feat(qw4): bench rigoureux LLM safety_checks → gemma4:latest par défaut

Bench 5 modèles × 5 scénarios × cold+warm sur RTX 5070 :
- gemma4:latest : warm 2.9s, JSON 92%, détection 46% → gagnant
- qwen2.5vl:7b : warm 6.6s, détection 23% (trop lent)
- qwen2.5vl:3b : warm 2.0s, détection 8% (vérifie pour vérifier)
- medgemma:4b : warm 0.5s, détection 0% (refuse de signaler) → mauvais
  défaut initial, corrigé
- qwen3-vl:8b : 0% JSON valide (ignore format=json Ollama) → écarté

Modifications safety_checks_provider.py :
- RPA_SAFETY_CHECKS_LLM_MODEL défaut: medgemma:4b → gemma4:latest
- RPA_SAFETY_CHECKS_LLM_TIMEOUT_S défaut: 5 → 7 (warm 2.9s + marge)

Doc complète : docs/BENCH_SAFETY_CHECKS_2026-05-06.md
Script : tools/bench_safety_checks_models.py (reproductible, ~10-15 min)

Limite assumée : 46% de détection. À présenter en démo comme aide médecin,
pas certification. Amélioration V2 = prompt plus dirigé sur champs à vérifier.

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

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/requirements.txt

@@ -5,6 +5,7 @@ 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/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_y = int(popup_coords["y_pct"] * height)
+                            real_x = int(popup_coords["x_pct"] * width) + mon_offset_x
+                            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_y = int(y_pct * height)
+                real_x = int(x_pct * width) + mon_offset_x
+                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_y = int(y_pct * height)
+                    real_x = int(x_pct * width) + mon_offset_x
+                    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_y = int(y_pct * height) if y_pct > 0 else int(0.5 * height)
+                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)) + 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

@@ -5,6 +5,7 @@ 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,
@@ -222,6 +224,7 @@ from .replay_engine import (
     _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,
@@ -359,6 +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(
@@ -511,6 +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):
@@ -765,6 +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)."
@@ -1962,6 +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
@@ -2892,8 +2928,54 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
 
             type_ = action.get("type")
 
-            # pause_for_human : no-op en mode autonome — on saute et on continue
+            # 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 "?"
@@ -2914,6 +2996,12 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
                             _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,
@@ -3117,6 +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,
@@ -3855,6 +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,
@@ -3941,8 +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,
@@ -3950,6 +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)
@@ -3968,6 +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"]
@@ -3976,6 +4216,10 @@ 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)
         # pause_for_human est une pause intentionnelle, pas une erreur — ne pas réinjecter

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

@@ -1381,6 +1381,14 @@ def _create_replay_state(
         # 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/safety_checks_provider.py

@@ -0,0 +1,195 @@
+# 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
+
+    # Défaut gemma4:latest : meilleur compromis détection/latence sur bench
+    # 2026-05-06 (cf. docs/BENCH_SAFETY_CHECKS_2026-05-06.md). medgemma:4b
+    # retournait systématiquement [] (refus de signaler).
+    model = _env("RPA_SAFETY_CHECKS_LLM_MODEL", "gemma4:latest")
+    # Timeout 7s : warm avg gemma4 = 2.9s + marge 4s. Cold start ~10s couvert
+    # si le modèle reste résident (OLLAMA_KEEP_ALIVE=24h recommandé prod).
+    timeout_s = _env_int("RPA_SAFETY_CHECKS_LLM_TIMEOUT_S", 7)
+    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.
@@ -157,11 +169,13 @@ def handle_detected_pattern(pattern: Dict[str, Any]) -> bool:
                 screenshot = sct.grab(monitor)
                 screen = Image.frombytes('RGB', screenshot.size, screenshot.bgra, 'raw', 'BGRX')
 
-            # EasyOCR (rapide, bonne qualité GUI) avec fallback docTR
+            # 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=False, verbose=False)
+                _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:
@@ -312,6 +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.
@@ -325,6 +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
@@ -347,6 +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)")
 
@@ -356,12 +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
 
@@ -411,20 +434,43 @@ def _describe_anchor_image(anchor_image_base64: str) -> Optional[str]:
         return None
 
 
-def _capture_screen():
-    """Capture l'écran principal et retourne (PIL.Image, width, height)."""
-    try:
-        import mss
-        from PIL import Image as PILImage
+def _capture_screen(monitor_idx=None):
+    """Capture l'écran et retourne (PIL.Image, width, height, offset_x, offset_y).
 
