -0,0 +1,343 @@
+# QW Suite Mai — Smoke tests pour validation manuelle
+
+**Date d'exécution prévue** : 2026-05-06 (matin)
+**Branche** : `feature/qw-suite-mai`
+**Durée estimée** : ~1h20 si tout passe, +30 min de debug par test KO
+
+> Coche au fur et à mesure. Si un test KO, applique le "Si KO" puis re-tente.
+> Tout test critique en KO bloquant → kill-switch (procédure §10).
+
+---
+
+## §0. Préflight (5 min)
+
+- [ ] **0.1** Vérifier branche : `git -C /home/dom/ai/rpa_vision_v3 branch --show-current`
+  Attendu : `feature/qw-suite-mai`
+
+- [ ] **0.2** Vérifier les commits récents : `git -C /home/dom/ai/rpa_vision_v3 log --oneline -15`
+  Attendu : voir tous les commits du sprint (spec, plan, QW1×4, QW2×2, QW4×3, docs, fixes A/B/C éventuels)
+
+- [ ] **0.3** Lancer la baseline rapide :
+  ```bash
+  cd /home/dom/ai/rpa_vision_v3
+  .venv/bin/pytest tests/unit/test_monitor_router.py \
+                   tests/unit/test_loop_detector.py \
+                   tests/unit/test_safety_checks_provider.py \
+                   tests/integration/test_grounding_offset.py \
+                   tests/integration/test_loop_detector_replay.py \
+                   tests/integration/test_replay_resume_acknowledgments.py \
+                   -q
+  ```
+  Attendu : `27 passed` (en ~5s).
+  Si KO : ne pas continuer, regarder l'erreur et m'appeler.
+
+- [ ] **0.4** Vérifier les services systemd :
+  ```bash
+  ./svc.sh status
+  ```
+  Attendu : `streaming`, `vwb-backend`, `vwb-frontend`, `dashboard` au minimum running.
+  Si KO : `./svc.sh start` puis re-vérifier.
+
+- [ ] **0.5** Ouvrir un terminal dédié pour `journalctl` (sera utilisé tout le long) :
+  ```bash
+  journalctl -u rpa-streaming -f
+  ```
+  Le laisser ouvert dans un coin de l'écran.
+
+---
+
+## §1. Test QW1 mono-écran (10 min) — RÉGRESSION
+
+**But** : prouver que le sprint n'a pas cassé un workflow Easily Assure existant.
+
+- [ ] **1.1** Ouvrir VWB : `https://vwb.labs.laurinebazin.design` (ou `http://localhost:3002` en local)
+
+- [ ] **1.2** Sélectionner un workflow validé le 30/04 sur Easily Assure (UHCD ou Forfait, le plus simple).
+
+- [ ] **1.3** Cliquer "→ Windows" pour lancer le replay sur Agent V1.
+
+- [ ] **1.4** Pendant l'exécution, dans le terminal `journalctl`, chercher la ligne :
+  ```
+  [BUS] lea:monitor_routed source=focus|composite_fallback ...
+  ```
+  Attendu : au moins 1 occurrence par action visuelle. Sur poste mono-écran, `source=composite_fallback` ou `source=focus` (les deux sont OK).
+
+- [ ] **1.5** Le replay doit terminer **identique** à avant (mêmes clics aux mêmes endroits).
+
+**Verdict** : ☐ OK ☐ KO
+**Si KO** : noter l'écart visuel, kill-switch QW2/QW4 (§10) puis re-tester. Si encore KO → rollback (§11).
+
+---
+
+## §2. Test QW1 multi-écrans (15 min, optionnel) — VALEUR AJOUTÉE
+
+**But** : prouver que le ciblage par écran fonctionne. **Skip si tu n'as qu'un seul écran sur le poste de démo.**
+
+- [ ] **2.1** Brancher un 2ème écran sur le poste Windows (Agent V1).
+
+- [ ] **2.2** Vérifier qu'Agent V1 voit les 2 écrans :
+  ```bash
+  ssh dom@192.168.1.11
+  C:\rpa_vision\.venv\Scripts\python.exe -c "from screeninfo import get_monitors; print([(m.x, m.y, m.width, m.height) for m in get_monitors()])"
+  ```
+  Attendu : 2 tuples affichés.
+
+- [ ] **2.3** Lancer le même workflow Easily Assure (§1.2).
+
+- [ ] **2.4** Dans `journalctl`, observer :
+  - Heartbeats Windows enrichis (cf. fix A) : la session reçoit `monitor_index` en continu.
+  - `[BUS] lea:monitor_routed source=focus idx=0` ou `idx=1` selon où Easily est ouvert.
+
+- [ ] **2.5** Déplacer la fenêtre Easily Assure sur le 2ème écran avant un nouveau replay → relancer → vérifier que le clic atterrit sur le 2ème écran (pas sur le composite).
+
+**Verdict** : ☐ OK ☐ KO ☐ Skipped (pas de 2ème écran)
+
+---
+
+## §3. Test QW2 LoopDetector — boucle artificielle (10 min)
+
+**But** : prouver que Léa s'arrête seule quand elle tourne en rond.
+
+- [ ] **3.1** Dupliquer un workflow simple (1-2 actions) dans VWB.
+
+- [ ] **3.2** Modifier la 1ère action `click` pour qu'elle cible un `target_text` impossible (ex: `target_text="ZZZZZ_INEXISTANT_999"`).
+
+- [ ] **3.3** Lancer le replay.
+
+- [ ] **3.4** Dans `journalctl`, attendre l'apparition de :
+  ```
+  LoopDetector: replay XXX mis en pause — signal=retry_threshold ...
+  [BUS] lea:loop_detected ...
+  ```
+  Délai attendu : ~30-60s (3 retries × ~10s par retry visuel).
+
+- [ ] **3.5** Côté VWB : la bulle `PauseDialog` doit apparaître avec `pause_reason=loop_detected`.
+
+- [ ] **3.6** Cliquer "Annuler" pour arrêter le replay propre.
+
+**Verdict** : ☐ OK ☐ KO
+**Si KO** : vérifier `RPA_LOOP_DETECTOR_ENABLED=1` (défaut). Si toujours KO → log dans `journalctl` doit donner la raison.
+
+---
+
+## §4. Test QW4 backward — workflow legacy (5 min)
+
+**But** : prouver qu'un `pause_for_human` existant continue à marcher exactement comme avant.
+
+- [ ] **4.1** Sélectionner un workflow ayant déjà une action `pause_for_human` (sans `safety_level` ni `safety_checks`).
+
+- [ ] **4.2** Lancer le replay.
+
+- [ ] **4.3** Quand la pause apparaît : la bulle doit être **identique** à avant (juste le `message`, boutons Continuer/Annuler, **PAS** de checklist).
+
+- [ ] **4.4** Dans `journalctl`, vérifier qu'**aucun** appel à Ollama `medgemma:4b` n'est lancé (pas de ligne avec ce modèle).
+
+- [ ] **4.5** Cliquer Continuer → le replay doit reprendre sans erreur.
+
+**Verdict** : ☐ OK ☐ KO
+**Si KO** : régression. Kill-switch QW4 (§10) + re-test.
+
+---
+
+## §5. Test QW4 safety_checks déclaratifs (15 min)
+
+**But** : prouver que la checklist s'affiche et bloque le Continue tant que les required ne sont pas cochés.
+
+- [ ] **5.1** Dans VWB, créer ou modifier un workflow pour insérer une action `pause_for_human` avec :
+  - `message` : "Validation patient"
+  - `safety_level` : `standard` (PAS medical_critical, on isole le déclaratif)
+  - `safety_checks` : 2 entrées
+    - `{id: "check_ipp", label: "IPP correct ?", required: true}`
+    - `{id: "check_diag", label: "Diagnostic confirmé ?", required: true}`
+
+- [ ] **5.2** Sauvegarder, lancer le replay.
+
+- [ ] **5.3** Quand la pause apparaît :
+  - ☐ Bulle "Pause supervisée" affichée
+  - ☐ 2 cases à cocher visibles avec badges `[obligatoire]`
+  - ☐ Bouton "Continuer" désactivé (grisé)
+  - ☐ Aucun badge `[Léa]` (pas de medical_critical → pas de LLM)
+
+- [ ] **5.4** Cocher 1 seule case → Continuer reste désactivé.
+- [ ] **5.5** Cocher la 2ème case → Continuer s'active.
+- [ ] **5.6** Cliquer Continuer → replay reprend.
+
+- [ ] **5.7** Test de sécurité : forcer un POST `/api/v3/replay/resume` sans cocher (via curl) :
+  ```bash
+  # Récupérer le replay_id en cours via VWB ou journalctl
+  curl -X POST http://localhost:5002/api/v3/replay/resume \
+       -H "Content-Type: application/json" \
+       -d '{"replay_id":"<replay_id>","acknowledged_check_ids":[]}'
+  ```
+  Attendu : `400 {"detail": {"error": "required_checks_missing", "missing": ["check_ipp","check_diag"]}}`
+
+**Verdict** : ☐ OK ☐ KO
+
+---
+
+## §6. Test QW4 medical_critical avec LLM (15 min)
+
+**But** : prouver que Léa appelle medgemma:4b en moins de 5s et ajoute des checks contextuels.
+
+- [ ] **6.1** Vérifier que `medgemma:4b` est dispo dans Ollama :
+  ```bash
+  ollama list | grep medgemma
+  ```
+  Attendu : `medgemma:4b` listé. Si absent : `ollama pull medgemma:4b` (3.3 GB).
+
+- [ ] **6.2** Reprendre le workflow §5.1 et changer `safety_level: medical_critical`.
+
+- [ ] **6.3** Lancer le replay.
+
+- [ ] **6.4** Quand la pause apparaît :
+  - ☐ Bulle affichée
+  - ☐ 2 checks déclaratifs (badges `[obligatoire]`)
+  - ☐ 0 à 3 checks supplémentaires avec badge `[Léa]` bleu (tooltip = evidence)
+  - ☐ Délai d'apparition < 5s (sinon le timeout a sauvé)
+
+- [ ] **6.5** Dans `journalctl`, vérifier la ligne :
+  ```
+  [BUS] lea:safety_checks_generated count=N sources=['declarative', 'declarative', 'llm_contextual', ...]
+  ```
+
+- [ ] **6.6** Si Ollama timeout ou crash, vérifier la ligne :
+  ```
+  [BUS] lea:safety_checks_llm_failed reason=... detail=...
+  ```
+  Et la pause s'affiche tout de même avec les 2 checks déclaratifs (fallback safe).
+
+**Verdict** : ☐ OK ☐ KO
+
+---
+
+## §7. Test bus events `lea:*` (5 min)
+
+**But** : agréger les events vus pour audit démo.
+
+- [ ] **7.1** Lancer un replay complet de A à Z (workflow §1 ou §6).
+
+- [ ] **7.2** À la fin, extraire tous les events `[BUS]` du journal :
+  ```bash
+  journalctl -u rpa-streaming --since "10 minutes ago" | grep "\[BUS\]" | tail -30
+  ```
+
+- [ ] **7.3** Vérifier la présence d'au moins :
+  - `lea:monitor_routed` (au moins 1 par action visuelle)
+  - `lea:safety_checks_generated` (si test §6 fait, au moins 1)
+  - `lea:loop_detected` (si test §3 fait)
+
+**Verdict** : ☐ OK ☐ KO
+
+---
+
+## §8. Test kill-switches (10 min) — RÉFLEXE DÉMO
+
+**But** : savoir désactiver QW2/QW4 en pleine démo si ça part en vrille.
+
+- [ ] **8.1** Désactiver QW2 + QW4 :
+  ```bash
+  sudo systemctl edit rpa-streaming
+  # Ajouter sous [Service] :
+  Environment=RPA_LOOP_DETECTOR_ENABLED=0
+  Environment=RPA_SAFETY_CHECKS_LLM_ENABLED=0
+  # Sauver, sortir
+  sudo systemctl restart rpa-streaming
+  ```
+
+- [ ] **8.2** Re-lancer un replay quelconque.
+
+- [ ] **8.3** Dans `journalctl` : vérifier qu'**aucun** event `lea:loop_detected` ni `lea:safety_checks_generated` n'apparaît.
+
+- [ ] **8.4** Réactiver (avant la démo réelle) :
+  ```bash
+  sudo systemctl edit rpa-streaming
+  # Supprimer les 2 lignes Environment=...
+  sudo systemctl restart rpa-streaming
+  ```
+
+- [ ] **8.5** Re-vérifier qu'un replay normal réémet les bus events.
+
+**Verdict** : ☐ OK ☐ KO
+
+---
+
+## §9. Test rollback complet (procédure) — RÉFLEXE D'URGENCE
+
+**À NE PAS exécuter sauf vraie urgence**, juste connaître la commande :
+
+```bash
+cd /home/dom/ai/rpa_vision_v3
+git checkout backup/pre-qw-suite-mai-2026-05-05
+./svc.sh restart
+```
+
+Pour revenir au sprint après rollback :
+```bash
+git checkout feature/qw-suite-mai
+./svc.sh restart
+```
+
+- [ ] **9.1** Lire la procédure, savoir où elle est documentée (`docs/QW_SUITE_MAI.md`).
+
+---
+
+## §10. Si problème en pleine démo
+
+Ordre des réflexes :
+
+1. **Kill-switch QW2 d'abord** (LoopDetector = couche passive, désactiver est sans risque) :
+   ```bash
+   sudo systemctl set-environment RPA_LOOP_DETECTOR_ENABLED=0
+   sudo systemctl restart rpa-streaming
+   ```
+   *(set-environment est plus rapide que `systemctl edit` mais ne survit pas au reboot — OK pour démo)*
+
+2. **Kill-switch QW4 ensuite** si toujours problème :
+   ```bash
+   sudo systemctl set-environment RPA_SAFETY_CHECKS_LLM_ENABLED=0
+   sudo systemctl restart rpa-streaming
+   ```
+
+3. **Rollback complet** si toujours KO (cf. §9).
+
+---
+
+## §11. Récap final
+
+À cocher après tous les tests pour acter "prêt démo" :
+
+- [ ] §1 mono-écran OK (régression zéro)
+- [ ] §2 multi-écrans OK ou skip assumé
+- [ ] §3 LoopDetector OK
+- [ ] §4 backward QW4 OK
+- [ ] §5 safety_checks déclaratifs OK
+- [ ] §6 medical_critical + LLM OK
+- [ ] §7 bus events visibles dans journalctl
+- [ ] §8 kill-switches testés et fonctionnels
+- [ ] §9 procédure rollback connue
+
+**Si tout coché → démo GHT GO** 🟢
+**Si §1 ou §3 ou §5 KO → démo NO-GO sans fix** 🔴
+**Si §2 ou §6 KO → démo OK avec kill-switch QW correspondant** 🟡
+
+---
+
+## Annexes
+
+- Spec : `docs/superpowers/specs/2026-05-05-qw-suite-mai-design.md`
+- Plan d'exécution : `docs/superpowers/plans/2026-05-05-qw-suite-mai.md`
+- Synthèse livraison : `docs/QW_SUITE_MAI.md`
+- Backup distant : `backup/pre-qw-suite-mai-2026-05-05` (Gitea)
+- Tests automatisés (référence 116 passed) :
+  ```bash
+  .venv/bin/pytest tests/unit/test_monitor_router.py \
+                   tests/unit/test_loop_detector.py \
+                   tests/unit/test_safety_checks_provider.py \
+                   tests/integration/test_grounding_offset.py \
+                   tests/integration/test_loop_detector_replay.py \
+                   tests/integration/test_replay_resume_acknowledgments.py \
+                   tests/test_pipeline_e2e.py \
+                   tests/test_phase0_integration.py \
+                   tests/integration/test_stream_processor.py \
+                   -q
+  ```

docs/QW_SUITE_MAI.md

@@ -0,0 +1,101 @@
+# QW Suite Mai 2026 — Synthèse de livraison
+
+Sprint d'amélioration RPA Vision V3, branche `feature/qw-suite-mai`,
+inspiré par exploration comparative de 5 frameworks computer-use
+(Simular Agent-S, browser-use, OpenAI CUA, Coasty, Showlab OOTB).
+
+## Trois quick wins livrés
+
+- **QW1 — Multi-écrans** : capture/grounding par `monitor_index` avec fallbacks
+  focus actif puis composite. Backward 100% sur workflows existants.
+  Ajoute `screeninfo>=0.8` aux dépendances Agent V1.
+- **QW2 — LoopDetector composite** : détection passive de stagnation via
+  3 signaux (CLIP screen_static + action_repeat + retry_threshold).
+  Bascule en `paused_need_help` automatique.
+- **QW4 — Safety checks hybrides** : `pause_for_human` enrichi de checks
+  déclaratifs (workflow) + LLM contextuels (`medgemma:4b` local, timeout 5s,
+  fallback safe). UX VWB avec ChecklistPanel acquittable + audit trail.
+
+## Kill-switches en cas de problème
+
+```bash
+sudo systemctl edit rpa-streaming
+# Ajouter sous [Service] :
+Environment=RPA_LOOP_DETECTOR_ENABLED=0
+Environment=RPA_SAFETY_CHECKS_LLM_ENABLED=0
+sudo systemctl restart rpa-streaming
+```
+
+Rollback complet : `git checkout backup/pre-qw-suite-mai-2026-05-05`.
+
+## Variables d'environnement utiles
+
+| Variable | Défaut | Effet |
+|---|---|---|
+| `RPA_LOOP_DETECTOR_ENABLED` | `1` | Kill-switch QW2 (composite) |
+| `RPA_LOOP_SCREEN_STATIC_THRESHOLD` | `0.99` | Seuil similarité CLIP |
+| `RPA_LOOP_SCREEN_STATIC_N` | `4` | Nb captures consécutives |
+| `RPA_LOOP_ACTION_REPEAT_N` | `3` | Nb actions identiques |
+| `RPA_LOOP_RETRY_THRESHOLD` | `3` | Nb retries cumulés |
+| `RPA_SAFETY_CHECKS_LLM_ENABLED` | `1` | Kill-switch QW4 LLM contextuel |
+| `RPA_SAFETY_CHECKS_LLM_MODEL` | `medgemma:4b` | Modèle Ollama |
+| `RPA_SAFETY_CHECKS_LLM_TIMEOUT_S` | `5` | Timeout dur (secondes) |
+| `RPA_SAFETY_CHECKS_LLM_MAX_CHECKS` | `3` | Max checks LLM ajoutés |
+
+## Smoke tests manuels à effectuer avant la démo GHT
+
+Ces tests demandent une interaction VWB et un Agent V1 actif — non automatisables.
+
+1. **QW1 multi-écrans** : rejouer un workflow Easily Assure validé. Vérifier
+   logs `[BUS] lea:monitor_routed` dans `journalctl -u rpa-streaming`. Le clic
+   doit atterrir au bon endroit même sur un poste à 2 écrans.
+2. **QW2 LoopDetector** : optionnel, difficile à reproduire fiable. Si tu
+   constates un bouclage en démo, vérifier que `paused_need_help` se déclenche
+   automatiquement avec `pause_reason="loop_detected"`.
+3. **QW4 safety_checks** :
+   - Workflow ancien sans `safety_checks` → bulle simple legacy s'affiche
+   - Workflow avec `safety_checks` déclaratifs → ChecklistPanel s'affiche,
+     bouton Continuer désactivé tant que required non cochés
+   - Workflow `safety_level: medical_critical` → checks LLM ajoutés en
+     plus (badge `[Léa]`), apparaissent dans les 5s
+   - POST `/api/v3/replay/resume` sans required acquitté → 400 toast UI
+
+## Tests automatisés (référence)
+
+```
+.venv/bin/pytest tests/unit/test_monitor_router.py \
+                 tests/integration/test_grounding_offset.py \
+                 tests/unit/test_loop_detector.py \
+                 tests/integration/test_loop_detector_replay.py \
+                 tests/unit/test_safety_checks_provider.py \
+                 tests/integration/test_replay_resume_acknowledgments.py \
+                 -v
+```
+
+Référence : 24 tests QW + 89 baseline = 113 passed.
+
+## Référence design
+
+`docs/superpowers/specs/2026-05-05-qw-suite-mai-design.md`
+
+## Référence plan d'exécution
+
+`docs/superpowers/plans/2026-05-05-qw-suite-mai.md`
+
+## Backup
+
+Branche backup poussée Gitea avant le sprint :
+`backup/pre-qw-suite-mai-2026-05-05` + tag `backup-pre-qw-suite-mai-2026-05-05`.
+
+## Statut au 2026-05-05
+
+| Composant | État | Smoke démo nécessaire |
+|---|---|---|
+| QW1 monitor_router + offsets | Livré, tests verts | Oui (multi-écran physique) |
+| QW1 enrichissement Agent V1 | Livré, fallback gracieux si screeninfo absent | Oui (Windows réel) |
+| QW1 hook serveur + cablage executor | Livré (commit fix fc01afa59) | Oui |
+| QW2 LoopDetector module | Livré, tests verts | Non (impossible à reproduire fiable) |
+| QW2 hook api_stream | Livré, tests verts | Non |
+| QW4 SafetyChecksProvider | Livré, tests verts | Oui (avec workflow `medical_critical`) |
+| QW4 endpoint /replay/resume + proxy VWB | Livré, tests verts | Oui (POST avec acknowledged_check_ids) |
+| QW4 PauseDialog + PropertiesPanel | Livré, 0 nouvelle erreur TS | Oui (rendre la bulle dans VWB) |