+    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]]:
@@ -439,7 +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
 
@@ -503,14 +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
@@ -519,7 +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
 
@@ -564,7 +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
@@ -624,7 +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
@@ -646,7 +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
 
@@ -668,7 +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

docs/BENCH_SAFETY_CHECKS_2026-05-06.md

@@ -0,0 +1,95 @@
+# Bench QW4 safety_checks — sélection du LLM contextuel
+
+**Date** : 2026-05-06
+**Contexte** : QW4 du sprint mai. La fonction `_call_llm_for_contextual_checks`
+appelle Ollama avec un screenshot + prompt court pour générer 0-3 checks de
+vérification supplémentaires que l'humain doit acquitter avant la reprise
+d'un replay en pause supervisée (`safety_level=medical_critical`).
+
+## Méthodologie
+
+- **5 scénarios** : screenshots synthétiques de dossiers patient avec UNE
+  anomalie volontaire chacun (date de naissance aberrante, IPP incohérent,
+  diagnostic vide, code CIM inadapté à l'âge, forfait incohérent avec durée).
+- **5 candidats** : `gemma4:latest`, `qwen3-vl:8b`, `qwen2.5vl:7b`,
+  `qwen2.5vl:3b`, `medgemma:4b`.
+- **Protocole par modèle** : déchargement VRAM (keep_alive=0 sur tous les
+  modèles loaded) → 1er appel = cold start chronométré → 4 autres screenshots
+  × 3 runs = 12 mesures warm.
+- **Métriques** : cold start, warm avg, warm p95, % JSON valide, % détection
+  (anomalie cible présente dans label/evidence d'au moins un check renvoyé).
+- **Script** : `tools/bench_safety_checks_models.py`.
+
+## Résultats
+
+| Modèle | Cold (s) | Warm avg (s) | Warm p95 (s) | JSON | Détection |
+|---|---:|---:|---:|---:|---:|
+| `gemma4:latest` | 10.6 | **2.9** | 3.4 | 92% (12/13) | **46% (6/13)** |
+| `qwen3-vl:8b` | 5.6 | — | — | **0%** (0/12) | 0% (0/12) |
+| `qwen2.5vl:7b` | 9.4 | 6.6 | 8.1 | 100% (13/13) | 23% (3/13) |
+| `qwen2.5vl:3b` | 6.0 | 2.0 | 2.5 | 100% (13/13) | 8% (1/13) |
+| `medgemma:4b` | 2.0 | 0.5 | 0.7 | 100% (13/13) | **0%** (0/13) |
+
+## Lecture
+
+- **`medgemma:4b` retourne systématiquement `[]`** sur les 13 mesures.
+  Trop obéissant à "Si rien d'inhabituel à signaler, retourne []", refuse
+  de pointer ne serait-ce qu'une date 1900-01-01. **Mauvais choix par défaut**
+  malgré sa rapidité et sa spécialisation médicale revendiquée.
+- **`qwen3-vl:8b` ignore `format=json` Ollama** : 0 réponse parsable. À écarter