docs/superpowers/specs/2026-05-05-qw-suite-mai-design.md

@@ -0,0 +1,467 @@
+# Spec — QW Suite Mai 2026
+
+| Champ | Valeur |
+|---|---|
+| Date | 2026-05-05 |
+| Auteur | Dom + Claude (brainstorming structuré) |
+| Branche | `feature/qw-suite-mai` (depuis `feature/feedback-bus`) |
+| Backup | `backup/pre-qw-suite-mai-2026-05-05` à pousser sur Gitea avant 1er commit |
+| Statut | Design approuvé — spec à valider par Dom avant `writing-plans` |
+| Cibles démo | GHT Sud 95 (1ère sem mai 2026, date à confirmer) |
+| Contraintes inviolables | 100% vision · 100% local (Ollama) · backward compatible |
+
+## 1. Contexte & motivation
+
+Suite à l'exploration comparative de 5 frameworks computer-use (Simular Agent-S, browser-use, OpenAI CUA sample, Coasty open-cu, Showlab OOTB), trois quick wins ont été identifiés comme améliorations à fort ratio valeur/risque pour RPA Vision V3, alignés avec la philosophie du projet (vision pure, souveraineté, supervision médicale) :
+
+- **QW1 — Multi-écrans propre** (inspiré OOTB) : capture et grounding sur l'écran cible plutôt que sur le composite tous écrans. Gain de perf grounding + correction des coordonnées.
+- **QW2 — LoopDetector composite** (inspiré browser-use) : détecter quand Léa exécute des actions techniquement valides mais que l'écran ne progresse pas, et escalader vers l'humain plutôt que de tourner en rond muettement.
+- **QW4 — Safety checks hybrides** (inspiré OpenAI CUA + browser-use Pydantic registry) : enrichir l'action `pause_for_human` avec une liste de vérifications à acquitter, mêlant déclaratif (workflow) et contextuel (LLM local).
+
+Effet cumulé attendu : Léa devient observable, robuste et auditable sans rien céder sur le 100% local.
+
+## 2. Décisions de design (récap)
+
+| Sujet | Décision |
+|---|---|
+| Activation | Default-ON pour tous les workflows (Dom recréera ce qui en a besoin) |
+| QW1 — Stratégie ciblage écran | `monitor_index` enregistré à la capture → fallback focus actif → fallback composite (backward) |
+| QW1 — Niveau de stack | Client Agent V1 (capture) + serveur (routeur) + `core/execution/input_handler.py` (capture locale) |
+| QW2 — Signal de boucle | Composite OR : screen_static (CLIP) + action_repeat + retry_threshold |
+| QW2 — Sortie | `replay_state["status"] = "paused_need_help"` avec `pause_reason` structuré |
+| QW4 — Source des checks | Hybride : déclaratif workflow + LLM contextuel sur `safety_level: "medical_critical"` |
+| QW4 — Robustesse LLM | `medgemma:4b` + timeout 5s + `format=json` Ollama + JSON Schema strict + fallback safe (zéro check additionnel) + kill-switch env var |
+| QW4 — UX VWB | Bulle existante préservée + `<ChecklistPanel>` au-dessus de Continuer (bouton désactivé tant que required non cochés) |
+| Ordre de livraison | QW1 → QW2 → QW4 (du moins invasif au plus visible) |
+| Plan timing | Option A : QW1+QW2 avant démo ; QW4 enchaîné dès validation des deux premiers |
+| Kill-switches | Env vars sur QW2 et QW4, surchargeables par `systemctl edit` |
+| Backward compatibility | 100% — aucun champ obligatoire ajouté au DSL ; workflows existants se comportent comme avant |
+
+## 3. Architecture globale
+
+```
+┌─────────────────────────┐         ┌─────────────────────────────────┐
+│   Agent V1 (Windows)    │         │     Serveur Streaming (5005)    │
+│                         │         │                                 │
+│  ┌──────────────────┐   │         │  ┌───────────────────────────┐  │
+│  │ ScreenCapture    │   │         │  │     LoopDetector  [QW2]   │  │
+│  │  + monitor_index │───┼────────▶│  │  • screen_static (CLIP)   │  │
+│  │     [QW1]        │   │  HTTP   │  │  • action_repeat          │  │
+│  └──────────────────┘   │         │  │  • retry_threshold        │  │
+│                         │         │  │  → paused_need_help       │  │
+│  ┌──────────────────┐   │         │  └───────────────────────────┘  │
+│  │ FeedbackBus lea:*│◀──┼─────────┤                                 │
+│  │   chat_window    │   │         │  ┌───────────────────────────┐  │
+│  └──────────────────┘   │         │  │   SafetyChecksProvider    │  │
+└─────────────────────────┘         │  │           [QW4]           │  │
+                                    │  │  • declarative (workflow) │  │
+                                    │  │  • LLM contextual         │  │
+                                    │  │    ‒ medgemma:4b 5s/JSON  │  │
+                                    │  │    ‒ fallback safe        │  │
+                                    │  │  • kill-switch env var    │  │
+                                    │  └───────────────────────────┘  │
+                                    │                                 │
+                                    │  ┌───────────────────────────┐  │
+                                    │  │ MonitorRouter   [QW1]     │  │
+                                    │  │  • cible monitor_index    │  │
+                                    │  │  • fallback focus actif   │  │
+                                    │  └───────────────────────────┘  │
+                                    └─────────────────────────────────┘
+                                                    │
+                                                    ▼
+                                    ┌─────────────────────────────────┐
+                                    │       VWB Frontend (3002)       │
+                                    │                                 │
+                                    │  PauseDialog (étendu) [QW4-UX]  │
+                                    │  • bulle existante préservée    │
+                                    │  • + ChecklistPanel             │
+                                    │    (cases à cocher acquittables)│
+                                    │  • + pause_reason si loop       │
+                                    │  Continuer désactivé tant que   │
+                                    │  required-checks non cochés     │
+                                    └─────────────────────────────────┘
+```
+
+### Principes invariants
+1. Aucun nouveau service, aucune nouvelle DB. Tout dans la stack existante (Agent V1 + serveur 5005 + VWB 3002).
+2. 3 modules serveur isolés (`monitor_router.py`, `loop_detector.py`, `safety_checks_provider.py`) — couplage faible, testables individuellement, désactivables par env var.
+3. Backward compatible : workflows sans nouveaux champs se comportent comme avant.
+4. Kill-switches env vars sur QW2 et QW4, override possible via `systemctl edit` pendant la démo.
+5. 100% vision : QW1 pure capture + grounding ; QW2 réutilise le `_clip_embedder` déjà chargé ; QW4 LLM = Ollama local strict.
+6. Bus `lea:*` étendu de 4 events d'observabilité : `lea:loop_detected`, `lea:safety_checks_generated`, `lea:safety_checks_llm_failed`, `lea:monitor_routed`.
+
+### Surface de modification (ordre A)
+
+| QW | Fichiers nouveaux | Fichiers modifiés |
+|---|---|---|
+| QW1 | `agent_v0/server_v1/monitor_router.py` | `agent_v0/agent_v1/capture/screen_capture.py`, `core/execution/input_handler.py`, `agent_v0/server_v1/api_stream.py` (~10 lignes) |
+| QW2 | `agent_v0/server_v1/loop_detector.py` | `agent_v0/server_v1/replay_engine.py` (~30 lignes), `agent_v0/server_v1/api_stream.py` (~20 lignes) |
+| QW4 | `agent_v0/server_v1/safety_checks_provider.py`, `visual_workflow_builder/frontend_v4/src/components/PauseDialog.tsx` | `agent_v0/server_v1/replay_engine.py`, `agent_v0/server_v1/api_stream.py` (`/replay/resume`), `visual_workflow_builder/frontend_v4/src/types.ts`, `visual_workflow_builder/frontend_v4/src/components/PropertiesPanel.tsx` |
+
+## 4. QW1 — Multi-écrans
+
+### 4.1 Composants
+
+**Client Agent V1** — `agent_v0/agent_v1/capture/screen_capture.py` (existant à modifier)
+- Enrichit chaque heartbeat / event avec :
+  - `monitor_index: int`
+  - `monitors_geometry: [{idx, x, y, w, h, primary}]`
+- Détection via `screeninfo` (port direct depuis Showlab OOTB)
+- Capture de l'écran *actif uniquement* (poids réseau identique à aujourd'hui)
+- Si `screeninfo` indisponible côté Windows : envoie `monitors_geometry: []`, comportement composite préservé
+
+**Serveur** — nouveau `agent_v0/server_v1/monitor_router.py` (~80 lignes)
+- API : `resolve_target_monitor(action: dict, session_state: dict) → MonitorTarget`
+- `MonitorTarget = {idx, offset_x, offset_y, w, h, source: "action" | "focus" | "composite_fallback"}`
+- Stratégie :
+  1. Lit `action.get("monitor_index")` si présent → cible cet écran
+  2. Sinon `session_state.get("last_focused_monitor")` → cible focus actif
+  3. Sinon `monitors[0]` composite (comportement actuel — backward)
+
+**Input local Linux** — `core/execution/input_handler.py` modifs ciblées
+- Signature changée : `_capture_screen(monitor_idx=None) → (image, w, h, offset_x, offset_y)`
+- Quand `monitor_idx` fourni : capture uniquement ce monitor
+- Toutes les fonctions `_grounding_*` (`_grounding_ocr`, `_grounding_ui_tars`, `_grounding_vlm`) propagent l'offset pour traduire les coords retournées en coords absolues écran
+
+### 4.2 Data flow replay
+
+```
+Action [monitor_index=1] reçue par serveur
+  → MonitorRouter.resolve()
+  → target_monitor = {idx:1, offset:(1920,0), w:1920, h:1080, source:"action"}
+  → grounding capture monitor 1 uniquement (image 1920×1080, pas 3840×1080)
+  → UI-TARS / OCR / VLM cherche cible → coords locales (640, 540)
+  → coords absolues = (640+1920, 540+0) = (2560, 540)
+  → pyautogui.click(2560, 540)
+  → bus.emit("lea:monitor_routed", {idx:1, source:"action"})
+```
+
+### 4.3 Error handling
+
+| Cas | Comportement |
+|---|---|
+| `monitor_index` absent (vieille session) | Fallback focus actif, log info `lea:monitor_routed source=focus` |
+| Monitor enregistré n'existe plus (2nd écran débranché) | Fallback focus actif, event `lea:monitor_unavailable` warning |
+| `mss.monitors[i]` hors limites | Fallback `monitors[0]` composite, event `lea:monitor_invalid_index` error |
+| `screeninfo` non installé côté Agent V1 | `monitors_geometry: []`, fallback composite (comportement actuel) — pas de blocage |
+
+### 4.4 Tests QW1
+
+- `tests/unit/test_monitor_router.py` : 4 cas (cible OK, fallback focus, fallback composite, monitor débranché)
+- `tests/integration/test_grounding_offset.py` : capture 1 monitor + clic résolu avec offset (mock pyautogui)
+- Smoke : 1 workflow Easily rejoué, vérification visuelle que le clic atterrit au bon endroit
+
+### 4.5 Compat workflows existants
+
+Aucune action n'a `monitor_index` aujourd'hui → 100% des workflows existants partent en fallback focus actif → comportement quasi-identique au composite actuel mais sur un seul écran (gain de perf grounding même sans recréation de workflow).
+
+## 5. QW2 — LoopDetector composite
+
+### 5.1 Composants
+
+**Nouveau** `agent_v0/server_v1/loop_detector.py` (~150 lignes)
+- Classe `LoopDetector` avec 3 sous-détecteurs
+- API : `evaluate(replay_state, screenshot_history, action_history) → LoopVerdict`
+- `LoopVerdict = {detected: bool, reason: str, signal: str, evidence: dict}`
+
+**Hook** dans `agent_v0/server_v1/api_stream.py`
+- Après chaque `report_action_result`, appel `loop_detector.evaluate(...)` si `RPA_LOOP_DETECTOR_ENABLED=1` (défaut)
+- Si `verdict.detected` :
+  - `replay_state["status"] = "paused_need_help"`
+  - `replay_state["pause_reason"] = verdict.reason`
+  - `replay_state["pause_message"] = f"Léa semble bloquée — {verdict.signal}"`
+  - bus.emit `lea:loop_detected` avec `{signal, evidence, replay_id}`
+
+**Étendu** dans `replay_engine.py` :
+- `_create_replay_state()` ajoute :
+  - `"_screenshot_history": []` (anneau de 5 derniers embeddings CLIP)
+  - `"_action_history": []` (anneau des 5 dernières actions)
+- `_pre_check_screen_state()` continue indépendamment (signal différent : check pré-action vs détection post-action de stagnation)
+
+### 5.2 Signaux composites
+
+| Signal | Détecteur | Seuil par défaut | Source |
+|---|---|---|---|
+| `screen_static` | A | 4 captures consécutives avec CLIP similarity > 0.99 | `_clip_embedder` déjà chargé serveur |
+| `action_repeat` | B | 3 actions consécutives identiques (type + coords) | `_action_history` |
+| `retry_threshold` | C | 3 retries sur même `action_id` | `replay_state["retried_actions"]` (déjà existant) |
+
+Un seul signal positif suffit à déclencher l'escalade.
+
+### 5.3 Data flow
+
+```
+Action exécutée → result reçu via /replay/result
+  ↓
+LoopDetector.evaluate(state, screenshots, actions) si RPA_LOOP_DETECTOR_ENABLED=1
+  ├─ A.check_screen_static() → embed(latest), compare aux N-1 derniers
+  ├─ B.check_action_repeat() → compare action_history[-3:]
+  └─ C.check_retry_threshold() → state["retried_actions"] >= 3
+  ↓
+Si verdict.detected:
+  state["status"] = "paused_need_help"
+  state["pause_reason"] = verdict.reason
+  state["pause_message"] = f"Léa semble bloquée — {verdict.signal} ({evidence})"
+  bus.emit("lea:loop_detected", {signal, evidence, replay_id})
+```
+
+### 5.4 Error handling
+
+| Cas | Comportement |
+|---|---|
+| CLIP embedder unavailable | Signal A désactivé (warning log 1×), B+C continuent. Pas de blocage. |
+| `_screenshot_history` < N | Signal A skip silencieusement (pas assez d'historique) |
+| `embed_image()` lève une exception | Catch + log warning, replay continue (verdict = `detected=False`) |
+| `RPA_LOOP_DETECTOR_ENABLED=0` | Module entier bypassé, comportement antérieur |
+| Faux positif détecté en pleine démo | `RPA_LOOP_DETECTOR_ENABLED=0` via `systemctl edit rpa-streaming` + restart → reprise immédiate |
+
+### 5.5 Configuration env vars
+
+- `RPA_LOOP_DETECTOR_ENABLED=1` (défaut)
+- `RPA_LOOP_SCREEN_STATIC_THRESHOLD=0.99`
+- `RPA_LOOP_SCREEN_STATIC_N=4`
+- `RPA_LOOP_ACTION_REPEAT_N=3`
+- `RPA_LOOP_RETRY_THRESHOLD=3`
+
+### 5.6 Tests QW2
+
+- `tests/unit/test_loop_detector.py` : 8 cas (chaque signal isolé, chaque combinaison, kill-switch, embedder absent)
+- `tests/integration/test_loop_detector_replay.py` : 3 cas — replay simulé qui boucle → vérifier transition `running → paused_need_help` avec bonne raison
+- Pas de smoke démo (impossible à reproduire fiable, on s'appuie sur les tests intégration)
+
+### 5.7 Compat VWB
+
+Aucune côté frontend pour QW2 : la pause `paused_need_help` existe déjà. Le `pause_reason` enrichi sera affiché par le composant `PauseDialog` étendu en QW4. Avant la livraison de QW4, la raison s'affichera en texte dans le `pause_message` (donc utile dès le commit QW2).
+
+## 6. QW4 — Safety checks hybrides
+
+### 6.1 Contrat de l'action étendue (rétro-compatible)
+
+```json
+{
+  "type": "pause_for_human",
+  "parameters": {
+    "message": "Validation T2A avant codage",
+    "safety_level": "medical_critical",
+    "safety_checks": [
+      {"id": "check_ipp", "label": "Vérifier IPP patient", "required": true},
+      {"id": "check_cim10", "label": "Confirmer code CIM-10", "required": true}
+    ]
+  }
+}
+```
+
+`safety_level` et `safety_checks` sont **optionnels**. Action sans ces champs → comportement actuel (bulle simple, aucun appel LLM).
+
+### 6.2 Composants serveur
+
+**Nouveau** `agent_v0/server_v1/safety_checks_provider.py` (~180 lignes)
+- API : `build_pause_payload(action, replay_state, last_screenshot) → PausePayload`
+- Concatène : checks déclaratifs (workflow) + checks contextuels (LLM si `safety_level == "medical_critical"`)
+- Chaque check porte sa source : `source: "declarative" | "llm_contextual"` et son `evidence` (vide pour déclaratif, justification courte pour LLM)
+- Format check final :
+  ```json
+  {
+    "id": "check_xxx",
+    "label": "...",
+    "required": true,
+    "source": "declarative" | "llm_contextual",
+    "evidence": null | "..."
+  }
+  ```
+
+**LLM contextual call** — sous-fonction `_call_llm_for_contextual_checks()`
+- Modèle : `medgemma:4b` (env `RPA_SAFETY_CHECKS_LLM_MODEL`)
+- Timeout dur : 5s (env `RPA_SAFETY_CHECKS_LLM_TIMEOUT_S`)
+- `format=json` natif Ollama + JSON Schema strict :
+  ```json
+  {"additional_checks": [{"label": "string", "evidence": "string"}]}
+  ```
+- Max 3 checks ajoutés (env `RPA_SAFETY_CHECKS_LLM_MAX_CHECKS`)
+- Prompt : screenshot heartbeat actuel + workflow message + liste des checks déclaratifs (évite doublons)
+- Tout échec (timeout, exception, JSON invalide post-schema) → `additional_checks = []`, event `lea:safety_checks_llm_failed`, replay continue
+
+**Hook** dans `replay_engine.py` — branche `action_type == "pause_for_human"`
+- Avant de basculer en `paused_need_help`, appel `safety_checks_provider.build_pause_payload(...)`
+- Stocke `replay_state["safety_checks"] = payload.checks`
+- Stocke `replay_state["pause_payload"] = payload` (pour debug/audit)
+
+**Modif** `api_stream.py` — endpoint `/replay/resume`
+- Reçoit `{acknowledged_check_ids: [...]}` dans le body POST
+- Vérifie : tous les checks `required=true` doivent être dans `acknowledged_check_ids`
+  - Sinon : `400 {error: "required_checks_missing", missing: [...]}`
+- Stocke `replay_state["checks_acknowledged"] = acknowledged_check_ids` (audit trail)
+- Reprise normale du replay
+
+### 6.3 Composants frontend VWB
+
+**Nouveau** `visual_workflow_builder/frontend_v4/src/components/PauseDialog.tsx` (~200 lignes)
+- Props : `pauseMessage`, `pauseReason`, `safetyChecks`, `onResume(ackIds)`, `onCancel`
+- Si `safetyChecks.length === 0` : rend la bulle existante (legacy, comportement actuel)
+- Sinon : bulle + `<ChecklistPanel>` avec checkboxes
+- Bouton Continuer disabled tant que `checks.filter(c => c.required && !checked).length > 0`
+- POST `/replay/resume` avec body `{acknowledged_check_ids: [...]}`
+- Visuel source :
+  - Badge `[Léa]` pour `source: "llm_contextual"` (avec tooltip `evidence`)
+  - Badge `[obligatoire]` pour `required: true`
+
+**Étendu** `types.ts`
+- `PauseAction['parameters']` : ajout `safety_level?`, `safety_checks?`
+- `Execution` : ajout `pause_reason?`, `safety_checks?`
+
+**Étendu** `PropertiesPanel.tsx:1356` — éditeur de l'action `pause_for_human`
+- Section "Niveau de sécurité" : dropdown `standard | medical_critical`
+- Section "Checks à valider" : liste éditable (id + label + required)
+
+### 6.4 Data flow complet
+
+```
+Action pause_for_human (medical_critical, 2 checks déclaratifs) atteinte
+  ↓
+SafetyChecksProvider.build_pause_payload()
+  ├─ checks = [...declarative]  (2 entrées)
+  ├─ if safety_level == "medical_critical" and RPA_SAFETY_CHECKS_LLM_ENABLED=1:
+  │     llm_checks = _call_llm_for_contextual_checks()  (max 3, timeout 5s)
+  │     checks += llm_checks
+  └─ return PausePayload(checks, pause_reason, message)
+  ↓
+replay_state["status"] = "paused_need_help"
+replay_state["safety_checks"] = checks
+bus.emit("lea:safety_checks_generated", {count, sources})
+  ↓
+Frontend VWB poll /replay/state → reçoit pause_payload
+  ↓
+<PauseDialog> rend ChecklistPanel
+  ↓
+Médecin coche les 4 checks → clique Continuer
+  ↓
+POST /replay/resume {acknowledged_check_ids: [4 ids]}
+  ↓
+Serveur valide (tous required acquittés) → reprise du replay
+replay_state["checks_acknowledged"] = [...] (audit trail conservé)
+```
+
+### 6.5 Error handling
+
+| Cas | Comportement |
+|---|---|
+| `safety_level` absent | Pas d'appel LLM ; checks déclaratifs uniquement (peut être `[]`) → bulle simple si vide, checklist sinon |
+| Ollama timeout 5s | Event `lea:safety_checks_llm_failed`, `additional_checks=[]`, fallback safe (déclaratifs seuls) |
+| Ollama JSON malformé (post `format=json` — théoriquement impossible) | Idem timeout, fallback safe |
+| LLM produit un check absurde | Accepté tel quel, le superviseur ignore (pas de filtrage en V1) |
+| Frontend reçoit `safety_checks=[]` | Bulle simple, comportement legacy |
+| `RPA_SAFETY_CHECKS_LLM_ENABLED=0` | Couche LLM bypassée, déclaratifs gardés |
+| `/replay/resume` sans `acknowledged_check_ids` sur required | `400 required_checks_missing` |
+| Frontend POST `/replay/resume` rejeté | Toast d'erreur côté UI, état pause conservé, possibilité de cocher manquants et réessayer |
+
+### 6.6 Configuration env vars
+
+- `RPA_SAFETY_CHECKS_LLM_ENABLED=1` (défaut)
+- `RPA_SAFETY_CHECKS_LLM_MODEL=medgemma:4b`
+- `RPA_SAFETY_CHECKS_LLM_TIMEOUT_S=5`
+- `RPA_SAFETY_CHECKS_LLM_MAX_CHECKS=3`
+
+### 6.7 Tests QW4
+
+- `tests/unit/test_safety_checks_provider.py` : 7 cas (déclaratif seul, hybride réussi, LLM timeout, LLM JSON invalide, kill-switch, max_checks respecté, déclaratif vide)
+- `tests/integration/test_replay_resume_acknowledgments.py` : 3 cas (resume OK, missing required → 400, audit trail enregistré dans `checks_acknowledged`)
+- Frontend : `tests/components/PauseDialog.test.tsx` si suite Vitest existe (à confirmer pendant l'implémentation), sinon test manuel avec checklist écrite
+- Smoke : 1 workflow Easily avec `pause_for_human medical_critical` enrichi → vérification full chain
+
+### 6.8 Compat workflows existants
+
+100% backward — `pause_for_human` actuels n'ont ni `safety_level` ni `safety_checks` → comportement strictement identique. Aucune recréation forcée. Dom enrichira uniquement les workflows qu'il veut promouvoir au niveau `medical_critical`.
+
+## 7. Tests, sécurité de la branche, livraison
+
+### 7.1 Filet de sécurité avant TOUT commit sur `feature/qw-suite-mai`
+
+1. Branche backup poussée Gitea : `backup/pre-qw-suite-mai-2026-05-05`
+2. Capture baseline E2E :
+   ```
+   pytest tests/test_pipeline_e2e.py \
+          tests/test_phase0_integration.py \
+          tests/integration/test_stream_processor.py \
+          -q 2>&1 | tee .qw-baseline.log
+   ```
+3. Smoke démo : 1 dérouler complet d'un workflow Easily Assure, archivage screenshot/vidéo de référence
+4. État VWB validé : démarrage Vite local, ouverture d'un workflow, lancement d'un replay simple, screenshot "tout va bien"
+
+### 7.2 Discipline TDD légère par QW
+
+- Test unitaire écrit AVANT le code de production (1 test rouge → 1 implémentation → vert)
+- Pas de TDD complet sur le frontend (Vitest + React = trop d'outillage à valider en parallèle), test manuel cadré avec checklist écrite
+- Re-run de la suite baseline après chaque commit QW, comparaison au log archivé
+- Toute régression bloque le passage au QW suivant tant qu'elle n'est pas comprise et résolue
+
+### 7.3 Compat VWB — checklist explicite avant commit QW4
+
+- [ ] Workflow ancien (sans `safety_checks`) → bulle simple s'affiche normalement
+- [ ] Workflow nouveau avec `safety_checks` déclaratifs uniquement → checklist visible, **pas** d'appel Ollama (vérification logs)
+- [ ] Workflow `medical_critical` → checklist + checks LLM apparaissent (vérification logs Ollama call dans les 5s)
+- [ ] Continuer désactivé tant que required non cochés
+- [ ] POST `/replay/resume` avec mauvais payload → toast d'erreur côté UI, pas de crash
+- [ ] PropertiesPanel : édition de `safety_checks` ne casse pas l'édition d'autres params de `pause_for_human`
+- [ ] DB `workflows.db` : ouverture après commit, aucune migration cassante (schéma JSON est libre)
+
+### 7.4 Plan de commits
+
+```
+1. test(qw1): tests monitor_router + grounding_offset (rouges)
+2. feat(qw1): multi-écrans piloté par monitor_index (verts)
+3. test(qw2): tests loop_detector composite (rouges)
+4. feat(qw2): LoopDetector composite avec kill-switch env
+5. test(qw4): tests safety_checks_provider + replay_resume (rouges)
+6. feat(qw4): safety_checks hybride déclaratif + LLM contextuel
+7. feat(vwb): PauseDialog + ChecklistPanel + extension PropertiesPanel
+8. docs(qw): docs/QW_SUITE_MAI.md + mise à jour MEMORY.md
+```
+
+Chaque commit signé Co-Authored-By Claude. Branche poussée régulièrement sur Gitea pour backup distant.
+
+### 7.5 Stratégie en cas de régression critique pendant la démo
+
+Kill-switches env vars surchargeables sans redéploiement code :
+
+```
+systemctl edit rpa-streaming
+# Ajouter sous [Service] :
+Environment=RPA_LOOP_DETECTOR_ENABLED=0
+Environment=RPA_SAFETY_CHECKS_LLM_ENABLED=0
+systemctl restart rpa-streaming
+```
+
+Si problème grave au-delà des kill-switches : rollback à `backup/pre-qw-suite-mai-2026-05-05`.
+
+```
+git checkout backup/pre-qw-suite-mai-2026-05-05
+./svc.sh restart
+```
+
+### 7.6 Plan de livraison (Option A validée)
+
+**Avant démo GHT (cette semaine) — Sprint priorité 1**
+- QW1 : tests + code + smoke (~1j)
+- QW2 : tests + code + tests intégration (~2j)
+- Capture baseline + replay smoke entre chaque
+- Si QW1+QW2 validés et probants → on enchaîne sur QW4 dès que possible (Dom accepte le weekend si "effet waouh" auprès de spécialistes RPA)
+
+**Après démo / dès validation QW1+QW2 — Sprint priorité 2**
+- QW4 serveur (provider + LLM + endpoint resume) (~3j)
+- QW4 frontend (PauseDialog + PropertiesPanel) (~2j)
+- Doc + mise à jour MEMORY.md
+
+**Total estimé** : ~8.5j-h ingénieur senior, étalable selon le retour démo.
+
+## 8. Ce qui n'est PAS dans ce spec (out of scope)
+
+- F1 (DSL d'actions Pydantic-first) : refactor de fond, sera son propre spec après la démo.
+- F2 (Mixture-of-Grounding routeur adaptatif) : nécessite F1, son propre spec.
+- F3 (Best-of-N + Reflection) : nécessite F1, son propre spec.
+- QW3 (`output_model_schema` Pydantic pour `extract_text`) : opportuniste, sera intégré quand on touchera `extract_text` pour autre chose.
+- Toute introduction de Pydantic-AI / instructor / Playwright / accessibility-tree : interdit (contraintes inviolables).
+- Refonte du composant pause en `<PauseDialog>` à 3 modes (option C de Q6) : reportée après démo si retour utilisateurs justifie l'investissement.
+
+## 9. Open questions
+
+Aucune. Toutes les décisions de design ont été tranchées via les 7 questions clarifiantes du brainstorming du 5 mai 2026.