+  pour cette tâche tant que le tooling Ollama / le modèle ne convergent pas.
+- **`qwen2.5vl:7b`** détecte mais 2× plus lent (warm 6.6s) que gemma4 et tend
+  à inventer des anomalies de format de date qui ne sont pas la vraie cible.
+- **`qwen2.5vl:3b`** rapide mais détection 8% — il "vérifie pour vérifier"
+  (renvoie souvent "vérification de la date de naissance" même quand la date
+  est correcte).
+- **`gemma4:latest` gagne** : meilleur taux de détection (46%) ET deuxième
+  meilleur warm (2.9s). Tend à raisonner cohérence motif/diagnostic plutôt
+  que valeurs aberrantes brutes.
+
+## Détail détection par scénario
+
+| Scénario | gemma4 | qwen2.5vl:7b | qwen2.5vl:3b | medgemma:4b |
+|---|:---:|:---:|:---:|:---:|
+| Date naissance aberrante (1900) | ❌ | ✅ | ✅ | ❌ |
+| IPP incohérent (`ABC@@##XYZ`) | ❌ | ❌ | ❌ | ❌ |
+| Diagnostic principal vide | ✅ | ❌ | ❌ | ❌ |
+| Code CIM inadapté à l'âge | ✅ | ❌ | ❌ | ❌ |
+| Forfait UHCD vs durée 1h | ❌ | ❌ | ❌ | ❌ |
+
+Aucun modèle ne détecte les 5 scénarios. **L'IPP corrompu et le forfait
+incohérent ne sont détectés par personne** — ces anomalies demanderaient
+soit un prompt plus dirigé (liste explicite des champs à vérifier), soit
+un modèle plus large.
+
+## Décision
+
+- **Défaut serveur** : `RPA_SAFETY_CHECKS_LLM_MODEL=gemma4:latest`
+- **Timeout** : `RPA_SAFETY_CHECKS_LLM_TIMEOUT_S=7` (warm 2.9s + marge)
+- **Persistance VRAM** : `OLLAMA_KEEP_ALIVE=24h` recommandé pour éviter le
+  cold start de 10s en démo
+
+Modifications appliquées dans `agent_v0/server_v1/safety_checks_provider.py`.
+
+## Limites & travail futur
+
+1. **46% de détection est faible** : à présenter comme aide au médecin, pas
+  comme certification. Le médecin reste le décideur.
+2. **Prompt actuel trop générique** : un prompt qui liste explicitement les
+  champs à vérifier (DDN, IPP, diagnostic, forfait, cohérence âge/diagnostic)
+  donnerait probablement de meilleurs résultats. À mesurer en V2.
+3. **Bench sur 5 anomalies seulement** : à étendre dès qu'on a un corpus de
+  vrais dossiers Easily Assure avec anomalies confirmées par Pauline / Amina.
+4. **Pas de test sur des dossiers SANS anomalie** (faux positifs) : à ajouter.
+5. **Pas de bench des modèles cloud** (gemma3:27b-cloud, deepseek, gpt-oss)
+   par contrainte 100% local — mais à explorer si on lève cette contrainte
+   pour les checks contextuels (qui ne contiennent pas de PII si on
+   anonymise les screenshots).
+
+## Reproductibilité
+
+```bash
+cd /home/dom/ai/rpa_vision_v3
+.venv/bin/python tools/bench_safety_checks_models.py
+# (BENCH_TIMEOUT=60 par défaut, ~10-15 min sur RTX 5070)
+```