tests/integration/test_chat_window_templates.py

@@ -0,0 +1,129 @@
+"""Tests des templates de bulles 'Léa exécute' (J3.4).
+
+On teste les fonctions _tpl_* et _extract_meta de chat_window.py — elles sont
+purement fonctionnelles (input payload → output tuple), aucune UI tkinter
+nécessaire.
+"""
+
+import pytest
+
+from agent_v0.agent_v1.ui import chat_window as cw
+
+
+# ----------------------------------------------------------------------
+# Templates _tpl_*
+# ----------------------------------------------------------------------
+
+def test_tpl_action_started_uses_workflow_name():
+    icon, color, title = cw._tpl_action_started({"workflow": "Demo Urgences UHCD"})
+    assert icon == "▶"
+    assert color == cw.ACTION_ICON_RUN
+    assert "Demo Urgences UHCD" in title
+
+
+def test_tpl_action_started_fallback_when_no_workflow():
+    _, _, title = cw._tpl_action_started({})
+    assert "?" in title
+
+
+def test_tpl_action_progress_uses_step_when_provided():
+    _, _, title = cw._tpl_action_progress({"step": "J'ouvre la fiche patient"})
+    assert title == "J'ouvre la fiche patient"
+
+
+def test_tpl_action_progress_fallback_to_counter():
+    _, _, title = cw._tpl_action_progress({"current": 4, "total": 7})
+    assert "4/7" in title
+
+
+def test_tpl_done_success():
+    icon, color, title = cw._tpl_done({"success": True, "message": "Codage terminé"})
+    assert icon == "✓"
+    assert color == cw.ACTION_ICON_OK
+    assert title == "Codage terminé"
+
+
+def test_tpl_done_failure():
+    icon, color, title = cw._tpl_done({"success": False, "message": "Action échouée"})
+    assert icon == "✗"
+    assert color == cw.ACTION_ICON_ERR
+    assert title == "Action échouée"
+
+
+def test_tpl_done_default_success_when_unspecified():
+    icon, _, _ = cw._tpl_done({})
+    assert icon == "✓"  # par défaut on suppose succès si non précisé
+
+
+def test_tpl_need_confirm_extracts_action_description():
+    icon, _, title = cw._tpl_need_confirm({
+        "action": {"description": "Cliquer sur l'IPP 25003284"}
+    })
+    assert icon == "?"
+    assert "25003284" in title
+
+
+def test_tpl_need_confirm_fallback():
+    _, _, title = cw._tpl_need_confirm({})
+    assert "Validation" in title
+
+
+def test_tpl_step_result_ok():
+    icon, color, _ = cw._tpl_step_result({"status": "ok", "message": "ok"})
+    assert icon == "✓"
+    assert color == cw.ACTION_ICON_OK
+
+
+def test_tpl_step_result_failed():
+    icon, color, _ = cw._tpl_step_result({"status": "failed", "message": "boom"})
+    assert icon == "✗"
+    assert color == cw.ACTION_ICON_ERR
+
+
+def test_tpl_step_result_neutral_status():
+    icon, color, _ = cw._tpl_step_result({"status": "skipped", "message": "passé"})
+    assert icon == "·"
+    assert color == cw.ACTION_ICON_INFO
+
+
+def test_tpl_resumed():
+    icon, color, title = cw._tpl_resumed({})
+    assert icon == "→"
+    assert color == cw.ACTION_ICON_OK
+    assert "Reprise" in title
+
+
+# ----------------------------------------------------------------------
+# Dispatch — chaque event lea:* (hors paused/acks) doit avoir un template
+# ----------------------------------------------------------------------
+
+def test_all_relevant_events_have_a_template():
+    expected = {
+        "lea:action_started", "lea:action_progress", "lea:done",
+        "lea:need_confirm", "lea:step_result", "lea:resumed",
+    }
+    assert set(cw._ACTION_TEMPLATES.keys()) == expected
+
+
+# ----------------------------------------------------------------------
+# _extract_meta
+# ----------------------------------------------------------------------
+
+def test_extract_meta_with_workflow():
+    meta = cw._extract_meta({"workflow": "Demo Urgences"})
+    assert meta == "Demo Urgences"
+
+
+def test_extract_meta_with_progress():
+    meta = cw._extract_meta({"workflow": "Demo Urgences", "current": 4, "total": 7})
+    assert "Demo Urgences" in meta
+    assert "étape 4/7" in meta
+
+
+def test_extract_meta_with_replay_id_truncated():
+    meta = cw._extract_meta({"replay_id": "rep_abcdef0123456789"})
+    assert "#789" in meta or "456789" in meta  # 6 derniers caractères
+
+
+def test_extract_meta_empty_payload():
+    assert cw._extract_meta({}) == ""

tests/integration/test_feedback_bus.py

@@ -0,0 +1,164 @@
+"""Tests du bus feedback Léa (events lea:* via Flask-SocketIO).
+
+Couvre J2.5 et J2.6 :
+- Flag LEA_FEEDBACK_BUS=0 → _emit_lea no-op, _emit_dual ne propage que l'event legacy
+- Flag LEA_FEEDBACK_BUS=1 → _emit_lea propage 'lea:{event}', _emit_dual propage les deux
+
+Approche : on intercepte socketio.emit avec monkeypatch (plus fiable que test_client
+de Flask-SocketIO qui ne capte pas toujours les broadcasts hors contexte requête).
+"""
+
+import importlib
+
+import pytest
+
+
+def _reload_app(monkeypatch, flag_value: str):
+    monkeypatch.setenv("LEA_FEEDBACK_BUS", flag_value)
+    import agent_chat.app as app_mod
+    importlib.reload(app_mod)
+    return app_mod
+
+
+def _capture_emits(monkeypatch, app_mod):
+    calls = []
+    monkeypatch.setattr(
+        app_mod.socketio, "emit",
+        lambda event, payload=None, **kwargs: calls.append((event, payload, kwargs)),
+    )
+    return calls
+
+
+@pytest.fixture
+def app_off(monkeypatch):
+    return _reload_app(monkeypatch, "0")
+
+
+@pytest.fixture
+def app_on(monkeypatch):
+    return _reload_app(monkeypatch, "1")
+
+
+def test_flag_off_by_default(monkeypatch):
+    monkeypatch.delenv("LEA_FEEDBACK_BUS", raising=False)
+    import agent_chat.app as app_mod
+    importlib.reload(app_mod)
+    assert app_mod.LEA_FEEDBACK_BUS is False
+
+
+def test_flag_accepts_truthy_values(monkeypatch):
+    for truthy in ["1", "true", "True", "yes", "on", "TRUE"]:
+        monkeypatch.setenv("LEA_FEEDBACK_BUS", truthy)
+        import agent_chat.app as app_mod
+        importlib.reload(app_mod)
+        assert app_mod.LEA_FEEDBACK_BUS is True, f"{truthy!r} devrait activer le flag"
+
+
+def test_emit_lea_noop_when_flag_off(app_off, monkeypatch):
+    calls = _capture_emits(monkeypatch, app_off)
+    app_off._emit_lea("paused", {"workflow": "demo", "reason": "test"})
+    assert calls == []
+
+
+def test_emit_lea_emits_when_flag_on(app_on, monkeypatch):
+    calls = _capture_emits(monkeypatch, app_on)
+    app_on._emit_lea("paused", {"workflow": "demo", "reason": "test"})
+    assert len(calls) == 1
+    event, payload, _ = calls[0]
+    assert event == "lea:paused"
+    assert payload == {"workflow": "demo", "reason": "test"}
+
+
+def test_emit_dual_emits_only_legacy_when_flag_off(app_off, monkeypatch):
+    calls = _capture_emits(monkeypatch, app_off)
+    app_off._emit_dual("execution_started", "action_started", {"workflow": "demo"})
+    assert len(calls) == 1
+    assert calls[0][0] == "execution_started"
+
+
+def test_emit_dual_emits_both_when_flag_on(app_on, monkeypatch):
+    calls = _capture_emits(monkeypatch, app_on)
+    payload = {"workflow": "demo", "params": {"k": "v"}}
+    app_on._emit_dual("execution_started", "action_started", payload)
+    events = [c[0] for c in calls]
+    assert "execution_started" in events
+    assert "lea:action_started" in events
+    assert len(calls) == 2
+
+
+def test_emit_dual_preserves_kwargs(app_on, monkeypatch):
+    """broadcast=True et autres kwargs Flask-SocketIO doivent être propagés au legacy."""
+    calls = _capture_emits(monkeypatch, app_on)
+    app_on._emit_dual("execution_cancelled", "cancelled", {}, broadcast=True)
+    legacy_call = next(c for c in calls if c[0] == "execution_cancelled")
+    assert legacy_call[2].get("broadcast") is True
+
+
+def test_emit_lea_silenced_on_socketio_error(app_on, monkeypatch):
+    """Une exception dans socketio.emit ne doit jamais remonter."""
+    def boom(*args, **kwargs):
+        raise RuntimeError("socketio fail")
+    monkeypatch.setattr(app_on.socketio, "emit", boom)
+    app_on._emit_lea("paused", {"x": 1})
+
+
+# ----------------------------------------------------------------------
+# J3.5 — Handlers SocketIO depuis ChatWindow
+# ----------------------------------------------------------------------
+
+class _FakeResponse:
+    def __init__(self, ok=True, status_code=200, text=""):
+        self.ok = ok
+        self.status_code = status_code
+        self.text = text
+
+
+def test_replay_resume_handler_relays_post_to_streaming(app_on, monkeypatch):
+    """Le handler 'lea:replay_resume' doit POSTer sur /replay/{id}/resume du streaming."""
+    captured = {}
+
+    def fake_post(url, headers=None, **kwargs):
+        captured["url"] = url
+        captured["headers"] = headers
+        return _FakeResponse(ok=True, status_code=200)
+
+    monkeypatch.setattr(app_on.http_requests, "post", fake_post)
+    emit_calls = _capture_emits(monkeypatch, app_on)
+
+    app_on.handle_lea_replay_resume({"replay_id": "rep_abc123"})
+
+    assert "rep_abc123" in captured["url"]
+    assert captured["url"].endswith("/api/v1/traces/stream/replay/rep_abc123/resume")
+    # Le bus doit propager un ack
+    acked = [c for c in emit_calls if c[0] == "lea:resume_acked"]
+    assert len(acked) == 1
+    assert acked[0][1]["status"] == "ok"
+
+
+def test_replay_resume_handler_emits_error_on_http_failure(app_on, monkeypatch):
+    monkeypatch.setattr(
+        app_on.http_requests, "post",
+        lambda *a, **k: _FakeResponse(ok=False, status_code=500, text="boom"),
+    )
+    emit_calls = _capture_emits(monkeypatch, app_on)
+    app_on.handle_lea_replay_resume({"replay_id": "rep_x"})
+    acked = [c for c in emit_calls if c[0] == "lea:resume_acked"]
+    assert acked[0][1]["status"] == "error"
+    assert acked[0][1]["http_status"] == 500
+
+
+def test_replay_resume_handler_emits_error_on_no_replay_id(app_on, monkeypatch):
+    emit_calls = _capture_emits(monkeypatch, app_on)
+    app_on.handle_lea_replay_resume({})
+    acked = [c for c in emit_calls if c[0] == "lea:resume_acked"]
+    assert acked[0][1]["status"] == "error"
+    assert "replay_id manquant" in acked[0][1]["detail"]
+
+
+def test_replay_abort_handler_stops_local_execution(app_on, monkeypatch):
+    app_on.execution_status["running"] = True
+    emit_calls = _capture_emits(monkeypatch, app_on)
+    app_on.handle_lea_replay_abort({"replay_id": "rep_y"})
+    assert app_on.execution_status["running"] is False
+    acked = [c for c in emit_calls if c[0] == "lea:abort_acked"]
+    assert acked[0][1]["status"] == "ok"

tests/integration/test_feedback_bus_client.py

@@ -0,0 +1,164 @@
+"""Tests FeedbackBusClient (J3.2).
+
+On mock python-socketio pour ne pas ouvrir de vraie connexion réseau.
+Le test E2E réel (vraie connexion bus 5004) est différé à J4.3.
+"""
+
+from unittest.mock import patch
+
+import pytest
+
+from agent_v0.agent_v1.network.feedback_bus import FeedbackBusClient, LEA_EVENTS
+
+
+def test_init_creates_socketio_client():
+    bus = FeedbackBusClient("http://localhost:5004")
+    assert bus._sio is not None
+    assert bus.connected is False
+
+
+def test_init_strips_trailing_slash():
+    bus = FeedbackBusClient("http://localhost:5004/")
+    assert bus._url == "http://localhost:5004"
+
+
+def test_lea_events_registered():
+    bus = FeedbackBusClient("http://localhost:5004")
+    handlers = bus._sio.handlers.get('/', {})
+    for ev in LEA_EVENTS:
+        assert ev in handlers, f"Handler {ev!r} non enregistré sur le client"
+
+
+def test_dispatch_calls_callback():
+    received = []
+    bus = FeedbackBusClient(
+        "http://localhost:5004",
+        on_event=lambda e, p: received.append((e, p)),
+    )
+    bus._dispatch('lea:paused', {'workflow': 'demo', 'reason': 'incertain'})
+    assert received == [('lea:paused', {'workflow': 'demo', 'reason': 'incertain'})]
+
+
+def test_dispatch_handles_none_payload():
+    received = []
+    bus = FeedbackBusClient(
+        "http://localhost:5004",
+        on_event=lambda e, p: received.append((e, p)),
+    )
+    bus._dispatch('lea:done', None)
+    assert received == [('lea:done', {})]
+
+
+def test_dispatch_silenced_on_callback_error():
+    """Une exception dans le callback consommateur ne doit jamais remonter."""
+    def boom(event, payload):
+        raise RuntimeError("callback fail")
+    bus = FeedbackBusClient("http://localhost:5004", on_event=boom)
+    bus._dispatch('lea:paused', {})  # ne doit pas raise
+
+
+def test_default_callback_is_silent():
+    """Sans callback fourni, le dispatch ne casse pas."""
+    bus = FeedbackBusClient("http://localhost:5004")
+    bus._dispatch('lea:paused', {'x': 1})  # ne doit pas raise
+
+
+def test_token_in_authorization_header():
+    bus = FeedbackBusClient("http://localhost:5004", token="abc123")
+    captured = {}
+
+    def fake_connect(url, headers=None, **kwargs):
+        captured['headers'] = headers
+        raise RuntimeError("stop here")
+
+    with patch.object(bus._sio, 'connect', side_effect=fake_connect):
+        bus._run()
+
+    assert captured['headers']['Authorization'] == 'Bearer abc123'
+
+
+def test_no_token_means_no_auth_header():
+    bus = FeedbackBusClient("http://localhost:5004")
+    captured = {}
+
+    def fake_connect(url, headers=None, **kwargs):
+        captured['headers'] = headers
+        raise RuntimeError("stop here")
+
+    with patch.object(bus._sio, 'connect', side_effect=fake_connect):
+        bus._run()
+
+    assert 'Authorization' not in captured['headers']
+
+
+def test_run_silenced_on_connect_error():
+    """connect() qui raise ne doit pas faire crasher le thread."""
+    bus = FeedbackBusClient("http://localhost:5004")
+    with patch.object(bus._sio, 'connect', side_effect=ConnectionError("boom")):
+        bus._run()  # ne doit pas raise
+
+
+def test_start_is_idempotent():
+    """Un second start() pendant que le thread tourne ne doit pas en créer un autre."""
+    import threading
+    bus = FeedbackBusClient("http://localhost:5004")
+    block = threading.Event()
+    with patch.object(bus, '_run', side_effect=lambda: block.wait(timeout=2)):
+        bus.start()
+        first_thread = bus._thread
+        bus.start()
+        second_thread = bus._thread
+        block.set()
+    assert first_thread is second_thread, "start() doit être idempotent quand un thread tourne"
+
+
+def test_stop_when_not_connected_is_silent():
+    bus = FeedbackBusClient("http://localhost:5004")
+    bus.stop()  # ne doit pas raise même si jamais connecté
+
+
+def test_stop_silenced_on_disconnect_error():
+    bus = FeedbackBusClient("http://localhost:5004")
+    # Forcer connected=True sur l'instance et faire raise disconnect()
+    with patch.object(bus._sio, 'disconnect', side_effect=RuntimeError("boom")):
+        bus._sio.connected = True
+        bus.stop()  # ne doit pas raise
+
+
+# ----------------------------------------------------------------------
+# J3.5 — Actions utilisateur (resume_replay / abort_replay)
+# ----------------------------------------------------------------------
+
+def test_resume_replay_emits_when_connected():
+    bus = FeedbackBusClient("http://localhost:5004")
+    bus._sio.connected = True
+    with patch.object(bus._sio, 'emit') as mock_emit:
+        ok = bus.resume_replay("rep_abc")
+    assert ok is True
+    mock_emit.assert_called_once_with("lea:replay_resume", {"replay_id": "rep_abc"})
+
+
+def test_resume_replay_returns_false_when_disconnected():
+    bus = FeedbackBusClient("http://localhost:5004")
+    # _sio.connected reste False par défaut
+    with patch.object(bus._sio, 'emit') as mock_emit:
+        ok = bus.resume_replay("rep_abc")
+    assert ok is False
+    mock_emit.assert_not_called()
+
+
+def test_abort_replay_emits_when_connected():
+    bus = FeedbackBusClient("http://localhost:5004")
+    bus._sio.connected = True
+    with patch.object(bus._sio, 'emit') as mock_emit:
+        ok = bus.abort_replay("rep_xyz")
+    assert ok is True
+    mock_emit.assert_called_once_with("lea:replay_abort", {"replay_id": "rep_xyz"})
+
+
+def test_safe_emit_silenced_on_error():
+    bus = FeedbackBusClient("http://localhost:5004")
+    bus._sio.connected = True
+    with patch.object(bus._sio, 'emit', side_effect=RuntimeError("boom")):
+        ok = bus.resume_replay("rep_abc")
+    assert ok is False  # erreur avalée silencieusement