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) |

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_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/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)

tools/bench_safety_checks_models.py

@@ -0,0 +1,437 @@
+#!/usr/bin/env python3
+"""Bench rigoureux des modèles candidats pour QW4 safety_checks contextuels.
+
+Méthodologie :
+- 5 screenshots synthétiques avec différentes anomalies cliniques
+- 4 modèles candidats (gemma4:e4b sur :11435, qwen2.5vl:7b/3b et medgemma:4b sur :11434)
+- Pour chaque modèle :
+    1. Décharger TOUS les modèles déjà en VRAM (keep_alive=0)
+    2. 1er appel = cold start chronométré (1er screenshot)
+    3. 12 appels warm = (4 autres screenshots × 3 runs)
+    4. Mesurer : cold_start, warm avg/p95, taux détection, JSON valide
+
+Usage : .venv/bin/python tools/bench_safety_checks_models.py
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import os
+import statistics
+import time
+from dataclasses import dataclass, field
+from typing import Any
+
+import requests
+from PIL import Image, ImageDraw, ImageFont
+
+
+OLLAMA_PRIMARY = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+OLLAMA_SECONDARY = os.environ.get("GEMMA4_URL", "http://localhost:11435")
+
+# Configuration des candidats : (nom, url, type)
+CANDIDATES = [
+    ("gemma4:latest",  OLLAMA_PRIMARY,   "vlm_default"),
+    ("qwen3-vl:8b",    OLLAMA_PRIMARY,   "vision_qwen3_8b"),
+    ("qwen2.5vl:7b",   OLLAMA_PRIMARY,   "vision_qwen25_7b"),
+    ("qwen2.5vl:3b",   OLLAMA_PRIMARY,   "vision_qwen25_3b"),
+    ("medgemma:4b",    OLLAMA_PRIMARY,   "medical_4b"),
+]
+
+TIMEOUT_S = int(os.environ.get("BENCH_TIMEOUT", "60"))   # large pour ne rien rater
+MAX_CHECKS = 3
+WORKFLOW_MESSAGE = "Validation T2A avant codage UHCD"
+EXISTING_LABELS: list[str] = []
+WARM_RUNS_PER_SCREENSHOT = 3  # warm = 4 autres screenshots × 3 runs = 12 mesures
+
+
+# ---------------------------------------------------------------------------
+# Scénarios : 5 screenshots avec anomalies différentes
+# ---------------------------------------------------------------------------
+
+@dataclass
+class Scenario:
+    label: str            # nom court
+    rows: list[tuple[str, str]]
+    anomaly_keywords: list[str]    # mots indiquant que l'anomalie est repérée
+
+
+SCENARIOS = [
+    Scenario(
+        label="ddn_aberrante",
+        rows=[
+            ("Nom :", "DUPONT Marie"),
+            ("IPP :", "25003284"),
+            ("Date de naissance :", "1900-01-01"),  # ANOMALIE
+            ("Sexe :", "F"),
+            ("Date d'admission :", "2026-05-05 14:32"),
+            ("Service :", "URGENCES"),
+            ("Motif :", "Douleur abdominale aiguë"),
+            ("Diagnostic principal :", "K35.8 - Appendicite aiguë"),
+            ("Forfait facturation :", "UHCD - Forfait 24h"),
+        ],
+        anomaly_keywords=["1900", "naissance", "ddn", "date"],
+    ),
+    Scenario(
+        label="ipp_incoherent",
+        rows=[
+            ("Nom :", "MARTIN Paul"),
+            ("IPP :", "ABC@@##XYZ"),  # ANOMALIE : non numérique
+            ("Date de naissance :", "1965-04-12"),
+            ("Sexe :", "M"),
+            ("Date d'admission :", "2026-05-06 09:15"),
+            ("Service :", "URGENCES"),
+            ("Motif :", "Chute mécanique"),
+            ("Diagnostic principal :", "S52.5 - Fracture du radius distal"),
+            ("Forfait facturation :", "UHCD - Forfait 24h"),
+        ],
+        anomaly_keywords=["ipp", "abc", "format", "incohérent", "incoherent", "invalide"],
+    ),
+    Scenario(
+        label="diagnostic_vide",
+        rows=[
+            ("Nom :", "BERNARD Sophie"),
+            ("IPP :", "25004191"),
+            ("Date de naissance :", "1972-11-08"),
+            ("Sexe :", "F"),
+            ("Date d'admission :", "2026-05-06 10:42"),
+            ("Service :", "URGENCES"),
+            ("Motif :", "Céphalées"),
+            ("Diagnostic principal :", ""),  # ANOMALIE : vide
+            ("Forfait facturation :", "UHCD - Forfait 24h"),
+        ],
+        anomaly_keywords=["diagnostic", "vide", "blanc", "absent", "manque", "non renseigné", "non renseigne"],
+    ),
+    Scenario(
+        label="cim_inadapte_age",
+        rows=[
+            ("Nom :", "PETIT Lucas"),
+            ("IPP :", "25004222"),
+            ("Date de naissance :", "2025-11-01"),  # nourrisson 6 mois
+            ("Sexe :", "M"),
+            ("Date d'admission :", "2026-05-06 11:00"),