tests/integration/test_grounding_offset.py

@@ -0,0 +1,41 @@
+# tests/integration/test_grounding_offset.py
+"""Tests intégration pour la propagation d'offset multi-écrans (QW1)."""
+import pytest
+from unittest.mock import patch, MagicMock
+
+from core.execution import input_handler
+
+
+@pytest.fixture
+def mock_screen():
+    """Mock une capture mss : retourne un PIL Image factice + offsets."""
+    from PIL import Image
+    img = Image.new("RGB", (1920, 1080), color="white")
+    return img
+
+
+def test_capture_screen_default_returns_composite_when_no_idx(mock_screen):
+    """_capture_screen() sans monitor_idx → composite, offset (0, 0)."""
+    with patch("core.execution.input_handler.mss") as mock_mss:
+        ctx = mock_mss.mss.return_value.__enter__.return_value
+        ctx.monitors = [{"left": 0, "top": 0, "width": 3840, "height": 1080}]
+        ctx.grab.return_value = MagicMock(size=(3840, 1080), bgra=b"\x00" * (3840 * 1080 * 4))
+        with patch("core.execution.input_handler.PILImage.frombytes", return_value=mock_screen):
+            screen, w, h, ox, oy = input_handler._capture_screen()
+    assert (w, h, ox, oy) == (3840, 1080, 0, 0)
+
+
+def test_capture_screen_targets_specific_monitor_with_offset(mock_screen):
+    """_capture_screen(monitor_idx=1) → cible monitors[2] (mss skip [0]), offset = monitor.left."""
+    with patch("core.execution.input_handler.mss") as mock_mss:
+        ctx = mock_mss.mss.return_value.__enter__.return_value
+        # mss layout : [0]=composite, [1]=primary, [2]=secondary
+        ctx.monitors = [
+            {"left": 0, "top": 0, "width": 3840, "height": 1080},
+            {"left": 0, "top": 0, "width": 1920, "height": 1080},
+            {"left": 1920, "top": 0, "width": 1920, "height": 1080},
+        ]
+        ctx.grab.return_value = MagicMock(size=(1920, 1080), bgra=b"\x00" * (1920 * 1080 * 4))
+        with patch("core.execution.input_handler.PILImage.frombytes", return_value=mock_screen):
+            screen, w, h, ox, oy = input_handler._capture_screen(monitor_idx=1)
+    assert (w, h, ox, oy) == (1920, 1080, 1920, 0)

tests/integration/test_loop_detector_replay.py

@@ -0,0 +1,61 @@
+# tests/integration/test_loop_detector_replay.py
+"""Tests intégration : un replay simulé qui boucle bascule en paused_need_help."""
+import pytest
+from unittest.mock import MagicMock
+
+from agent_v0.server_v1.loop_detector import LoopDetector
+
+
+def test_replay_state_transitions_to_paused_on_screen_static():
+    """Cas : 4 screenshots identiques → replay passe à paused_need_help."""
+    embedder = MagicMock()
+    embedder.embed_image.return_value = [1.0, 0.0, 0.0]  # constant
+    detector = LoopDetector(clip_embedder=embedder)
+
+    state = {
+        "replay_id": "r_test",
+        "status": "running",
+        "retried_actions": 0,
+        "_screenshot_history": ["img1", "img2", "img3", "img4"],  # 4 images factices
+        "_action_history": [
+            {"type": "click", "x_pct": 0.1, "y_pct": 0.1},
+            {"type": "type", "x_pct": 0.2, "y_pct": 0.2},
+        ],
+    }
+    verdict = detector.evaluate(state, state["_screenshot_history"], state["_action_history"])
+
+    # Simuler ce que ferait api_stream après verdict
+    if verdict.detected:
+        state["status"] = "paused_need_help"
+        state["pause_reason"] = verdict.reason
+        state["pause_message"] = f"signal={verdict.signal}"
+
+    assert state["status"] == "paused_need_help"
+    assert state["pause_reason"] == "loop_detected"
+    assert "screen_static" in state["pause_message"]
+
+
+def test_replay_state_transitions_on_action_repeat():
+    """Cas : 3 actions identiques → paused_need_help signal action_repeat."""
+    detector = LoopDetector(clip_embedder=None)
+    actions = [{"type": "click", "x_pct": 0.5, "y_pct": 0.5}] * 3
+    state = {"replay_id": "r2", "status": "running", "retried_actions": 0,
+             "_screenshot_history": [], "_action_history": actions}
+
+    verdict = detector.evaluate(state, [], actions)
+    assert verdict.detected and verdict.signal == "action_repeat"
+
+
+def test_kill_switch_keeps_replay_running(monkeypatch):
+    """Avec RPA_LOOP_DETECTOR_ENABLED=0 le replay continue même en boucle."""
+    monkeypatch.setenv("RPA_LOOP_DETECTOR_ENABLED", "0")
+    embedder = MagicMock()
+    embedder.embed_image.return_value = [1.0, 0.0, 0.0]
+    detector = LoopDetector(clip_embedder=embedder)
+
+    state = {"retried_actions": 10,
+             "_screenshot_history": ["img1"] * 10,
+             "_action_history": [{"type": "click", "x_pct": 0.5, "y_pct": 0.5}] * 10}
+
+    verdict = detector.evaluate(state, state["_screenshot_history"], state["_action_history"])
+    assert verdict.detected is False

tests/integration/test_pause_for_human.py

@@ -0,0 +1,131 @@
+"""Tests de l'action pause_for_human (C.5).
+
+Vérifie la chaîne :
+- Validation côté replay_engine accepte le nouveau type
+- Conversion edge → action normalisée préserve le message
+- Bridge VWB → core mappe correctement
+- Le bridge VWB construit bien un edge avec action.type='pause_for_human'
+"""
+
+from agent_v0.server_v1.replay_engine import (
+    _ALLOWED_ACTION_TYPES,
+    _validate_replay_action,
+    _edge_to_normalized_actions,
+)
+from visual_workflow_builder.backend.services.learned_workflow_bridge import (
+    VWB_ACTION_TO_CORE,
+    convert_vwb_to_core_workflow,
+    _vwb_params_to_core,
+)
+
+
+# ----------------------------------------------------------------------
+# Validation pipeline (replay_engine)
+# ----------------------------------------------------------------------
+
+def test_pause_for_human_in_allowed_types():
+    assert "pause_for_human" in _ALLOWED_ACTION_TYPES
+
+
+def test_validate_pause_for_human_action_valid():
+    action = {"type": "pause_for_human", "parameters": {"message": "Valider UHCD ?"}}
+    assert _validate_replay_action(action) is None
+
+
+def test_validate_pause_for_human_no_params_still_valid():
+    """Le validateur ne doit pas exiger 'message' (fallback côté handler)."""
+    action = {"type": "pause_for_human"}
+    assert _validate_replay_action(action) is None
+
+
+# ----------------------------------------------------------------------
+# Conversion edge → action normalisée
+# ----------------------------------------------------------------------
+
+class _FakeAction:
+    def __init__(self, type_, parameters=None):
+        self.type = type_
+        self.target = None
+        self.parameters = parameters or {}
+
+
+class _FakeEdge:
+    def __init__(self, action, edge_id="e1", from_node="n1", to_node="n2"):
+        self.edge_id = edge_id
+        self.from_node = from_node
+        self.to_node = to_node
+        self.action = action
+
+
+def test_edge_to_action_pause_for_human_preserves_message():
+    edge = _FakeEdge(_FakeAction(
+        "pause_for_human",
+        parameters={"message": "Tu valides UHCD ?"},
+    ))
+    actions = _edge_to_normalized_actions(edge, params={})
+    assert len(actions) == 1
+    a = actions[0]
+    assert a["type"] == "pause_for_human"
+    assert a["parameters"]["message"] == "Tu valides UHCD ?"
+    assert "x_pct" not in a  # action logique, pas de coords
+    assert "y_pct" not in a
+
+
+def test_edge_to_action_pause_for_human_default_message():
+    edge = _FakeEdge(_FakeAction("pause_for_human", parameters={}))
+    actions = _edge_to_normalized_actions(edge, params={})
+    assert actions[0]["parameters"]["message"] == "Validation requise"
+
+
+def test_edge_to_action_pause_for_human_carries_edge_metadata():
+    edge = _FakeEdge(
+        _FakeAction("pause_for_human", parameters={"message": "x"}),
+        edge_id="edge_42", from_node="n_src", to_node="n_dst",
+    )
+    actions = _edge_to_normalized_actions(edge, params={})
+    a = actions[0]
+    assert a["edge_id"] == "edge_42"
+    assert a["from_node"] == "n_src"
+    assert a["to_node"] == "n_dst"
+    assert "action_id" in a
+
+
+# ----------------------------------------------------------------------
+# Bridge VWB → core
+# ----------------------------------------------------------------------
+
+def test_vwb_action_to_core_passthrough():
+    assert VWB_ACTION_TO_CORE["pause_for_human"] == "pause_for_human"
+
+
+def test_vwb_params_to_core_preserves_message():
+    core_params = _vwb_params_to_core("pause_for_human", {"message": "Coucou"})
+    assert core_params == {"message": "Coucou"}
+
+
+def test_vwb_params_to_core_default_message():
+    core_params = _vwb_params_to_core("pause_for_human", {})
+    assert core_params["message"] == "Validation requise"
+
+
+def test_export_vwb_workflow_with_pause_step():
+    """Un workflow VWB contenant une step pause_for_human doit produire un edge
+    avec action.type='pause_for_human' et message dans parameters."""
+    workflow_data = {"id": "wf_demo", "name": "Demo Urgences", "description": ""}
+    steps_data = [
+        {"id": "s1", "action_type": "click_anchor", "parameters": {"target_text": "25003284"}, "label": "Clic IPP"},
+        {"id": "s2", "action_type": "pause_for_human", "parameters": {"message": "Valider UHCD ?"}, "label": "Pause"},
+        {"id": "s3", "action_type": "click_anchor", "parameters": {"target_text": "Enregistrer"}, "label": "Clic Enregistrer"},
+    ]
+    core = convert_vwb_to_core_workflow(workflow_data, steps_data)
+    assert core["learning_state"] == "COACHING"
+    assert len(core["nodes"]) == 3
+    assert len(core["edges"]) == 2
+
+    # L'edge sortant du node de pause doit avoir le bon type + message
+    pause_edges = [
+        e for e in core["edges"]
+        if e["action"]["type"] == "pause_for_human"
+    ]
+    assert len(pause_edges) == 1
+    assert pause_edges[0]["action"]["parameters"]["message"] == "Valider UHCD ?"

tests/integration/test_replay_resume_acknowledgments.py

@@ -0,0 +1,52 @@
+# tests/integration/test_replay_resume_acknowledgments.py
+"""Tests intégration : /replay/resume valide les acquittements de safety_checks (QW4)."""
+import pytest
+
+
+def test_resume_accepts_when_all_required_acknowledged():
+    """État pause + tous required acquittés → reprise OK."""
+    state = {
+        "status": "paused_need_help",
+        "safety_checks": [
+            {"id": "c1", "label": "X", "required": True, "source": "declarative", "evidence": None},
+            {"id": "c2", "label": "Y", "required": True, "source": "declarative", "evidence": None},
+        ],
+        "checks_acknowledged": [],
+    }
+    # Simuler la validation côté serveur
+    acknowledged = ["c1", "c2"]
+    required_ids = {c["id"] for c in state["safety_checks"] if c["required"]}
+    missing = required_ids - set(acknowledged)
+    assert missing == set()  # rien ne manque → reprise OK
+
+
+def test_resume_rejects_when_required_missing():
+    """État pause + un required non acquitté → 400 required_checks_missing."""
+    state = {
+        "status": "paused_need_help",
+        "safety_checks": [
+            {"id": "c1", "label": "X", "required": True, "source": "declarative", "evidence": None},
+            {"id": "c2", "label": "Y", "required": False, "source": "llm_contextual", "evidence": "..."},
+        ],
+        "checks_acknowledged": [],
+    }
+    acknowledged = ["c2"]  # only optional
+    required_ids = {c["id"] for c in state["safety_checks"] if c["required"]}
+    missing = required_ids - set(acknowledged)
+    assert missing == {"c1"}  # c1 manquant → resume doit retourner 400
+
+
+def test_resume_audit_trail_stored():
+    """checks_acknowledged contient les ids reçus (audit)."""
+    state = {
+        "status": "paused_need_help",
+        "safety_checks": [
+            {"id": "c1", "required": True, "label": "X", "source": "declarative", "evidence": None},
+        ],
+        "checks_acknowledged": [],
+    }
+    acknowledged = ["c1"]
+    state["checks_acknowledged"] = acknowledged
+    state["status"] = "running"
+    assert state["checks_acknowledged"] == ["c1"]
+    assert state["status"] == "running"

tests/integration/test_t2a_extract.py

@@ -0,0 +1,282 @@
+"""Tests des actions extract_text et t2a_decision (C+.5/.6).
+
+Couvre :
+- _resolve_runtime_vars : templating {{var}} / {{var.field}}
+- _handle_extract_text_action : OCR mocké, stockage variable
+- _handle_t2a_decision_action : analyze_dpi mocké, stockage JSON
+- _edge_to_normalized_actions pour les 2 types
+- Bridge VWB → core (mapping + paramètres)
+"""
+
+from unittest.mock import patch
+
+import pytest
+
+from agent_v0.server_v1.replay_engine import (
+    _ALLOWED_ACTION_TYPES,
+    _SERVER_SIDE_ACTION_TYPES,
+    _resolve_runtime_vars,
+    _handle_extract_text_action,
+    _handle_t2a_decision_action,
+    _edge_to_normalized_actions,
+    _create_replay_state,
+)
+from visual_workflow_builder.backend.services.learned_workflow_bridge import (
+    VWB_ACTION_TO_CORE,
+    convert_vwb_to_core_workflow,
+    _vwb_params_to_core,
+)
+
+
+# ----------------------------------------------------------------------
+# Templating runtime
+# ----------------------------------------------------------------------
+
+def test_resolve_simple_var():
+    r = _resolve_runtime_vars("Patient {{ipp}}", {"ipp": "25003284"})
+    assert r == "Patient 25003284"
+
+
+def test_resolve_field_access():
+    r = _resolve_runtime_vars(
+        "{{result.decision}} car {{result.justification}}",
+        {"result": {"decision": "UHCD", "justification": "asthme + insuf coro"}},
+    )
+    assert "UHCD car asthme + insuf coro" == r
+
+
+def test_resolve_missing_var_kept_intact():
+    r = _resolve_runtime_vars("Hello {{absent}} world", {"x": "y"})
+    assert r == "Hello {{absent}} world"
+
+
+def test_resolve_missing_field_kept_intact():
+    r = _resolve_runtime_vars("{{var.absent}}", {"var": {"present": "x"}})
+    assert r == "{{var.absent}}"
+
+
+def test_resolve_in_dict_recursive():
+    r = _resolve_runtime_vars(
+        {"msg": "IPP {{ipp}}", "nested": {"k": "{{ipp}}"}, "list": ["{{age}}"]},
+        {"ipp": "X", "age": 77},
+    )
+    assert r == {"msg": "IPP X", "nested": {"k": "X"}, "list": ["77"]}
+
+
+def test_resolve_empty_vars_noop():
+    val = {"k": "{{var}}"}
+    assert _resolve_runtime_vars(val, {}) == val
+    assert _resolve_runtime_vars(val, None) == val
+
+
+def test_resolve_non_string_passthrough():
+    assert _resolve_runtime_vars(42, {"x": "y"}) == 42
+    assert _resolve_runtime_vars(None, {"x": "y"}) is None
+
+
+def test_resolve_handles_whitespace_in_braces():
+    r = _resolve_runtime_vars("{{ ipp }}", {"ipp": "X"})
+    assert r == "X"
+
+
+# ----------------------------------------------------------------------
+# Action types & types serveur
+# ----------------------------------------------------------------------
+
+def test_extract_text_in_allowed():
+    assert "extract_text" in _ALLOWED_ACTION_TYPES
+
+
+def test_t2a_decision_in_allowed():
+    assert "t2a_decision" in _ALLOWED_ACTION_TYPES
+
+
+def test_server_side_types():
+    assert _SERVER_SIDE_ACTION_TYPES == {"extract_text", "t2a_decision"}
+
+
+# ----------------------------------------------------------------------
+# Handler extract_text
+# ----------------------------------------------------------------------
+
+def test_handle_extract_text_stores_variable():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    last_hb = {"sess": {"path": "/fake/heartbeat.png", "timestamp": 0}}
+    action = {
+        "type": "extract_text",
+        "parameters": {"output_var": "texte_motif", "paragraph": True},
+    }
+    with patch(
+        "core.llm.extract_text_from_image",
+        return_value="Patient asthme peakflow 260",
+    ):
+        ok = _handle_extract_text_action(action, state, "sess", last_hb)
+    assert ok is True
+    assert state["variables"]["texte_motif"] == "Patient asthme peakflow 260"
+
+
+def test_handle_extract_text_no_heartbeat_stores_empty():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    last_hb = {}  # pas de heartbeat
+    action = {"type": "extract_text", "parameters": {"output_var": "v"}}
+    ok = _handle_extract_text_action(action, state, "sess", last_hb)
+    assert ok is False
+    assert state["variables"]["v"] == ""
+
+
+def test_handle_extract_text_default_var_name():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    last_hb = {"sess": {"path": "/x.png", "timestamp": 0}}
+    action = {"type": "extract_text", "parameters": {}}
+    with patch("core.llm.extract_text_from_image", return_value="abc"):
+        _handle_extract_text_action(action, state, "sess", last_hb)
+    assert "extracted_text" in state["variables"]
+
+
+# ----------------------------------------------------------------------
+# Handler t2a_decision
+# ----------------------------------------------------------------------
+
+def test_handle_t2a_decision_stores_json():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    action = {
+        "type": "t2a_decision",
+        "parameters": {
+            "input_template": "Patient 78 ans, asthme, peakflow 260",
+            "output_var": "decision_t2a",
+            "model": "qwen2.5:7b",
+        },
+    }
+    fake_result = {
+        "decision": "REQUALIFICATION_HOSPITALISATION",
+        "justification": "Surveillance continue requise",
+        "confiance": "elevee",
+        "_elapsed_s": 4.2,
+    }
+    with patch("core.llm.analyze_dpi", return_value=fake_result):
+        ok = _handle_t2a_decision_action(action, state)
+    assert ok is True
+    assert state["variables"]["decision_t2a"]["decision"] == "REQUALIFICATION_HOSPITALISATION"
+
+
+def test_handle_t2a_decision_empty_input_returns_indetermine():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    action = {"type": "t2a_decision", "parameters": {"input_template": "", "output_var": "r"}}
+    ok = _handle_t2a_decision_action(action, state)
+    assert ok is False
+    assert state["variables"]["r"]["decision"] == "INDETERMINE"
+
+
+def test_handle_t2a_decision_analyze_exception():
+    state = _create_replay_state("rep1", "wf", "sess", 3)
+    action = {"type": "t2a_decision", "parameters": {"input_template": "x", "output_var": "r"}}
+    with patch("core.llm.analyze_dpi", side_effect=RuntimeError("ollama down")):
+        ok = _handle_t2a_decision_action(action, state)
+    assert ok is False
+    assert state["variables"]["r"]["decision"] == "INDETERMINE"
+    assert "ollama down" in state["variables"]["r"]["_error"]
+
+
+# ----------------------------------------------------------------------
+# Edge → action normalisée
+# ----------------------------------------------------------------------
+
+class _FakeAction:
+    def __init__(self, type_, parameters=None):
+        self.type = type_
+        self.target = None
+        self.parameters = parameters or {}
+
+
+class _FakeEdge:
+    def __init__(self, action, edge_id="e1", from_node="n1", to_node="n2"):
+        self.edge_id = edge_id
+        self.from_node = from_node
+        self.to_node = to_node
+        self.action = action
+
+
+def test_edge_to_action_extract_text():
+    edge = _FakeEdge(_FakeAction(
+        "extract_text",
+        parameters={"output_var": "texte_examens", "paragraph": True},
+    ))
+    actions = _edge_to_normalized_actions(edge, params={})
+    assert len(actions) == 1
+    a = actions[0]
+    assert a["type"] == "extract_text"
+    assert a["parameters"]["output_var"] == "texte_examens"
+    assert a["parameters"]["paragraph"] is True
+
+
+def test_edge_to_action_t2a_decision():
+    edge = _FakeEdge(_FakeAction(
+        "t2a_decision",
+        parameters={
+            "input_template": "{{texte_motif}}",
+            "output_var": "result",
+            "model": "qwen2.5:7b",
+        },
+    ))
+    actions = _edge_to_normalized_actions(edge, params={})
+    a = actions[0]
+    assert a["type"] == "t2a_decision"
+    assert a["parameters"]["input_template"] == "{{texte_motif}}"
+    assert a["parameters"]["output_var"] == "result"
+    assert a["parameters"]["model"] == "qwen2.5:7b"
+
+
+# ----------------------------------------------------------------------
+# Bridge VWB → core
+# ----------------------------------------------------------------------
+
+def test_vwb_extract_text_passthrough():
+    assert VWB_ACTION_TO_CORE["extract_text"] == "extract_text"
+
+
+def test_vwb_t2a_decision_passthrough():
+    assert VWB_ACTION_TO_CORE["t2a_decision"] == "t2a_decision"
+
+
+def test_vwb_params_extract_text_preserves_output_var():
+    p = _vwb_params_to_core("extract_text", {"output_var": "v", "paragraph": False})
+    assert p == {"output_var": "v", "paragraph": False}
+
+
+def test_vwb_params_extract_text_legacy_variable_name():
+    """Compat avec l'ancien paramètre variable_name côté VWB."""
+    p = _vwb_params_to_core("extract_text", {"variable_name": "v_legacy"})
+    assert p["output_var"] == "v_legacy"
+
+
+def test_vwb_params_t2a_decision_preserves_all():
+    p = _vwb_params_to_core("t2a_decision", {
+        "input_template": "DPI {{ipp}}",
+        "output_var": "dec",
+        "model": "qwen2.5:7b",
+    })
+    assert p == {"input_template": "DPI {{ipp}}", "output_var": "dec", "model": "qwen2.5:7b"}
+
+
+def test_export_workflow_with_t2a_chain():
+    """Workflow VWB extract_text → t2a_decision → pause_for_human export propre."""
+    workflow_data = {"id": "wf_t2a", "name": "Demo T2A"}
+    steps_data = [
+        {"id": "s1", "action_type": "click_anchor", "parameters": {"target_text": "25003284"}, "label": "Clic IPP"},
+        {"id": "s2", "action_type": "extract_text", "parameters": {"output_var": "dpi"}, "label": "OCR"},
+        {"id": "s3", "action_type": "t2a_decision", "parameters": {
+            "input_template": "{{dpi}}", "output_var": "dec", "model": "qwen2.5:7b",
+        }, "label": "Analyse"},
+        {"id": "s4", "action_type": "pause_for_human", "parameters": {
+            "message": "Décision : {{dec.decision}} — {{dec.justification}}",
+        }, "label": "Validation"},
+        {"id": "s5", "action_type": "click_anchor", "parameters": {"target_text": "Enregistrer"}, "label": "Clic Enregistrer"},
+    ]
+    core = convert_vwb_to_core_workflow(workflow_data, steps_data)
+    edge_types = [e["action"]["type"] for e in core["edges"]]
+    assert "extract_text" in edge_types
+    assert "t2a_decision" in edge_types
+    assert "pause_for_human" in edge_types
+    # Vérifier que le templating est bien transporté
+    t2a_edge = next(e for e in core["edges"] if e["action"]["type"] == "t2a_decision")
+    assert t2a_edge["action"]["parameters"]["input_template"] == "{{dpi}}"