+            ("Service :", "URGENCES PEDIATRIQUES"),
+            ("Motif :", "Pleurs persistants"),
+            ("Diagnostic principal :", "M19.9 - Arthrose, sans précision"),  # ANOMALIE
+            ("Forfait facturation :", "UHCD - Forfait 24h"),
+        ],
+        anomaly_keywords=["arthrose", "âge", "age", "nourrisson", "incohérent", "incoherent", "m19", "incompatible"],
+    ),
+    Scenario(
+        label="forfait_incoherent_duree",
+        rows=[
+            ("Nom :", "ROUSSEAU Jean"),
+            ("IPP :", "25004317"),
+            ("Date de naissance :", "1958-03-22"),
+            ("Sexe :", "M"),
+            ("Date d'admission :", "2026-05-06 08:00"),
+            ("Date de sortie :", "2026-05-06 09:00"),  # 1h
+            ("Service :", "URGENCES"),
+            ("Motif :", "Bilan biologique"),
+            ("Diagnostic principal :", "Z00.0 - Examen médical général"),
+            ("Forfait facturation :", "UHCD - Forfait 24h"),  # ANOMALIE : 1h ≠ UHCD 24h
+        ],
+        anomaly_keywords=["forfait", "uhcd", "durée", "duree", "1h", "incohérent", "incoherent", "24h"],
+    ),
+]
+
+
+# ---------------------------------------------------------------------------
+# Génération des screenshots
+# ---------------------------------------------------------------------------
+
+def make_screenshot(scenario: Scenario, path: str) -> None:
+    """Crée un PNG du dossier patient pour un scénario donné."""
+    img = Image.new("RGB", (1024, 600), color="white")
+    draw = ImageDraw.Draw(img)
+    try:
+        font_title = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 22)
+        font_body = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 18)
+    except OSError:
+        font_title = ImageFont.load_default()
+        font_body = ImageFont.load_default()
+
+    draw.text((20, 20), "DOSSIER PATIENT - URGENCES UHCD", fill="black", font=font_title)
+    draw.line([(20, 55), (1004, 55)], fill="black", width=2)
+    y = 80
+    for label, value in scenario.rows:
+        draw.text((30, y), label, fill="black", font=font_body)
+        draw.text((280, y), value, fill="#1f2937", font=font_body)
+        y += 35
+    img.save(path, format="PNG")
+
+
+def encode_image(path: str) -> str:
+    with open(path, "rb") as f:
+        return base64.b64encode(f.read()).decode("ascii")
+
+
+def build_prompt() -> str:
+    existing = ", ".join(EXISTING_LABELS) if EXISTING_LABELS else "aucun"
+    return 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"}}
+  ]
+}}
+"""
+
+
+# ---------------------------------------------------------------------------
+# Gestion VRAM Ollama (déchargement)
+# ---------------------------------------------------------------------------
+
+def list_loaded_models(url: str) -> list[str]:
+    """Retourne la liste des modèles actuellement en VRAM sur cet Ollama."""
+    try:
+        resp = requests.get(f"{url}/api/ps", timeout=5)
+        if resp.status_code == 200:
+            data = resp.json()
+            return [m["name"] for m in data.get("models", [])]
+    except Exception:
+        pass
+    return []
+
+
+def unload_all_models() -> None:
+    """Décharge tous les modèles en VRAM sur les 2 Ollama (keep_alive=0)."""
+    for url in (OLLAMA_PRIMARY, OLLAMA_SECONDARY):
+        loaded = list_loaded_models(url)
+        for model_name in loaded:
+            try:
+                requests.post(
+                    f"{url}/api/generate",
+                    json={"model": model_name, "prompt": "", "keep_alive": 0, "stream": False},
+                    timeout=10,
+                )
+            except Exception:
+                pass
+    # Petit temps pour laisser le GC GPU faire son travail
+    time.sleep(2)
+
+
+# ---------------------------------------------------------------------------
+# Appel modèle + parsing
+# ---------------------------------------------------------------------------
+
+@dataclass
+class CallResult:
+    elapsed_s: float
+    error: str = ""
+    raw: str = ""
+    checks: list[dict] = field(default_factory=list)
+
+
+def call_model(model: str, url: str, prompt: str, image_b64: str) -> CallResult:
+    payload = {
+        "model": model,
+        "prompt": prompt,
+        "stream": False,
+        "format": "json",
+        "options": {"temperature": 0.1, "num_predict": 250},
+        "images": [image_b64],
+    }
+    t0 = time.perf_counter()
+    try:
+        resp = requests.post(f"{url}/api/generate", json=payload, timeout=TIMEOUT_S)
+        elapsed = time.perf_counter() - t0