tests/unit/test_loop_detector.py

@@ -0,0 +1,96 @@
+# tests/unit/test_loop_detector.py
+"""Tests unitaires pour LoopDetector composite (QW2)."""
+import os
+import pytest
+from unittest.mock import MagicMock
+
+from agent_v0.server_v1.loop_detector import LoopDetector, LoopVerdict
+
+
+@pytest.fixture
+def detector():
+    """LoopDetector avec embedder mocké (signal A toujours dispo)."""
+    embedder = MagicMock()
+    # Par défaut : 4 embeddings tous identiques → similarity 1.0
+    embedder.embed_image.return_value = [1.0, 0.0, 0.0]
+    return LoopDetector(clip_embedder=embedder)
+
+
+def _state(retried=0, n_screenshots=0, n_actions=0):
+    return {
+        "retried_actions": retried,
+        "_screenshot_history": [[1.0, 0.0, 0.0]] * n_screenshots,
+        "_action_history": [{"type": "click", "x_pct": 0.5, "y_pct": 0.5}] * n_actions,
+    }
+
+
+def test_screen_static_triggers_when_n_identical_embeddings(detector):
+    """Signal A : 4 captures identiques (similarity > 0.99) → detected."""
+    state = _state(n_screenshots=4)
+    verdict = detector.evaluate(state, screenshots=state["_screenshot_history"], actions=[])
+    assert verdict.detected is True
+    assert verdict.signal == "screen_static"
+
+
+def test_screen_static_skipped_when_history_too_short(detector):
+    """Signal A : moins de N captures → pas de détection."""
+    state = _state(n_screenshots=2)
+    verdict = detector.evaluate(state, screenshots=state["_screenshot_history"], actions=[])
+    # Si seul A pourrait déclencher mais skip, et B/C pas remplis : detected=False
+    assert verdict.detected is False
+
+
+def test_action_repeat_triggers_when_n_identical_actions(detector):
+    """Signal B : 3 actions consécutives identiques → detected."""
+    state = _state(n_actions=3)
+    verdict = detector.evaluate(state, screenshots=[], actions=state["_action_history"])
+    assert verdict.detected is True
+    assert verdict.signal == "action_repeat"
+
+
+def test_action_repeat_skipped_when_actions_differ(detector):
+    """Signal B : actions différentes → pas de détection."""
+    actions = [
+        {"type": "click", "x_pct": 0.1, "y_pct": 0.1},
+        {"type": "click", "x_pct": 0.2, "y_pct": 0.2},
+        {"type": "click", "x_pct": 0.3, "y_pct": 0.3},
+    ]
+    verdict = detector.evaluate(_state(), screenshots=[], actions=actions)
+    assert verdict.detected is False
+
+
+def test_retry_threshold_triggers_at_3(detector):
+    """Signal C : retried_actions >= 3 → detected."""
+    state = _state(retried=3)
+    verdict = detector.evaluate(state, screenshots=[], actions=[])
+    assert verdict.detected is True
+    assert verdict.signal == "retry_threshold"
+
+
+def test_kill_switch_disables_all_signals(monkeypatch, detector):
+    """Si RPA_LOOP_DETECTOR_ENABLED=0 → toujours detected=False."""
+    monkeypatch.setenv("RPA_LOOP_DETECTOR_ENABLED", "0")
+    state = _state(retried=10, n_screenshots=10, n_actions=10)
+    verdict = detector.evaluate(state, screenshots=state["_screenshot_history"],
+                                actions=state["_action_history"])
+    assert verdict.detected is False
+
+
+def test_embedder_unavailable_skips_signal_A_continues_others():
+    """Si CLIP embedder None → signal A skip, B et C continuent."""
+    detector = LoopDetector(clip_embedder=None)
+    # Trigger signal C
+    state = _state(retried=3)
+    verdict = detector.evaluate(state, screenshots=[], actions=[])
+    assert verdict.detected is True
+    assert verdict.signal == "retry_threshold"
+
+
+def test_embedder_exception_does_not_crash(detector):
+    """Si embed_image lève une exception → log + verdict detected=False."""
+    detector.clip_embedder.embed_image.side_effect = RuntimeError("CUDA OOM")
+    state = _state(n_screenshots=4)
+    # Ne doit PAS lever : signal A devient inerte
+    verdict = detector.evaluate(state, screenshots=state["_screenshot_history"], actions=[])
+    # Signal A inerte, B/C pas remplis → detected False
+    assert verdict.detected is False

tests/unit/test_monitor_router.py

@@ -0,0 +1,51 @@
+# tests/unit/test_monitor_router.py
+"""Tests unitaires pour MonitorRouter (QW1)."""
+import pytest
+
+from agent_v0.server_v1.monitor_router import resolve_target_monitor, MonitorTarget
+
+
+# Geometry de référence pour les 3 tests : 2 écrans côte à côte
+TWO_MONITORS = [
+    {"idx": 0, "x": 0, "y": 0, "w": 1920, "h": 1080, "primary": True},
+    {"idx": 1, "x": 1920, "y": 0, "w": 1920, "h": 1080, "primary": False},
+]
+
+
+def test_resolve_uses_action_monitor_index_when_present():
+    """Si action.monitor_index présent et valide → cible cet écran."""
+    action = {"monitor_index": 1}
+    session_state = {"monitors_geometry": TWO_MONITORS, "last_focused_monitor": 0}
+    result = resolve_target_monitor(action, session_state)
+    assert result.idx == 1
+    assert result.offset_x == 1920
+    assert result.offset_y == 0
+    assert result.source == "action"
+
+
+def test_resolve_falls_back_to_focused_monitor_when_action_missing():
+    """Si action.monitor_index absent → fallback focus actif."""
+    action = {}  # pas de monitor_index
+    session_state = {"monitors_geometry": TWO_MONITORS, "last_focused_monitor": 1}
+    result = resolve_target_monitor(action, session_state)
+    assert result.idx == 1
+    assert result.source == "focus"
+
+
+def test_resolve_falls_back_to_composite_when_geometry_empty():
+    """Si geometry vide (vieux Agent V1) → fallback composite (idx=-1, offset=0)."""
+    action = {}
+    session_state = {"monitors_geometry": [], "last_focused_monitor": None}
+    result = resolve_target_monitor(action, session_state)
+    assert result.source == "composite_fallback"
+    assert result.offset_x == 0
+    assert result.offset_y == 0
+
+
+def test_resolve_falls_back_when_action_index_out_of_range():
+    """Si action.monitor_index hors limites (écran débranché) → fallback focus."""
+    action = {"monitor_index": 5}  # n'existe pas
+    session_state = {"monitors_geometry": TWO_MONITORS, "last_focused_monitor": 0}
+    result = resolve_target_monitor(action, session_state)
+    assert result.idx == 0
+    assert result.source == "focus"

tests/unit/test_safety_checks_provider.py

@@ -0,0 +1,111 @@
+# tests/unit/test_safety_checks_provider.py
+"""Tests unitaires SafetyChecksProvider (QW4)."""
+import json
+import pytest
+from unittest.mock import patch, MagicMock
+
+from agent_v0.server_v1.safety_checks_provider import build_pause_payload, PausePayload
+
+
+def _action(safety_level=None, declarative_checks=None, message="Validation"):
+    params = {"message": message}
+    if safety_level:
+        params["safety_level"] = safety_level
+    if declarative_checks is not None:
+        params["safety_checks"] = declarative_checks
+    return {"type": "pause_for_human", "parameters": params}
+
+
+def test_only_declarative_when_no_safety_level():
+    """Pas de safety_level → uniquement les checks déclaratifs, pas d'appel LLM."""
+    decl = [{"id": "c1", "label": "Vérifier IPP", "required": True}]
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks") as mock_llm:
+        payload = build_pause_payload(_action(declarative_checks=decl), {}, last_screenshot=None)
+    mock_llm.assert_not_called()
+    assert len(payload.checks) == 1
+    assert payload.checks[0]["source"] == "declarative"
+
+
+def test_hybrid_appends_llm_checks_on_medical_critical(monkeypatch):
+    """safety_level=medical_critical → LLM appelé, checks concaténés."""
+    decl = [{"id": "c1", "label": "Vérifier IPP", "required": True}]
+    llm_resp = [{"label": "Nom patient suspect à l'écran", "evidence": "vu un nom différent"}]
+
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks",
+               return_value=llm_resp) as mock_llm:
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=decl),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    mock_llm.assert_called_once()
+    assert len(payload.checks) == 2
+    assert payload.checks[0]["source"] == "declarative"
+    assert payload.checks[1]["source"] == "llm_contextual"
+    assert payload.checks[1]["evidence"] == "vu un nom différent"
+
+
+def test_llm_timeout_falls_back_to_declarative_only():
+    """LLM timeout → additional_checks=[], pas de crash, déclaratifs gardés."""
+    decl = [{"id": "c1", "label": "Vérifier IPP", "required": True}]
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks",
+               return_value=[]) as mock_llm:
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=decl),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    assert len(payload.checks) == 1
+    assert payload.checks[0]["source"] == "declarative"
+
+
+def test_llm_invalid_response_falls_back():
+    """Si _call_llm retourne [] (parse échoué en interne) → fallback safe."""
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks",
+               return_value=[]):
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=[]),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    assert payload.checks == []
+
+
+def test_kill_switch_disables_llm_call(monkeypatch):
+    """RPA_SAFETY_CHECKS_LLM_ENABLED=0 → LLM jamais appelé."""
+    monkeypatch.setenv("RPA_SAFETY_CHECKS_LLM_ENABLED", "0")
+    decl = [{"id": "c1", "label": "X", "required": True}]
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks") as mock_llm:
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=decl),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    mock_llm.assert_not_called()
+    assert len(payload.checks) == 1
+
+
+def test_max_checks_respected(monkeypatch):
+    """RPA_SAFETY_CHECKS_LLM_MAX_CHECKS=2 → max 2 checks LLM ajoutés."""
+    monkeypatch.setenv("RPA_SAFETY_CHECKS_LLM_MAX_CHECKS", "2")
+    decl = []
+    llm_resp = [
+        {"label": f"Check {i}", "evidence": f"e{i}"} for i in range(5)
+    ]
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks",
+               return_value=llm_resp[:2]):  # provider tronque déjà
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=decl),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    assert len(payload.checks) == 2
+
+
+def test_empty_declarative_with_llm_returns_only_llm():
+    """Pas de déclaratif + LLM ajoute 2 checks → payload contient les 2."""
+    llm_resp = [{"label": "Vérifier date", "evidence": "date 1900 suspecte"},
+                {"label": "Vérifier devise", "evidence": "montant en USD au lieu d'EUR"}]
+    with patch("agent_v0.server_v1.safety_checks_provider._call_llm_for_contextual_checks",
+               return_value=llm_resp):
+        payload = build_pause_payload(
+            _action(safety_level="medical_critical", declarative_checks=[]),
+            {}, last_screenshot="/tmp/fake.png",
+        )
+    assert len(payload.checks) == 2
+    assert all(c["source"] == "llm_contextual" for c in payload.checks)

tests/unit/test_template_matcher.py

@@ -0,0 +1,311 @@
+"""Tests pour core/grounding/template_matcher.py"""
+
+import base64
+import io
+import time
+from unittest.mock import MagicMock, patch
+
+import cv2
+import numpy as np
+import pytest
+from PIL import Image
+
+from core.grounding.template_matcher import MatchResult, TemplateMatcher
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _make_image(w: int, h: int, color: tuple = (128, 128, 128)) -> Image.Image:
+    """Crée une image PIL unie."""
+    img = Image.new('RGB', (w, h), color)
+    return img
+
+
+def _pil_to_b64(img: Image.Image) -> str:
+    """Encode une image PIL en base64 PNG."""
+    buf = io.BytesIO()
+    img.save(buf, format='PNG')
+    return base64.b64encode(buf.getvalue()).decode()
+
+
+def _make_screen_with_target(
+    screen_w: int = 800,
+    screen_h: int = 600,
+    target_x: int = 300,
+    target_y: int = 200,
+    target_w: int = 60,
+    target_h: int = 40,
+):
+    """Crée un screen bruité avec un motif unique et l'ancre correspondante.
+
+    Le screen a un fond aléatoire (bruit) pour que le template matching
+    ne puisse matcher qu'à l'endroit exact du motif injecté.
+    """
+    rng = np.random.RandomState(42)
+    # Fond bruité — chaque pixel est différent, pas de faux match possible
+    screen = rng.randint(0, 256, (screen_h, screen_w, 3), dtype=np.uint8)
+
+    # Injecter un motif déterministe unique (damier rouge/bleu)
+    target = np.zeros((target_h, target_w, 3), dtype=np.uint8)
+    for r in range(target_h):
+        for c in range(target_w):
+            if (r + c) % 2 == 0:
+                target[r, c] = [255, 0, 0]   # rouge
+            else:
+                target[r, c] = [0, 0, 255]   # bleu
+    screen[target_y:target_y + target_h, target_x:target_x + target_w] = target
+    screen_pil = Image.fromarray(screen)
+
+    # L'ancre est exactement le même motif
+    anchor_pil = Image.fromarray(target)
+
+    expected_cx = target_x + target_w // 2
+    expected_cy = target_y + target_h // 2
+
+    return screen_pil, anchor_pil, expected_cx, expected_cy
+
+
+# ---------------------------------------------------------------------------
+# Tests MatchResult
+# ---------------------------------------------------------------------------
+
+class TestMatchResult:
+    def test_fields(self):
+        r = MatchResult(x=100, y=200, score=0.85, method='template', time_ms=5.0)
+        assert r.x == 100
+        assert r.y == 200
+        assert r.score == 0.85
+        assert r.method == 'template'
+        assert r.time_ms == 5.0
+        assert r.scale == 1.0  # default
+
+    def test_with_scale(self):
+        r = MatchResult(x=10, y=20, score=0.9, method='template_multiscale', time_ms=12.0, scale=0.95)
+        assert r.scale == 0.95
+
+
+# ---------------------------------------------------------------------------
+# Tests TemplateMatcher — init
+# ---------------------------------------------------------------------------
+
+class TestTemplateMatcherInit:
+    def test_defaults(self):
+        m = TemplateMatcher()
+        assert m.threshold == 0.75
+        assert m.multiscale is False
+        assert m.grayscale is False
+
+    def test_custom_params(self):
+        m = TemplateMatcher(threshold=0.5, multiscale=True, grayscale=True, scales=[1.0, 0.8])
+        assert m.threshold == 0.5
+        assert m.multiscale is True
+        assert m.grayscale is True
+        assert m.scales == [1.0, 0.8]
+
+
+# ---------------------------------------------------------------------------
+# Tests TemplateMatcher — _decode_anchor
+# ---------------------------------------------------------------------------
+
+class TestDecodeAnchor:
+    def test_pil_passthrough(self):
+        img = _make_image(50, 50)
+        result = TemplateMatcher._decode_anchor(None, img)
+        assert result is img
+
+    def test_b64_decode(self):
+        img = _make_image(50, 50, (255, 0, 0))
+        b64 = _pil_to_b64(img)
+        result = TemplateMatcher._decode_anchor(b64, None)
+        assert result is not None
+        assert result.size == (50, 50)
+
+    def test_b64_with_data_prefix(self):
+        img = _make_image(30, 30)
+        b64 = "data:image/png;base64," + _pil_to_b64(img)
+        result = TemplateMatcher._decode_anchor(b64, None)
+        assert result is not None
+
+    def test_none_inputs(self):
+        result = TemplateMatcher._decode_anchor(None, None)
+        assert result is None
+
+    def test_invalid_b64(self):
+        result = TemplateMatcher._decode_anchor("not-valid-base64!!!", None)
+        assert result is None
+
+
+# ---------------------------------------------------------------------------
+# Tests TemplateMatcher — match_screen avec screen_pil fourni
+# ---------------------------------------------------------------------------
+
+class TestMatchScreenWithPIL:
+    def test_exact_match(self):
+        screen, anchor, cx, cy = _make_screen_with_target()
+        m = TemplateMatcher(threshold=0.75)
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        assert result is not None
+        assert abs(result.x - cx) <= 1
+        assert abs(result.y - cy) <= 1
+        assert result.score > 0.9
+        assert result.method == 'template'
+        assert result.time_ms >= 0
+
+    def test_no_match(self):
+        # Screen bruité, ancre = damier unique absent du screen
+        rng = np.random.RandomState(123)
+        screen_np = rng.randint(0, 256, (600, 800, 3), dtype=np.uint8)
+        screen = Image.fromarray(screen_np)
+
+        # Ancre = damier régulier non présent dans le bruit
+        anchor_np = np.zeros((40, 60, 3), dtype=np.uint8)
+        for r in range(40):
+            for c in range(60):
+                anchor_np[r, c] = [255, 255, 0] if (r + c) % 2 == 0 else [0, 255, 255]
+        anchor = Image.fromarray(anchor_np)
+
+        m = TemplateMatcher(threshold=0.75)
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        assert result is None
+
+    def test_b64_anchor(self):
+        screen, anchor, cx, cy = _make_screen_with_target()
+        b64 = _pil_to_b64(anchor)
+        m = TemplateMatcher(threshold=0.75)
+        result = m.match_screen(anchor_b64=b64, screen_pil=screen)
+        assert result is not None
+        assert abs(result.x - cx) <= 1
+
+    def test_anchor_bigger_than_screen(self):
+        screen = _make_image(100, 100)
+        anchor = _make_image(200, 200)
+        m = TemplateMatcher()
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        assert result is None
+
+    def test_threshold_configurable(self):
+        screen, anchor, cx, cy = _make_screen_with_target()
+        # Avec un seuil de 0.999, le match exact devrait quand même passer (score=1.0)
+        m = TemplateMatcher(threshold=0.999)
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        # Le score d'un match pixel-perfect peut être 1.0 ou très proche
+        # On accepte les deux cas
+        if result:
+            assert result.score >= 0.999
+
+
+# ---------------------------------------------------------------------------
+# Tests TemplateMatcher — multi-scale
+# ---------------------------------------------------------------------------
+
+class TestMultiscale:
+    def test_multiscale_exact(self):
+        screen, anchor, cx, cy = _make_screen_with_target()
+        m = TemplateMatcher(threshold=0.75, multiscale=True)
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        assert result is not None
+        assert abs(result.x - cx) <= 2
+        assert abs(result.y - cy) <= 2
+        assert result.score > 0.9
+
+    def test_multiscale_scaled_anchor(self):
+        """L'ancre a été capturée à une échelle légèrement différente.
+
+        On utilise un motif plus gros (bloc de couleur unie) pour que le resize
+        ne détruise pas le pattern comme avec un damier fin.
+        """
+        # Screen bruité + gros bloc rouge
+        rng = np.random.RandomState(42)
+        screen_np = rng.randint(50, 200, (600, 800, 3), dtype=np.uint8)
+        target = np.full((80, 120, 3), dtype=np.uint8, fill_value=0)
+        target[:, :] = [220, 30, 30]  # rouge vif unique
+        # Ajouter un bord vert pour le rendre encore plus unique
+        target[:5, :] = [30, 220, 30]
+        target[-5:, :] = [30, 220, 30]
+        screen_np[200:280, 300:420] = target
+        screen = Image.fromarray(screen_np)
+
+        # L'ancre d'origine
+        anchor_original = Image.fromarray(target)
+        # L'ancre à 105% (scale modeste pour que ça reste réaliste)
+        w, h = anchor_original.size
+        scaled_anchor = anchor_original.resize((int(w * 1.05), int(h * 1.05)), Image.BILINEAR)
+
+        m_multi = TemplateMatcher(threshold=0.60, multiscale=True)
+        result_multi = m_multi.match_screen(anchor_pil=scaled_anchor, screen_pil=screen)
+        assert result_multi is not None
+        assert result_multi.method == 'template_multiscale'
+
+    def test_multiscale_anchor_too_small(self):
+        """Ancre très petite — certaines échelles sont sautées."""
+        screen = _make_image(800, 600)
+        anchor = _make_image(5, 5, (255, 0, 0))
+        m = TemplateMatcher(threshold=0.99, multiscale=True, scales=[0.5, 0.3])
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        # Pas de crash même avec des échelles qui produisent < 8px
+        # Le résultat peut être None ou un match selon le contenu
+
+
+# ---------------------------------------------------------------------------
+# Tests TemplateMatcher — match_in_region
+# ---------------------------------------------------------------------------
+
+class TestMatchInRegion:
+    def test_region_match(self):
+        # Créer une region BGR bruitée avec un motif damier injecté
+        rng = np.random.RandomState(77)
+        region = rng.randint(0, 256, (200, 300, 3), dtype=np.uint8)
+        # Motif damier en BGR
+        anchor = np.zeros((40, 60, 3), dtype=np.uint8)
+        for r in range(40):
+            for c in range(60):
+                if (r + c) % 2 == 0:
+                    anchor[r, c] = [255, 0, 0]
+                else:
+                    anchor[r, c] = [0, 0, 255]
+        region[50:90, 100:160] = anchor
+
+        m = TemplateMatcher(threshold=0.75)
+        result = m.match_in_region(region, anchor)
+        assert result is not None
+        assert abs(result.x - 130) <= 1  # 100 + 60//2
+        assert abs(result.y - 70) <= 1   # 50 + 40//2
+
+    def test_region_no_match(self):
+        # Region bruitée, ancre damier absente
+        rng = np.random.RandomState(88)
+        region = rng.randint(0, 256, (200, 300, 3), dtype=np.uint8)
+        anchor = np.zeros((40, 60, 3), dtype=np.uint8)
+        for r in range(40):
+            for c in range(60):
+                anchor[r, c] = [255, 255, 0] if (r + c) % 2 == 0 else [0, 255, 255]
+
+        m = TemplateMatcher(threshold=0.75)
+        result = m.match_in_region(region, anchor)
+        assert result is None
+
+
+# ---------------------------------------------------------------------------
+# Tests grayscale mode
+# ---------------------------------------------------------------------------
+
+class TestGrayscale:
+    def test_grayscale_match(self):
+        screen, anchor, cx, cy = _make_screen_with_target()
+        m = TemplateMatcher(threshold=0.75, grayscale=True)
+        result = m.match_screen(anchor_pil=anchor, screen_pil=screen)
+        assert result is not None
+        assert abs(result.x - cx) <= 1
+
+
+# ---------------------------------------------------------------------------
+# Tests _capture_screen (mocké)
+# ---------------------------------------------------------------------------
+
+class TestCaptureScreen:
+    @patch('core.grounding.template_matcher._MSS', False)
+    def test_no_mss(self):
+        result = TemplateMatcher._capture_screen()
+        assert result is None

tools/benchmark_grounding.py

@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+"""
+Benchmark complet des méthodes de grounding visuel.
+À lancer avec la VM Windows visible à l'écran, bureau avec dossier Demo.
+
+Usage:
+    cd ~/ai/rpa_vision_v3
+    .venv/bin/python3 tools/benchmark_grounding.py
+"""
+import mss, io, base64, requests, time, re, cv2, numpy as np, os, glob, json
+from PIL import Image
+
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+ANCHOR_DIR = 'visual_workflow_builder/backend/data/anchors'
+
+
+def capture_screen():
+    with mss.mss() as sct:
+        grab = sct.grab(sct.monitors[0])
+        screen = Image.frombytes('RGB', grab.size, grab.rgb)
+    return screen
+
+
+def screen_to_b64(screen):
+    buf = io.BytesIO()
+    screen.save(buf, format='JPEG', quality=70)
+    return base64.b64encode(buf.getvalue()).decode()
+
+
+def parse_coords(text, screen_w, screen_h):
+    for pat in [
+        r"start_box='?\<?\|?box_start\|?\>?\((\d+),(\d+)\)",
+        r'\((\d+(?:\.\d+)?)\s*,\s*(\d+(?:\.\d+)?)\)',
+        r'\[(\d+(?:\.\d+)?)\s*,\s*(\d+(?:\.\d+)?)\]',
+    ]:
+        m = re.search(pat, text)
+        if m:
+            rx, ry = float(m.group(1)), float(m.group(2))
+            if rx <= 1.0 and ry <= 1.0:
+                return int(rx * screen_w), int(ry * screen_h)
+            elif rx <= 1000 and ry <= 1000:
+                return int(rx * screen_w / 1000), int(ry * screen_h / 1000)
+            return int(rx), int(ry)
+    return None
+
+
+def test_vlm(model, prompt, b64, screen_w, screen_h):
+    t0 = time.time()
+    try:
+        resp = requests.post(f'{OLLAMA_URL}/api/generate', json={
+            'model': model, 'prompt': prompt, 'images': [b64],
+            'stream': False, 'options': {'temperature': 0.0, 'num_predict': 50}
+        }, timeout=60)
+        elapsed = time.time() - t0
+        if resp.status_code != 200:
+            return elapsed, None, f"HTTP {resp.status_code}"
+        text = resp.json().get('response', '').strip()
+        coords = parse_coords(text, screen_w, screen_h)
+        return elapsed, coords, text[:120]
+    except Exception as e:
+        return time.time() - t0, None, str(e)[:80]
+
+
+def test_template(screen_gray, anchor_path):
+    anchor = cv2.imread(anchor_path, cv2.IMREAD_GRAYSCALE)
+    if anchor is None:
+        return None
+    ah, aw = anchor.shape[:2]
+    if ah >= screen_gray.shape[0] or aw >= screen_gray.shape[1]:
+        return None
+    t0 = time.time()
+    result = cv2.matchTemplate(screen_gray, anchor, cv2.TM_CCOEFF_NORMED)
+    _, max_val, _, max_loc = cv2.minMaxLoc(result)
+    elapsed = (time.time() - t0) * 1000
+    return {
+        'method': 'template', 'time_ms': elapsed,
+        'score': max_val, 'pos': (max_loc[0] + aw//2, max_loc[1] + ah//2)
+    }
+
+
+def test_template_multiscale(screen_gray, anchor_path, scales=(0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3)):
+    anchor = cv2.imread(anchor_path, cv2.IMREAD_GRAYSCALE)
+    if anchor is None:
+        return None
+    ah, aw = anchor.shape[:2]
+    t0 = time.time()
+    best_val, best_loc, best_scale = 0, None, 1.0
+    for s in scales:
+        resized = cv2.resize(anchor, None, fx=s, fy=s)
+        rh, rw = resized.shape[:2]
+        if rh >= screen_gray.shape[0] or rw >= screen_gray.shape[1]:
+            continue
+        res = cv2.matchTemplate(screen_gray, resized, cv2.TM_CCOEFF_NORMED)
+        _, mv, _, ml = cv2.minMaxLoc(res)
+        if mv > best_val:
+            best_val, best_loc, best_scale = mv, ml, s
+    elapsed = (time.time() - t0) * 1000
+    if best_loc is None:
+        return None
+    rh, rw = int(ah * best_scale), int(aw * best_scale)
+    return {
+        'method': 'template_multiscale', 'time_ms': elapsed,
+        'score': best_val, 'pos': (best_loc[0] + rw//2, best_loc[1] + rh//2),
+        'scale': best_scale
+    }
+
+
+def test_orb(screen_gray, anchor_path, max_distance=50):
+    anchor = cv2.imread(anchor_path, cv2.IMREAD_GRAYSCALE)
+    if anchor is None:
+        return None
+    t0 = time.time()
+    orb = cv2.ORB_create(nfeatures=1000)
+    kp1, des1 = orb.detectAndCompute(anchor, None)
+    kp2, des2 = orb.detectAndCompute(screen_gray, None)
+    if des1 is None or des2 is None or len(des1) < 2 or len(des2) < 2:
+        return {'method': 'ORB', 'time_ms': (time.time()-t0)*1000, 'matches': 0, 'pos': None}
+    bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
+    matches = bf.match(des1, des2)
+    good = sorted([m for m in matches if m.distance < max_distance], key=lambda m: m.distance)
+    elapsed = (time.time() - t0) * 1000
+    pos = None
+    if len(good) >= 4:
+        pts = np.float32([kp2[m.trainIdx].pt for m in good])
+        pos = (int(np.median(pts[:, 0])), int(np.median(pts[:, 1])))
+    return {'method': 'ORB', 'time_ms': elapsed, 'matches': len(good), 'pos': pos}
+
+
+def test_akaze(screen_gray, anchor_path, max_distance=80):
+    anchor = cv2.imread(anchor_path, cv2.IMREAD_GRAYSCALE)
+    if anchor is None:
+        return None
+    t0 = time.time()
+    akaze = cv2.AKAZE_create()
+    kp1, des1 = akaze.detectAndCompute(anchor, None)
+    kp2, des2 = akaze.detectAndCompute(screen_gray, None)
+    if des1 is None or des2 is None or len(des1) < 2 or len(des2) < 2:
+        return {'method': 'AKAZE', 'time_ms': (time.time()-t0)*1000, 'matches': 0, 'pos': None}
+    bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
+    matches = bf.match(des1, des2)
+    good = sorted([m for m in matches if m.distance < max_distance], key=lambda m: m.distance)
+    elapsed = (time.time() - t0) * 1000
+    pos = None
+    if len(good) >= 4:
+        pts = np.float32([kp2[m.trainIdx].pt for m in good])
+        pos = (int(np.median(pts[:, 0])), int(np.median(pts[:, 1])))
+    return {'method': 'AKAZE', 'time_ms': elapsed, 'matches': len(good), 'pos': pos}
+
+
+def main():
+    print("="*70)
+    print("BENCHMARK GROUNDING — Léa RPA Vision")
+    print("="*70)
+
+    screen = capture_screen()
+    screen_w, screen_h = screen.size
+    b64 = screen_to_b64(screen)
+    screen_cv = cv2.cvtColor(np.array(screen), cv2.COLOR_RGB2BGR)
+    screen_gray = cv2.cvtColor(screen_cv, cv2.COLOR_BGR2GRAY)
+    print(f"Écran: {screen_w}x{screen_h}\n")
+
+    # ── VLM grounding ──
+    print("─── VLM GROUNDING (cible: 'Demo folder') ───")
+    vlm_tests = [
+        ("qwen3-vl:8b", 'Click on "Demo folder". Return the action in format: click(start_box="(x,y)") with coordinates normalized 0-1000.'),
+        ("qwen2.5vl:7b", 'Click on "Demo folder". Return the action in format: click(start_box="(x,y)") with coordinates normalized 0-1000.'),
+        ("moondream:latest", 'Where is the "Demo" folder icon? Give coordinates as (x, y) in pixels.'),
+        ("gemma4:latest", 'Click on "Demo folder". Return the action in format: click(start_box="(x,y)") with coordinates normalized 0-1000.'),
+    ]
+    for model, prompt in vlm_tests:
+        elapsed, coords, text = test_vlm(model, prompt, b64, screen_w, screen_h)
+        coord_str = f"({coords[0]:4d}, {coords[1]:4d})" if coords else "       —       "
+        print(f"  {model:35s} {elapsed:5.1f}s  {coord_str}  {text[:60]}")
+
+    # ── OpenCV ──
+    print(f"\n─── OPENCV (ancres de {ANCHOR_DIR}) ───")
+    thumbs = sorted(glob.glob(f'{ANCHOR_DIR}/*_thumb.png'))[:5]
+    full_imgs = sorted(glob.glob(f'{ANCHOR_DIR}/*_full.png'))[:5]
+
+    for thumb_path in thumbs:
+        name = os.path.basename(thumb_path).replace('_thumb.png', '')[:30]
+        ah, aw = cv2.imread(thumb_path, cv2.IMREAD_GRAYSCALE).shape[:2] if cv2.imread(thumb_path) is not None else (0,0)
+        print(f"\n  Ancre: {name} ({aw}x{ah})")
+
+        r = test_template(screen_gray, thumb_path)
+        if r:
+            print(f"    Template:          {r['time_ms']:6.1f}ms  score={r['score']:.3f}  pos={r['pos']}")
+
+        r = test_template_multiscale(screen_gray, thumb_path)
+        if r:
+            print(f"    Template multi-s:  {r['time_ms']:6.1f}ms  score={r['score']:.3f}  pos={r['pos']}  scale={r['scale']}")
+
+        r = test_orb(screen_gray, thumb_path)
+        if r:
+            print(f"    ORB:               {r['time_ms']:6.1f}ms  matches={r['matches']:3d}  pos={r['pos']}")
+
+        r = test_akaze(screen_gray, thumb_path)
+        if r:
+            print(f"    AKAZE:             {r['time_ms']:6.1f}ms  matches={r['matches']:3d}  pos={r['pos']}")
+
+    # ── Résumé ──
+    print(f"\n{'='*70}")
+    print("RÉSUMÉ")
+    print("="*70)
+    print("""
+Pipeline recommandé (du plus rapide au plus lent) :
+  1. Template matching classique     ~20-50ms   (score > 0.75 = direct)
+  2. Template multi-scale            ~80-150ms  (robuste aux changements de taille)
+  3. OCR (docTR)                     ~500-1000ms (texte uniquement)
+  4. Static fallback                 ~0ms       (coordonnées d'origine)
+
+Note : les feature matchers (ORB/AKAZE) ne sont pas adaptés aux petites
+ancres UI (< 200x200px) — trop peu de keypoints distinctifs.
+""")
+
+
+if __name__ == '__main__':
+    main()

tools/start_grounding_server.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# Lancement du serveur de grounding UI-TARS (port 8200)
+#
+# Le serveur charge UI-TARS-1.5-7B en 4-bit NF4 dans son propre process
+# Python avec un contexte CUDA propre. Le backend Flask VWB et la boucle
+# ORA appellent ce serveur en HTTP.
+#
+# Usage :
+#   ./tools/start_grounding_server.sh          # premier plan
+#   ./tools/start_grounding_server.sh --bg     # arriere-plan (log dans /tmp)
+
+set -e
+
+cd /home/dom/ai/rpa_vision_v3
+
+VENV=".venv/bin/python3"
+LOG="/tmp/grounding_server.log"
+
+if [ ! -f "$VENV" ]; then
+    echo "ERREUR: venv non trouve a $VENV"
+    exit 1
+fi
+
+echo "=== Serveur de Grounding UI-TARS ==="
+echo "Port: 8200"
+echo "Modele: ByteDance-Seed/UI-TARS-1.5-7B (4-bit NF4)"
+echo ""
+
+if [ "$1" = "--bg" ]; then
+    echo "Lancement en arriere-plan (logs dans $LOG)"
+    nohup $VENV -m core.grounding.server > "$LOG" 2>&1 &
+    PID=$!
+    echo "PID: $PID"
+    echo "$PID" > /tmp/grounding_server.pid
+    echo "Verifier: curl http://localhost:8200/health"
+    echo "Logs: tail -f $LOG"
+else
+    $VENV -m core.grounding.server
+fi

visual_workflow_builder/backend/api_v3/dag_execute.py

@@ -868,6 +868,60 @@ def _load_anchor_metadata(anchor_id: str) -> Optional[Dict]:
     return None
 
 
+def _inject_anchor_targeting(action: Dict, anchor_id: str) -> None:
+    """Enrichit une action avec la cible visuelle (x_pct/y_pct + visual_mode/target_spec).
+
+    Mutation in-place de `action`. Utilisé pour click_anchor*, type_text et
+    type_secret — toute action qui doit cibler une zone visuelle précise avant
+    d'agir (clic ou frappe avec focus).
+
+    Sans cette injection, l'agent côté Windows ne peut pas faire le pre-click
+    de focus avant `_type_text`, et le texte tape dans le vide.
+    """
+    if not anchor_id:
+        return
+
+    anchor_meta = _load_anchor_metadata(anchor_id)
+
+    # Coordonnées du centre du bbox (fallback si template matching échoue)
+    if anchor_meta:
+        bbox = anchor_meta.get('bounding_box', {})
+        orig = anchor_meta.get('original_size', {})
+        orig_w = orig.get('width', 1920)
+        orig_h = orig.get('height', 1080)
+        if bbox.get('x') is not None and orig_w > 0 and orig_h > 0:
+            cx = (bbox['x'] + bbox.get('width', 0) / 2) / orig_w
+            cy = (bbox['y'] + bbox.get('height', 0) / 2) / orig_h
+            action['x_pct'] = round(cx, 4)
+            action['y_pct'] = round(cy, 4)
+
+    # Image de l'ancre pour template matching côté agent
+    anchor_b64 = _load_anchor_image_b64(anchor_id)
+    if anchor_b64:
+        target_spec = {
+            'anchor_image_base64': anchor_b64,
+            'anchor_id': anchor_id,
+        }
+        if anchor_meta:
+            target_spec['anchor_bbox'] = anchor_meta.get('bounding_box', {})
+            target_spec['original_size'] = anchor_meta.get('original_size', {})
+
+        action['visual_mode'] = True
+        action['target_spec'] = target_spec
+        logger.info(
+            "Action %s : ancre '%s' chargée (%d Ko), visual_mode activé",
+            action.get('action_id', '?'),
+            anchor_id,
+            len(anchor_b64) // 1024,
+        )
+    else:
+        logger.warning(
+            "Action %s : ancre '%s' introuvable, fallback blind mode",
+            action.get('action_id', '?'),
+            anchor_id,
+        )
+
+
 @api_v3_bp.route('/execute-windows', methods=['POST'])
 def execute_windows():
     """Proxy les actions du workflow vers le streaming server pour exécution sur Windows.
@@ -932,45 +986,14 @@ def execute_windows():
             if vwb_type in _ANCHOR_CLICK_TYPES:
                 anchor_id = action.get('anchor_id')
                 if anchor_id:
-                    anchor_meta = _load_anchor_metadata(anchor_id)
+                    _inject_anchor_targeting(action, anchor_id)
 
-                    # Calculer les coordonnées du centre du bbox (fallback si visual échoue)
+                # Propagation du by_text (ciblage textuel prioritaire sur template)
-                    if anchor_meta:
+                _by_text = params.get('by_text', '')
-                        bbox = anchor_meta.get('bounding_box', {})
+                if _by_text:
-                        orig = anchor_meta.get('original_size', {})
+                    action['by_text'] = _by_text
-                        orig_w = orig.get('width', 1920)
+                    if 'target_spec' in action:
-                        orig_h = orig.get('height', 1080)
+                        action['target_spec']['by_text'] = _by_text
-                        if bbox.get('x') is not None and orig_w > 0 and orig_h > 0:
-                            cx = (bbox['x'] + bbox.get('width', 0) / 2) / orig_w
-                            cy = (bbox['y'] + bbox.get('height', 0) / 2) / orig_h
-                            action['x_pct'] = round(cx, 4)
-                            action['y_pct'] = round(cy, 4)
-
-                    # Tenter aussi le visual_mode (template matching)
-                    anchor_b64 = _load_anchor_image_b64(anchor_id)
-                    if anchor_b64:
-                        target_spec = {
-                            'anchor_image_base64': anchor_b64,
-                            'anchor_id': anchor_id,
-                        }
-                        if anchor_meta:
-                            target_spec['anchor_bbox'] = anchor_meta.get('bounding_box', {})
-                            target_spec['original_size'] = anchor_meta.get('original_size', {})
-
-                        action['visual_mode'] = True
-                        action['target_spec'] = target_spec
-                        logger.info(
-                            "Action %s : ancre '%s' chargée (%d Ko), visual_mode activé",
-                            action.get('action_id', '?'),
-                            anchor_id,
-                            len(anchor_b64) // 1024,
-                        )
-                    else:
-                        logger.warning(
-                            "Action %s : ancre '%s' introuvable, fallback blind mode",
-                            action.get('action_id', '?'),
-                            anchor_id,
-                        )
 
                 # Mapper le bouton selon le type de clic VWB
                 if vwb_type == 'double_click_anchor':
@@ -979,13 +1002,18 @@ def execute_windows():
                     action['button'] = 'right'
 
             # ---------------------------------------------------------------
-            # type_text / type_secret → extraire le texte
+            # type_text / type_secret → extraire le texte + cibler la zone
+            # de saisie si une ancre visuelle est associée au step.
+            # Sans ancre, l'agent tape là où le focus se trouve déjà
+            # (compatibilité avec les workflows historiques sans anchor).
             # ---------------------------------------------------------------
             if vwb_type in ('type_text', 'type_secret') and 'text' in params:
                 action['text'] = params['text']
-                # Ne pas forcer un clic préalable à (0,0) si pas de coordonnées
+                anchor_id = action.get('anchor_id') or (
-                # L'exécuteur ne cliquera que si x_pct > 0 et y_pct > 0
+                    params.get('visual_anchor') or {}
-                # (le clic de positionnement est fait par l'action click_anchor précédente)
+                ).get('anchor_id')
+                if anchor_id:
+                    _inject_anchor_targeting(action, anchor_id)
 
             # ---------------------------------------------------------------
             # keyboard_shortcut / hotkey → extraire les touches
@@ -1043,11 +1071,26 @@ def execute_windows():
             # Sinon, retirer les actions fichiers du flux principal
             data['actions'] = non_file_actions
 
+    # Token Bearer pour le streaming server (auth obligatoire)
+    _stream_token = os.environ.get('RPA_API_TOKEN', '')
+    _stream_headers = {'Authorization': f'Bearer {_stream_token}'} if _stream_token else {}
+
+    # L'agent Windows poll sous session "agent_demo_user" (= agent_{user_id}, user_id="demo_user")
+    # On injecte directement dans cette session pour éviter le transfer cross-session
+    # et pour que /replay/raw ne tente pas l'auto-détection d'une session "sess_*"
+    # (qui échoue avec "Aucune session Agent V1 active" si l'agent n'a pas créé de session V1).
+    if not data.get('session_id'):
+        data['session_id'] = 'agent_demo_user'
+
     # Injecter le machine_id pour le ciblage multi-machine
     # Chercher la première machine Windows connectée si pas spécifié
     if 'machine_id' not in data or not data.get('machine_id'):
         try:
-            machines_resp = req.get('http://localhost:5005/api/v1/traces/stream/machines', timeout=3)
+            machines_resp = req.get(
+                'http://localhost:5005/api/v1/traces/stream/machines',
+                headers=_stream_headers,
+                timeout=3,
+            )
             if machines_resp.ok:
                 machines = machines_resp.json().get('machines', [])
                 for m in machines:
@@ -1062,6 +1105,7 @@ def execute_windows():
         resp = req.post(
             'http://localhost:5005/api/v1/traces/stream/replay/raw',
             json=data,
+            headers=_stream_headers,
             timeout=30,
         )
         return jsonify(resp.json()), resp.status_code
@@ -1069,3 +1113,76 @@ def execute_windows():
         return jsonify({'error': 'Streaming server (port 5005) non disponible'}), 503
     except Exception as e:
         return jsonify({'error': str(e)}), 500
+
+
+# ---------------------------------------------------------------------------
+# QW4 — Proxy /api/v3/replay/resume → streaming /replay/{id}/resume
+# Forward Bearer token + body { replay_id, acknowledged_check_ids }.
+# Le frontend (PauseDialog) appelle /api/v3/replay/resume via le VWB ;
+# on relaye au streaming server pour valider les acquittements safety_checks.
+# ---------------------------------------------------------------------------
+@api_v3_bp.route('/replay/resume', methods=['POST'])
+def replay_resume_proxy():
+    """Proxy QW4 vers le serveur streaming pour la reprise avec safety_checks."""
+    import requests as req
+
+    data = request.get_json() or {}
+    replay_id = data.get('replay_id')
+    if not replay_id:
+        return jsonify({'error': 'replay_id manquant'}), 400
+
+    streaming_url = os.environ.get('RPA_STREAMING_URL', 'http://localhost:5005')
+    token = os.environ.get('RPA_API_TOKEN', '')
+    headers = {'Content-Type': 'application/json'}
+    if token:
+        headers['Authorization'] = f'Bearer {token}'
+
+    # Body forwardé : uniquement acknowledged_check_ids (replay_id est dans l'URL)
+    forward_body = {
+        'acknowledged_check_ids': data.get('acknowledged_check_ids') or [],
+    }
+
+    try:
+        resp = req.post(
+            f'{streaming_url}/api/v1/traces/stream/replay/{replay_id}/resume',
+            json=forward_body,
+            headers=headers,
+            timeout=10,
+        )
+        return resp.content, resp.status_code, {'Content-Type': 'application/json'}
+    except req.ConnectionError:
+        return jsonify({'error': 'streaming_unreachable',
+                        'detail': f'Streaming server non disponible ({streaming_url})'}), 502
+    except req.RequestException as e:
+        return jsonify({'error': 'streaming_unreachable', 'detail': str(e)}), 502
+
+
+# ---------------------------------------------------------------------------
+# QW4 — Proxy GET /api/v3/replay/state/<replay_id> → streaming /replay/{id}
+# Forward Bearer token vers le serveur streaming.
+# Permet à App.tsx de récupérer le state du replay actif (Agent V1 Windows)
+# pour afficher PauseDialog quand status = paused_need_help avec safety_checks.
+# ---------------------------------------------------------------------------
+@api_v3_bp.route('/replay/state/<replay_id>', methods=['GET'])
+def replay_state_proxy(replay_id):
+    """Proxy QW4 vers le serveur streaming pour récupérer le state replay actif."""
+    import requests as req
+
+    streaming_url = os.environ.get('RPA_STREAMING_URL', 'http://localhost:5005')
+    token = os.environ.get('RPA_API_TOKEN', '')
+    headers = {}
+    if token:
+        headers['Authorization'] = f'Bearer {token}'
+
+    try:
+        resp = req.get(
+            f'{streaming_url}/api/v1/traces/stream/replay/{replay_id}',
+            headers=headers,
+            timeout=5,
+        )
+        return resp.content, resp.status_code, {'Content-Type': 'application/json'}
+    except req.ConnectionError:
+        return jsonify({'error': 'streaming_unreachable',
+                        'detail': f'Streaming server non disponible ({streaming_url})'}), 502
+    except req.RequestException as e:
+        return jsonify({'error': 'streaming_unreachable', 'detail': str(e)}), 502

visual_workflow_builder/backend/api_v3/execute.py

@@ -896,15 +896,15 @@ def execute_action(action_type: str, params: dict) -> dict:
                         _fc_target_text = params.get('_step_label', '')
                     _action_types = {'click_anchor', 'double_click_anchor', 'right_click_anchor',
                                      'hover_anchor', 'focus_anchor', 'scroll_to_anchor'}
-                    if _fc_target_text in _action_types and screenshot_base64:
+                    # Note: plus d'appel à _describe_anchor_image() (qwen2.5vl) ici.
-                        try:
+                    # Le crop d'ancre (screenshot_base64) est utilisé directement par
-                            from core.execution.input_handler import _describe_anchor_image
+                    # le template matching pixel-perfect en avant-poste, puis InfiGUI
-                            _desc = _describe_anchor_image(screenshot_base64)
+                    # en mode fusionné si nécessaire (option 2.c+2.a). Économise ~9.4 GB
-                            if _desc:
+                    # de VRAM Ollama qui rentrait en conflit avec InfiGUI.
-                                print(f"🏷️ [Vision] Ancre décrite: '{_desc}'")
+                    if _fc_target_text in _action_types:
-                                _fc_target_text = _desc
+                        # Marquer le label comme garbage pour que le pipeline
-                        except Exception:
+                        # bascule sur le mode fusionné via template_b64.
-                            pass
+                        _fc_target_text = ''
                     _fc_target_desc = params.get('visual_anchor', {}).get('description', '')
 
                     x, y, confidence, method_used = None, None, 0, ''
@@ -1431,7 +1431,7 @@ def run_workflow_verified(execution_id: str, workflow_id: str, app):
             from core.execution.observe_reason_act import ORALoop
 
             ora = ORALoop(
-                max_retries=2, max_steps=50, verify_level='auto',
+                max_retries=2, max_steps=50, verify_level='none',
                 should_continue=lambda: not _execution_state.get('should_stop', False)
             )
             ora._variables = _execution_state.get('variables', {})

visual_workflow_builder/backend/api_v3/learned_workflows.py

@@ -40,6 +40,17 @@ if _ROOT not in sys.path:
 STREAMING_SERVER_URL = "http://localhost:5005"
 
 
+def _stream_headers() -> Dict[str, str]:
+    """Bearer token pour les appels proxy VWB → streaming server.
+
+    Retourne un dict vide si RPA_API_TOKEN n'est pas défini ; dans ce cas
+    les appels échoueront en 401 (auth obligatoire côté streaming).
+    """
+    import os as _os
+    token = _os.environ.get("RPA_API_TOKEN", "")
+    return {"Authorization": f"Bearer {token}"} if token else {}
+
+
 # ---------------------------------------------------------------------------
 # Helpers — nom par défaut à l'import
 # ---------------------------------------------------------------------------
@@ -162,6 +173,7 @@ def list_learned_workflows():
         resp = http_requests.get(
             f"{STREAMING_SERVER_URL}/api/v1/traces/stream/workflows",
             params=params,
+            headers=_stream_headers(),
             timeout=3,
         )
         if resp.ok:
@@ -526,6 +538,7 @@ def _load_core_workflow(
         resp = http_requests.get(
             f"{STREAMING_SERVER_URL}/api/v1/traces/stream/workflows",
             params=params,
+            headers=_stream_headers(),
             timeout=3,
         )
         if resp.ok:
@@ -538,6 +551,7 @@ def _load_core_workflow(
                     try:
                         detail_resp = http_requests.get(
                             f"{STREAMING_SERVER_URL}/api/v1/traces/stream/workflow/{workflow_id}",
+                            headers=_stream_headers(),
                             timeout=5,
                         )
                         if detail_resp.ok:
@@ -573,6 +587,7 @@ def _notify_streaming_reload():
     try:
         http_requests.post(
             f"{STREAMING_SERVER_URL}/api/v1/traces/stream/reload-workflows",
+            headers=_stream_headers(),
             timeout=2,
         )
         logger.debug("Streaming server notifié pour rechargement des workflows")

visual_workflow_builder/backend/app.py

@@ -13,11 +13,17 @@ from flask_caching import Cache
 from flask_migrate import Migrate
 import os
 import logging
+from pathlib import Path
 from logging.handlers import RotatingFileHandler
 from dotenv import load_dotenv
 
-# Load environment variables
+# Charger .env.local depuis la racine du projet AVANT tout : il contient
-load_dotenv()
+# RPA_API_TOKEN utilisé pour le proxy VWB → streaming server. Sans cela,
+# le token est absent après chaque restart manuel du backend et tous les
+# appels proxy renvoient 401 « Token API invalide ».
+_PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent
+load_dotenv(_PROJECT_ROOT / '.env.local')
+load_dotenv()  # fallback .env dans cwd (n'écrase pas les vars déjà définies)
 
 # Initialize Flask app
 app = Flask(__name__)

visual_workflow_builder/backend/services/learned_workflow_bridge.py

@@ -57,7 +57,9 @@ VWB_ACTION_TO_CORE = {
     "scroll_to_anchor": "scroll",
     "visual_condition": "evaluate_condition",
     "screenshot_evidence": "screenshot",
-    "extract_text": "extract_data",
+    "extract_text": "extract_text",         # passthrough — handler serveur OCR + variable
+    "pause_for_human": "pause_for_human",   # passthrough — intercepté par api_stream /replay/next
+    "t2a_decision": "t2a_decision",         # passthrough — handler serveur LLM T2A + variable
 }
 
 
@@ -660,6 +662,23 @@ def _vwb_params_to_core(action_type: str, params: Dict[str, Any]) -> Dict[str, A
     elif action_type == "wait_for_anchor":
         core_params["duration_ms"] = params.get("duration_ms", 2000)
 
+    elif action_type == "pause_for_human":
+        core_params["message"] = params.get("message", "Validation requise")
+
+    elif action_type == "extract_text":
+        # variable_name côté VWB → output_var côté core (compat avec
+        # le catalogue VWB existant qui utilise variable_name)
+        var = params.get("output_var") or params.get("variable_name") or "extracted_text"
+        core_params["output_var"] = var
+        if "paragraph" in params:
+            core_params["paragraph"] = bool(params["paragraph"])
+
+    elif action_type == "t2a_decision":
+        core_params["input_template"] = params.get("input_template", "")
+        core_params["output_var"] = params.get("output_var", "t2a_result")
+        if params.get("model"):
+            core_params["model"] = params["model"]
+
     return core_params
 
 

visual_workflow_builder/frontend_v4/src/App.tsx

@@ -25,6 +25,7 @@ import ExecutionOverlay from './components/ExecutionOverlay';
 import type { Variable } from './components/VariableManager';
 import RightPanel from './components/RightPanel';
 import SelfHealingDialog from './components/SelfHealingDialog';
+import PauseDialog from './components/PauseDialog';
 import ConfidenceDashboard from './components/ConfidenceDashboard';
 import WorkflowValidation from './components/WorkflowValidation';
 import ReviewModal from './components/ReviewModal';
@@ -61,6 +62,13 @@ function App() {
   const [healingCandidates, setHealingCandidates] = useState<any[]>([]);
   const [healingStepInfo, setHealingStepInfo] = useState<any>(null);
 
+  // QW4 — Replay streaming Windows en cours (Agent V1 distant).
+  // Quand un replay distant est lancé via ExecutionControls "→ Windows",
+  // ExecutionControls appelle setStreamingReplayId(replay_id) et un useEffect
+  // poll /api/v3/replay/state/<id> pour fusionner safety_checks + pause_*
+  // dans appState.execution → PauseDialog s'affiche.
+  const [streamingReplayId, setStreamingReplayId] = useState<string | null>(null);
+
   // Charger l'état initial
   const loadState = useCallback(async () => {
     try {
@@ -122,6 +130,62 @@ function App() {
     return () => clearInterval(interval);
   }, [isExecutionRunning, loadState]);
 
+  // QW4 — Polling state replay streaming (Agent V1 Windows distant)
+  // Tourne dès qu'un replay distant a été lancé. Récupère safety_checks,
+  // pause_message, pause_reason et les fusionne dans appState.execution
+  // pour que PauseDialog s'affiche quand status = paused_need_help.
+  useEffect(() => {
+    if (!streamingReplayId) return;
+
+    let stopped = false;
+    const pollReplay = async () => {
+      try {
+        const resp = await fetch(`/api/v3/replay/state/${streamingReplayId}`);
+        if (!resp.ok) return;
+        const state = await resp.json();
+        if (stopped) return;
+
+        // Fusionner dans appState.execution sans écraser le reste.
+        setAppState(prev => {
+          if (!prev) return prev;
+          const prevExec = prev.execution || {
+            id: streamingReplayId,
+            workflow_id: prev.session?.active_workflow_id || '',
+            status: 'pending',
+            progress: 0,
+            current_step_index: 0,
+            completed_steps: 0,
+            failed_steps: 0,
+            total_steps: 0,
+          };
+          return {
+            ...prev,
+            execution: {
+              ...prevExec,
+              status: state.status || prevExec.status,
+              pause_message: state.pause_message || state.message,
+              pause_reason: state.pause_reason,
+              safety_checks: state.safety_checks || [],
+              replay_id: streamingReplayId,
+            },
+          };
+        });
+
+        // Stopper le polling si le replay est terminé / annulé.
+        if (state.status && ['completed', 'error', 'cancelled'].includes(state.status)) {
+          setStreamingReplayId(null);
+        }
+      } catch (err) {
+        // ignore (le serveur streaming peut être momentanément indispo)
+      }
+    };
+
+    // Tick immédiat puis toutes les 1s.
+    pollReplay();
+    const interval = setInterval(pollReplay, 1000);
+    return () => { stopped = true; clearInterval(interval); };
+  }, [streamingReplayId]);
+
   // Convertir les étapes en nœuds React Flow
   // Les edges ne sont générées automatiquement que lors du premier chargement
   // d'un workflow. Ensuite, les connexions manuelles de l'utilisateur sont préservées.
@@ -451,6 +515,7 @@ function App() {
           execution={appState?.execution || null}
           onStart={handleStartExecution}
           onStop={handleStopExecution}
+          onWindowsReplayStarted={(replayId) => setStreamingReplayId(replayId)}
         />
         <ConfidenceDashboard
           isExecutionRunning={isExecutionRunning}
@@ -569,6 +634,47 @@ function App() {
         }}
       />
 
+      {/* QW4 — Pause supervisée (safety_checks).
+          Affiché si le serveur renvoie status == paused_need_help, ou
+          status == paused avec un payload de checks. Backward 100% : si
+          safety_checks vide, PauseDialog rend la bulle simple legacy. */}
+      {(appState?.execution?.status === 'paused_need_help' ||
+        (appState?.execution?.status === 'paused' &&
+         (appState?.execution?.safety_checks?.length ?? 0) > 0)) && (
+        <div className="pause-dialog-overlay">
+          <PauseDialog
+            pauseMessage={appState.execution.pause_message || 'Validation requise'}
+            pauseReason={appState.execution.pause_reason}
+            safetyChecks={appState.execution.safety_checks || []}
+            onResume={async (ackIds) => {
+              const replayId = appState.execution?.replay_id || appState.execution?.id;
+              if (replayId) {
+                // Voie streaming server (Agent V1 / replay distant)
+                const resp = await fetch('/api/v3/replay/resume', {
+                  method: 'POST',
+                  headers: { 'Content-Type': 'application/json' },
+                  body: JSON.stringify({
+                    replay_id: replayId,
+                    acknowledged_check_ids: ackIds,
+                  }),
+                });
+                if (!resp.ok) {
+                  const err = await resp.json().catch(() => ({}));
+                  throw new Error(err?.detail?.error || resp.statusText);
+                }
+              } else {
+                // Voie locale (execute/resume)
+                await api.resumeExecution();
+              }
+              await loadState();
+            }}
+            onCancel={() => {
+              handleStopExecution();
+            }}
+          />
+        </div>
+      )}
+
       {/* ConfidenceDashboard déplacé dans le header */}
     </div>
   );

visual_workflow_builder/frontend_v4/src/components/CapturePanel.tsx

@@ -4,7 +4,8 @@ import type { UIElement } from '../services/uiDetection';
 import {
   loadLibraryAsync,
   saveLibrary,
-  compressThumbnail,
+  addCaptureToLibrary,
+  removeCaptureFromLibrary,
 } from '../services/captureLibraryStorage';
 
 /**
@@ -40,6 +41,8 @@ interface LibraryItem {
   timestamp: Date;
   sessionId?: string;
   favorite?: boolean;
+  format?: 'v2';
+  fullImageUrl?: string;
 }
 
 export default function CapturePanel({
@@ -55,7 +58,7 @@ export default function CapturePanel({
   const [showLibraryGallery, setShowLibraryGallery] = useState(false);
   const [library, setLibrary] = useState<LibraryItem[]>([]);
   const [currentCapture, setCurrentCapture] = useState<Capture | null>(null);
-  const [timerSeconds, setTimerSeconds] = useState(0);
+  const [timerSeconds, setTimerSeconds] = useState(5);
   const [countdown, setCountdown] = useState<number | null>(null);
   // Elements detectes sur l'apercu miniature
   const [previewElements, setPreviewElements] = useState<UIElement[]>([]);
@@ -89,24 +92,35 @@ export default function CapturePanel({
     }
   }, [library, libraryLoaded]);
 
-  // Ajouter capture a la bibliotheque (thumbnail compresse JPEG 320x240)
+  // Helper : ajoute une capture à la bibliothèque (PNG HD upload backend +
+  // mise à jour de l'état local). Utilisé par le useEffect [capture] et par
+  // doSmartCapture (capture locale Windows qui ne passe pas par la prop parente).
+  const addToLibrary = useCallback(async (cap: Capture) => {
+    try {
+      const item = await addCaptureToLibrary(cap, { id: `cap_${Date.now()}` });
+      setLibrary(prev => [
+        {
+          id: item.id,
+          capture: item.capture,
+          timestamp: typeof item.timestamp === 'string' ? new Date(item.timestamp) : item.timestamp,
+          sessionId: item.sessionId,
+          favorite: item.favorite ?? false,
+          format: item.format,
+          fullImageUrl: item.fullImageUrl,
+        },
+        ...prev.slice(0, 19),
+      ]);
+    } catch (e) {
+      console.warn('[CapturePanel] Échec ajout bibliothèque', e);
+    }
+  }, []);
+
+  // Capture venant du parent (path "fallback local" via prop capture)
   useEffect(() => {
     if (!capture) return;
     setCurrentCapture(capture);
-    let cancelled = false;
+    void addToLibrary(capture);
-    (async () => {
+  }, [capture, addToLibrary]);
-      const compressed = await compressThumbnail(capture.screenshot_base64);
-      if (cancelled) return;
-      const newItem: LibraryItem = {
-        id: `cap_${Date.now()}`,
-        capture: { ...capture, screenshot_base64: compressed },
-        timestamp: new Date(),
-        favorite: false,
-      };
-      setLibrary(prev => [newItem, ...prev.slice(0, 19)]);
-    })();
-    return () => { cancelled = true; };
-  }, [capture]);
 
   // Detecter les elements UI quand une capture arrive
   useEffect(() => {
@@ -152,19 +166,24 @@ export default function CapturePanel({
       const resp = await fetch('/api/screen-capture/capture-windows', { method: 'POST' });
       const data = await resp.json();
       if (resp.ok && data.image) {
-        setCurrentCapture({
+        const cap: Capture = {
           screenshot_base64: data.image,
           width: data.width,
           height: data.height,
           source: data.source || 'windows',
-        } as any);
+        } as any;
+        setCurrentCapture(cap);
+        // Ajouter à la bibliothèque (le useEffect [capture] ne tire pas
+        // ici car on ne passe pas par la prop parente)
+        void addToLibrary(cap);
         return;
       }
       console.warn('Agent Windows indisponible, fallback local:', data.error);
     } catch (err) {
       console.warn('Erreur capture Windows, fallback local:', err);
     }
-    // Fallback : capture locale (ecran du serveur Linux)
+    // Fallback : capture locale (ecran du serveur Linux) — passe par la prop
+    // parente, l'ajout se fera dans le useEffect [capture]
     onCapture();
   };
 
@@ -189,13 +208,44 @@ export default function CapturePanel({
     }, 1000);
   };
 
-  const handleLibrarySelect = (item: LibraryItem) => {
+  const handleLibrarySelect = async (item: LibraryItem) => {
-    setCurrentCapture(item.capture);
+    // Format v2 : remplacer le thumbnail par le PNG HD téléchargé du backend
+    // pour que la sélection d'ancre utilise une image non pixélisée.
+    if (item.format === 'v2' && item.fullImageUrl) {
+      try {
+        const resp = await fetch(item.fullImageUrl);
+        if (!resp.ok) throw new Error(`HTTP ${resp.status}`);
+        const blob = await resp.blob();
+        const base64 = await new Promise<string>((resolve, reject) => {
+          const reader = new FileReader();
+          reader.onload = () => {
+            const result = reader.result as string;
+            // FileReader → "data:image/png;base64,..." → on retire le préfixe
+            const idx = result.indexOf(',');
+            resolve(idx >= 0 ? result.slice(idx + 1) : result);
+          };
+          reader.onerror = () => reject(reader.error);
+          reader.readAsDataURL(blob);
+        });
+        setCurrentCapture({ ...item.capture, screenshot_base64: base64 });
+      } catch (e) {
+        console.warn('[CaptureLibrary] Échec chargement HD, fallback thumbnail', e);
+        setCurrentCapture(item.capture);
+      }
+    } else {
+      setCurrentCapture(item.capture);
+    }
     setIsFullscreen(true);
   };
 
   const handleDeleteLibraryItem = (id: string) => {
+    const target = library.find(it => it.id === id);
     setLibrary(prev => prev.filter(item => item.id !== id));
+    // v2 : supprimer aussi le PNG côté backend (le saveLibrary auto-déclenché
+    // par le useEffect ne nettoie que le JSON, pas les fichiers PNG orphelins).
+    if (target?.format === 'v2') {
+      void removeCaptureFromLibrary(id, true);
+    }
   };
 
   return (
@@ -204,17 +254,35 @@ export default function CapturePanel({
 
       {/* Capture — auto-detection OS navigateur */}
       <div className="capture-controls">
-        <button disabled={countdown !== null} onClick={doSmartCapture}>
+        <button disabled={countdown !== null} onClick={doSmartCapture} title="Capture immédiate (sans délai)">
           Capturer
         </button>
-        <select value={timerSeconds} onChange={(e) => setTimerSeconds(Number(e.target.value))}>
+        <label style={{ display: 'flex', alignItems: 'center', gap: 4, fontSize: 12 }}>
-          <option value="0">Immediat</option>
+          Délai :
-          <option value="3">3 sec</option>
+          <select
-          <option value="5">5 sec</option>
+            value={String(timerSeconds)}
-          <option value="10">10 sec</option>
+            onChange={(e) => {
-        </select>
+              const v = Number(e.target.value);
-        <button onClick={handleTimerCapture} disabled={countdown !== null}>
+              console.log('[CapturePanel] timerSeconds →', v);
-          {countdown !== null ? countdown : 'Timer'}
+              setTimerSeconds(v);
+            }}
+          >
+            <option value="0">Immediat</option>
+            <option value="3">3 sec</option>
+            <option value="5">5 sec</option>
+            <option value="10">10 sec</option>
+          </select>
+        </label>
+        <button
+          onClick={handleTimerCapture}
+          disabled={countdown !== null}
+          title={`Capture après ${timerSeconds}s — utile pour préparer l'écran avant la prise`}
+        >
+          {countdown !== null
+            ? `${countdown}…`
+            : timerSeconds === 0
+              ? 'Timer'
+              : `Capturer dans ${timerSeconds}s`}
         </button>
       </div>
 

visual_workflow_builder/frontend_v4/src/components/ExecutionControls.tsx

@@ -9,9 +9,12 @@ interface Props {
   execution: Execution | null;
   onStart: () => void;
   onStop: () => void;
+  // QW4 — Notifie App.tsx quand un replay streaming Windows est lancé,
+  // pour qu'il poll /api/v3/replay/state/<id> et affiche PauseDialog au besoin.
+  onWindowsReplayStarted?: (replayId: string) => void;
 }
 
-export default function ExecutionControls({ execution, onStart, onStop }: Props) {
+export default function ExecutionControls({ execution, onStart, onStop, onWindowsReplayStarted }: Props) {
   const isRunning = execution?.status === 'running' || execution?.status === 'paused';
   const [windowsStatus, setWindowsStatus] = useState<'idle' | 'sending' | 'sent' | 'error'>('idle');
 
@@ -56,6 +59,11 @@ export default function ExecutionControls({ execution, onStart, onStop }: Props)
       const result = await resp.json();
       if (result.replay_id) {
         setWindowsStatus('sent');
+        // QW4 — propage le replay_id à App.tsx pour activer le polling
+        // /api/v3/replay/state/<id> (PauseDialog si paused_need_help).
+        if (onWindowsReplayStarted) {
+          try { onWindowsReplayStarted(result.replay_id); } catch {}
+        }
         alert('Replay lancé ! Réduisez cette fenêtre maintenant.\nLes actions commenceront dans 5 secondes.');
         setTimeout(() => setWindowsStatus('idle'), 5000);
       } else {
@@ -75,9 +83,27 @@ export default function ExecutionControls({ execution, onStart, onStop }: Props)
       {!isRunning ? (
         <div style={{ display: 'flex', gap: '4px', alignItems: 'center' }}>
           {userOS === 'linux' ? (
-            <button className="btn-start" onClick={onStart} title="Exécuter sur cet écran">
+            <>
-              Exécuter
+              <button className="btn-start" onClick={onStart} title="Exécuter sur cet écran (Linux local)">
-            </button>
+                Exécuter
+              </button>
+              <button
+                className="btn-start"
+                onClick={handleExecuteWindows}
+                disabled={windowsStatus === 'sending'}
+                style={{
+                  background: windowsStatus === 'sent' ? '#22c55e' : windowsStatus === 'error' ? '#ef4444' : '#0078d4',
+                  fontSize: '12px',
+                  opacity: windowsStatus === 'sending' ? 0.6 : 1,
+                }}
+                title="Exécuter sur le PC Windows (Agent V1)"
+              >
+                {windowsStatus === 'sending' ? 'Envoi...' :
+                 windowsStatus === 'sent' ? 'Lancé !' :
+                 windowsStatus === 'error' ? 'Erreur' :
+                 '→ Windows'}
+              </button>
+            </>
           ) : (
             <button
               className="btn-start"

visual_workflow_builder/frontend_v4/src/components/PauseDialog.tsx

@@ -0,0 +1,126 @@
+// QW4 — PauseDialog : bulle de pause supervisée avec ChecklistPanel intégré.
+//
+// 2 modes de rendu :
+//   - safety_checks vide  -> bulle simple legacy (Continuer / Annuler)
+//   - safety_checks fournis -> ChecklistPanel ; bouton Continuer désactivé
+//                              tant qu'un check `required` n'est pas coché.
+//
+// Les checks `llm_contextual` portent un badge [Léa] avec evidence en tooltip.
+
+import { useState, useMemo } from 'react';
+import type { SafetyCheck } from '../types';
+
+interface Props {
+  pauseMessage: string;
+  pauseReason?: string;
+  safetyChecks: SafetyCheck[];
+  onResume: (acknowledgedIds: string[]) => Promise<void>;
+  onCancel: () => void;
+}
+
+export default function PauseDialog({
+  pauseMessage,
+  pauseReason,
+  safetyChecks,
+  onResume,
+  onCancel,
+}: Props) {
+  const [checked, setChecked] = useState<Record<string, boolean>>({});
+  const [submitting, setSubmitting] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  const allRequiredOK = useMemo(() => {
+    return safetyChecks
+      .filter((c) => c.required)
+      .every((c) => checked[c.id] === true);
+  }, [safetyChecks, checked]);
+
+  const toggle = (id: string) => {
+    setChecked((prev) => ({ ...prev, [id]: !prev[id] }));
+  };
+
+  const handleResume = async () => {
+    setSubmitting(true);
+    setError(null);
+    try {
+      const acknowledgedIds = Object.entries(checked)
+        .filter(([, v]) => v)
+        .map(([k]) => k);
+      await onResume(acknowledgedIds);
+    } catch (e: any) {
+      setError(e?.message || 'Erreur lors de la reprise');
+    } finally {
+      setSubmitting(false);
+    }
+  };
+
+  // Backward compat : pas de checks -> bulle simple legacy
+  if (safetyChecks.length === 0) {
+    return (
+      <div className="pause-dialog-simple">
+        <p>{pauseMessage}</p>
+        {pauseReason && <small className="pause-reason">Raison : {pauseReason}</small>}
+        <div className="pause-actions">
+          <button onClick={() => onResume([])} disabled={submitting}>
+            Continuer
+          </button>
+          <button onClick={onCancel} disabled={submitting}>
+            Annuler
+          </button>
+        </div>
+      </div>
+    );
+  }
+
+  return (
+    <div className="pause-dialog-checks">
+      <h3>Pause supervisée</h3>
+      <p className="pause-message">{pauseMessage}</p>
+      {pauseReason && (
+        <div className="pause-reason-banner">
+          <strong>Raison :</strong> {pauseReason}
+        </div>
+      )}
+
+      <ul className="checklist-panel">
+        {safetyChecks.map((c) => (
+          <li key={c.id} className={`check-item ${c.required ? 'required' : 'optional'}`}>
+            <label>
+              <input
+                type="checkbox"
+                checked={!!checked[c.id]}
+                onChange={() => toggle(c.id)}
+                disabled={submitting}
+              />
+              <span className="check-label">{c.label}</span>
+              {c.required && <span className="badge badge-required">obligatoire</span>}
+              {c.source === 'llm_contextual' && (
+                <span className="badge badge-lea" title={c.evidence || ''}>
+                  Léa
+                </span>
+              )}
+            </label>
+            {c.source === 'llm_contextual' && c.evidence && (
+              <small className="check-evidence">-&gt; {c.evidence}</small>
+            )}
+          </li>
+        ))}
+      </ul>
+
+      {error && <div className="pause-error">{error}</div>}
+
+      <div className="pause-actions">
+        <button
+          onClick={handleResume}
+          disabled={!allRequiredOK || submitting}
+          title={!allRequiredOK ? 'Coche tous les checks obligatoires' : 'Reprendre le replay'}
+        >
+          {submitting ? 'Reprise...' : 'Continuer'}
+        </button>
+        <button onClick={onCancel} disabled={submitting}>
+          Annuler
+        </button>
+      </div>
+    </div>
+  );
+}

visual_workflow_builder/frontend_v4/src/components/PropertiesPanel.tsx

@@ -1353,6 +1353,136 @@ export default function PropertiesPanel({ step, onUpdateParams, onDelete }: Prop
           </>
         );
 
+      case 'pause_for_human': {
+        // QW4 — éditeur safety_level + safety_checks (déclaratifs)
+        const safetyChecks = Array.isArray(params.safety_checks)
+          ? (params.safety_checks as Array<{ id?: string; label?: string; required?: boolean }>)
+          : [];
+        return (
+          <>
+            <div className="prop-field">
+              <label>Message affiché à l'opérateur</label>
+              <textarea
+                rows={4}
+                value={String(params.message || '')}
+                onChange={(e) => updateParam('message', e.target.value)}
+                placeholder="Ex: Décision : {{dec.decision}}&#10;&#10;{{dec.justification}}"
+                style={{ width: '100%', fontFamily: 'monospace', fontSize: '12px' }}
+              />
+            </div>
+
+            {/* QW4 — Niveau de sécurité */}
+            <div className="prop-field">
+              <label>Niveau de sécurité</label>
+              <select
+                value={String(params.safety_level || 'standard')}
+                onChange={(e) => updateParam('safety_level', e.target.value)}
+              >
+                <option value="standard">Standard (pas de LLM)</option>
+                <option value="medical_critical">Médical critique (LLM contextuel)</option>
+              </select>
+            </div>
+
+            {/* QW4 — Liste éditable de checks déclaratifs */}
+            <div className="prop-field">
+              <label>Checks à valider (déclaratifs)</label>
+              {safetyChecks.map((check, i) => (
+                <div key={i} className="check-editor-row">
+                  <input
+                    placeholder="ID (ex: check_ipp)"
+                    value={check.id || ''}
+                    style={{ width: '30%' }}
+                    onChange={(e) => {
+                      const next = [...safetyChecks];
+                      next[i] = { ...check, id: e.target.value };
+                      updateParam('safety_checks', next);
+                    }}
+                  />
+                  <input
+                    placeholder="Libellé"
+                    value={check.label || ''}
+                    style={{ flex: 1 }}
+                    onChange={(e) => {
+                      const next = [...safetyChecks];
+                      next[i] = { ...check, label: e.target.value };
+                      updateParam('safety_checks', next);
+                    }}
+                  />
+                  <label style={{ display: 'flex', alignItems: 'center', gap: '4px' }}>
+                    <input
+                      type="checkbox"
+                      checked={!!check.required}
+                      onChange={(e) => {
+                        const next = [...safetyChecks];
+                        next[i] = { ...check, required: e.target.checked };
+                        updateParam('safety_checks', next);
+                      }}
+                    />
+                    Obligatoire
+                  </label>
+                  <button
+                    type="button"
+                    onClick={() => {
+                      const next = safetyChecks.filter((_, j) => j !== i);
+                      updateParam('safety_checks', next);
+                    }}
+                    title="Supprimer ce check"
+                  >
+                    −
+                  </button>
+                </div>
+              ))}
+              <button
+                type="button"
+                onClick={() => {
+                  const next = [
+                    ...safetyChecks,
+                    { id: '', label: '', required: true },
+                  ];
+                  updateParam('safety_checks', next);
+                }}
+              >
+                + Ajouter un check
+              </button>
+            </div>
+          </>
+        );
+      }
+
+      case 't2a_decision':
+        return (
+          <>
+            <div className="prop-field">
+              <label>Template d'entrée (supporte {'{{var}}'})</label>
+              <textarea
+                rows={5}
+                value={String(params.input_template || '')}
+                onChange={(e) => updateParam('input_template', e.target.value)}
+                placeholder={'{{t0}}\n---\n{{t1}}\n{{t2}}\n{{t3}}\n{{t4}}'}
+                style={{ width: '100%', fontFamily: 'monospace', fontSize: '12px' }}
+              />
+            </div>
+            <div className="prop-field">
+              <label>Variable de sortie (ex: dec)</label>
+              <input
+                type="text"
+                value={String(params.output_var || '')}
+                onChange={(e) => updateParam('output_var', e.target.value)}
+                placeholder="dec"
+              />
+            </div>
+            <div className="prop-field">
+              <label>Modèle Ollama</label>
+              <input
+                type="text"
+                value={String(params.model || 'qwen2.5:7b')}
+                onChange={(e) => updateParam('model', e.target.value)}
+                placeholder="qwen2.5:7b"
+              />
+            </div>
+          </>
+        );
+
       default:
         return <div className="prop-info">Pas de paramètres supplémentaires</div>;
     }

visual_workflow_builder/frontend_v4/src/components/ToolPalette.tsx

@@ -54,11 +54,11 @@ export default function ToolPalette() {
                       className="tool-item"
                       draggable
                       onDragStart={(e) => onDragStart(e, action.type)}
-                      title={action.label}
+                      title={`${action.label}\n\n${action.description}${action.needsAnchor ? '\n\n🎯 Nécessite une ancre visuelle' : ''}${action.params.length > 0 ? '\n\nParamètres : ' + action.params.map(p => p.name).join(', ') : ''}`}
                     >
                       <span className="tool-icon">{action.icon}</span>
                       <span className="tool-label">{action.label}</span>
-                      {action.needsAnchor && <span className="tool-anchor-badge">🎯</span>}
+                      {action.needsAnchor && <span className="tool-anchor-badge" title="Nécessite de viser un élément à l'écran">🎯</span>}
                     </div>
                   ))}
                 </div>

visual_workflow_builder/frontend_v4/src/styles.css

@@ -4491,3 +4491,86 @@ body {
 .right-panel-tabbed .capture-library {
   border-top: 1px solid var(--border);
 }
+
+/* === QW4 — PauseDialog & ChecklistPanel === */
+.pause-dialog-overlay {
+  position: fixed;
+  inset: 0;
+  background: rgba(15, 23, 42, 0.45);
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  z-index: 9999;
+}
+.pause-dialog-simple,
+.pause-dialog-checks {
+  padding: 16px;
+  max-width: 480px;
+  background: #fff;
+  border: 2px solid #f59e0b;
+  border-radius: 8px;
+  box-shadow: 0 10px 40px rgba(0, 0, 0, 0.25);
+}
+.pause-dialog-checks h3 { margin: 0 0 8px; color: #92400e; }
+.pause-message { margin: 0 0 12px; }
+.pause-reason-banner {
+  background: #fef3c7;
+  padding: 8px;
+  margin-bottom: 12px;
+  border-radius: 4px;
+}
+.pause-reason { color: #6b7280; display: block; margin-top: 4px; }
+.checklist-panel {
+  list-style: none;
+  padding: 0;
+  margin: 0 0 12px;
+}
+.check-item {
+  padding: 6px 0;
+  border-bottom: 1px solid #f3f4f6;
+}
+.check-item.required { background: #fef9c3; }
+.check-item label {
+  cursor: pointer;
+  display: flex;
+  align-items: center;
+  gap: 6px;
+}
+.badge {
+  font-size: 10px;
+  padding: 2px 6px;
+  border-radius: 10px;
+  margin-left: 6px;
+}
+.badge-required { background: #dc2626; color: #fff; }
+.badge-lea { background: #2563eb; color: #fff; cursor: help; }
+.check-evidence {
+  display: block;
+  font-style: italic;
+  color: #6b7280;
+  margin-left: 24px;
+}
+.pause-error {
+  color: #dc2626;
+  padding: 8px;
+  background: #fef2f2;
+  border-radius: 4px;
+  margin-bottom: 8px;
+}
+.pause-actions {
+  display: flex;
+  gap: 8px;
+  justify-content: flex-end;
+}
+.pause-actions button:disabled {
+  opacity: 0.5;
+  cursor: not-allowed;
+}
+
+/* QW4 — éditeur de safety_checks dans PropertiesPanel */
+.check-editor-row {
+  display: flex;
+  gap: 4px;
+  margin-bottom: 4px;
+  align-items: center;
+}

visual_workflow_builder/frontend_v4/src/types.ts

@@ -1,5 +1,16 @@
 // Types pour l'API v3
 
+// === QW4 — Safety checks (pause supervisée) ===
+export type SafetyLevel = 'standard' | 'medical_critical';
+
+export interface SafetyCheck {
+  id: string;
+  label: string;
+  required: boolean;
+  source: 'declarative' | 'llm_contextual';
+  evidence?: string | null;
+}
+
 // Mode d'exécution
 export type ExecutionMode = 'basic' | 'intelligent' | 'debug' | 'verified';
 
@@ -43,6 +54,8 @@ export type ActionType =
   | 'screenshot_evidence'
   | 'visual_condition'
   | 'loop_visual'
+  | 'pause_for_human'
+  | 't2a_decision'
   | 'download_to_folder'
   | 'ai_analyze_text'
   | 'ai_ocr'
@@ -108,8 +121,9 @@ export const ACTIONS: ActionDefinition[] = [
   ] },
 
   // === EXTRACTION DE DONNÉES ===
-  { type: 'extract_text', label: 'Extraire texte', icon: '📋', description: 'Extrait le texte visible dans la zone de l\'ancre via OCR.', category: 'data', needsAnchor: true, params: [
+  { type: 'extract_text', label: 'Extraire texte (OCR écran)', icon: '📋', description: 'OCR EasyOCR fr+en sur le dernier screenshot. Stocke le texte dans une variable réutilisable plus loin via {{output_var}}. Pas d\'ancre nécessaire — extrait toute la page visible.', category: 'data', needsAnchor: false, params: [
-    { name: 'variable_name', type: 'string', description: 'Nom de la variable pour stocker le résultat' }
+    { name: 'output_var', type: 'string', description: 'Nom de la variable de sortie (ex: texte_motif). Réutilisable via {{nom}}.' },
+    { name: 'paragraph', type: 'boolean', description: 'Regrouper en paragraphes (true) ou lignes brutes (false)' }
   ] },
   { type: 'extract_table', label: 'Extraire tableau', icon: '📊', description: 'Extrait un tableau structuré depuis la zone de l\'ancre.', category: 'data', needsAnchor: true, params: [
     { name: 'variable_name', type: 'string', description: 'Nom de la variable pour stocker le tableau' }
@@ -129,6 +143,16 @@ export const ACTIONS: ActionDefinition[] = [
   { 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' }
   ] },
+  { type: 'pause_for_human', label: 'Pause supervisée', icon: '⏸', description: 'Léa s\'arrête et demande validation humaine via une bulle interactive (boutons Continuer / Annuler).', category: 'logic', needsAnchor: false, params: [
+    { name: 'message', type: 'string', description: 'Message affiché dans la bulle (ex: "Je ne suis pas sûre du critère 3, validez-vous UHCD ?")' },
+    { name: 'safety_level', type: 'select', description: 'Niveau de sécurité : standard (pas de LLM) ou medical_critical (LLM contextuel)' },
+    { name: 'safety_checks', type: 'safety_checks_editor', description: 'Liste de checks à valider avant reprise (id, libellé, obligatoire ?). Édité dans le panneau Propriétés.' }
+  ] },
+  { type: 't2a_decision', label: 'Décision T2A (LLM)', icon: '🧠', description: 'Analyse un DPI urgences via LLM local (qwen2.5:7b par défaut) et propose FORFAIT_URGENCE ou REQUALIFICATION_HOSPITALISATION. Retourne JSON {decision, justification, elements_pour/contre, confiance}. Bench validé 100% accuracy.', category: 'logic', needsAnchor: false, params: [
+    { name: 'input_template', type: 'string', description: 'DPI à analyser. Supporte le templating {{var}} pour concaténer plusieurs extractions (ex: "{{texte_motif}}\\n{{texte_examens}}\\n{{texte_notes}}")' },
+    { name: 'output_var', type: 'string', description: 'Variable de sortie (ex: decision_t2a). Accès aux champs : {{decision_t2a.decision}}, {{decision_t2a.justification}}' },
+    { name: 'model', type: 'string', description: 'Modèle Ollama (default qwen2.5:7b). Autres : t2a-gemma3-27b-q4, gpt-oss:120b-cloud...' }
+  ] },
 
   // === INTELLIGENCE ARTIFICIELLE ===
   { type: 'ai_ocr', label: 'OCR Intelligent', icon: '📝', description: 'Reconnaissance de texte par IA sur la zone de l\'ancre.', category: 'ai', needsAnchor: true, params: [
@@ -301,13 +325,19 @@ export interface WorkflowSummary {
 export interface Execution {
   id: string;
   workflow_id: string;
-  status: 'pending' | 'running' | 'paused' | 'completed' | 'error' | 'cancelled';
+  status: 'pending' | 'running' | 'paused' | 'paused_need_help' | 'completed' | 'error' | 'cancelled';
   progress: number;
   current_step_index: number;
   completed_steps: number;
   failed_steps: number;
   total_steps: number;
   error_message?: string;
+  // === QW4 — Pause supervisée (renvoyés par /replay/state quand status = paused_need_help) ===
+  pause_reason?: string;
+  pause_message?: string;
+  safety_checks?: SafetyCheck[];
+  // ID du replay (utile pour appeler /replay/resume avec acknowledged_check_ids)
+  replay_id?: string;
 }
 
 export interface Session {