+    except requests.Timeout:
+        return CallResult(elapsed_s=TIMEOUT_S, error="TIMEOUT")
+    except Exception as e:
+        return CallResult(elapsed_s=time.perf_counter() - t0, error=f"NETWORK:{type(e).__name__}")
+
+    if resp.status_code != 200:
+        return CallResult(elapsed_s=elapsed, error=f"HTTP_{resp.status_code}", raw=resp.text[:200])
+
+    raw = resp.json().get("response", "").strip()
+    try:
+        parsed = json.loads(raw)
+        checks = parsed.get("additional_checks") or []
+        if not isinstance(checks, list):
+            checks = []
+        return CallResult(elapsed_s=elapsed, raw=raw[:300], checks=checks)
+    except json.JSONDecodeError as e:
+        return CallResult(elapsed_s=elapsed, error=f"JSON:{type(e).__name__}", raw=raw[:200])
+
+
+def detects_anomaly(scenario: Scenario, checks: list[dict]) -> bool:
+    blob = " ".join(
+        f"{c.get('label', '')} {c.get('evidence', '')}".lower()
+        for c in checks
+    )
+    return any(pat.lower() in blob for pat in scenario.anomaly_keywords)
+
+
+# ---------------------------------------------------------------------------
+# Bench main
+# ---------------------------------------------------------------------------
+
+@dataclass
+class ModelStats:
+    model: str
+    cold_s: float = 0.0
+    warm_times: list[float] = field(default_factory=list)
+    detection_count: int = 0
+    detection_total: int = 0
+    json_valid_count: int = 0
+    json_valid_total: int = 0
+    errors: list[str] = field(default_factory=list)
+    sample_checks: list[tuple[str, list[dict]]] = field(default_factory=list)  # (scenario_label, checks)
+
+
+def run_bench_for_model(model: str, url: str, screenshots: list[tuple[Scenario, str]]) -> ModelStats:
+    print(f"\n══════════════════════════════════════════════════════════")
+    print(f"  MODEL: {model}  ({url})")
+    print(f"══════════════════════════════════════════════════════════")
+
+    # Décharger tout
+    print(f"  [1/3] Déchargement VRAM...", end=" ", flush=True)
+    unload_all_models()
+    loaded_after = list_loaded_models(OLLAMA_PRIMARY) + list_loaded_models(OLLAMA_SECONDARY)
+    print(f"OK (loaded={loaded_after if loaded_after else 'aucun'})")
+
+    stats = ModelStats(model=model)
+    prompt = build_prompt()
+
+    # Cold start sur le 1er screenshot
+    scen0, path0 = screenshots[0]
+    img_b64 = encode_image(path0)
+    print(f"  [2/3] Cold start ({scen0.label})...", end=" ", flush=True)
+    r0 = call_model(model, url, prompt, img_b64)
+    stats.cold_s = r0.elapsed_s
+    if r0.error:
+        print(f"❌ {r0.error} ({r0.elapsed_s:.1f}s)")
+        stats.errors.append(f"cold:{scen0.label}:{r0.error}")
+    else:
+        det = detects_anomaly(scen0, r0.checks)
+        stats.detection_count += int(det)
+        stats.detection_total += 1
+        stats.json_valid_count += 1
+        stats.json_valid_total += 1
+        stats.sample_checks.append((scen0.label, r0.checks))
+        print(f"{'✅' if det else '⚠️'} {len(r0.checks)} check(s) en {r0.elapsed_s:.1f}s (det={det})")
+
+    # Warm runs sur les 4 autres screenshots × N runs
+    print(f"  [3/3] Warm runs ({len(screenshots)-1} scenarios × {WARM_RUNS_PER_SCREENSHOT} runs)...")
+    for scen, path in screenshots[1:]:
+        img_b64 = encode_image(path)
+        for run_idx in range(WARM_RUNS_PER_SCREENSHOT):
+            r = call_model(model, url, prompt, img_b64)
+            if r.error:
+                stats.errors.append(f"{scen.label}:run{run_idx}:{r.error}")
+                stats.json_valid_total += 1
+                stats.detection_total += 1
+                print(f"      {scen.label} run{run_idx}: ❌ {r.error}")
+                continue
+            stats.warm_times.append(r.elapsed_s)
+            stats.json_valid_count += 1
+            stats.json_valid_total += 1
+            det = detects_anomaly(scen, r.checks)
+            stats.detection_count += int(det)
+            stats.detection_total += 1
+            if run_idx == 0:
+                stats.sample_checks.append((scen.label, r.checks))
+            print(f"      {scen.label} run{run_idx}: {'✅' if det else '⚠️'} {len(r.checks)} check(s) en {r.elapsed_s:.1f}s")
+    return stats
+
+
+def print_summary_table(all_stats: list[ModelStats]) -> None:
+    print("\n\n══════════════════════════════════════════════════════════")
+    print("  SYNTHÈSE")
+    print("══════════════════════════════════════════════════════════\n")
+    print("| Modèle | Cold (s) | Warm avg (s) | Warm p95 (s) | JSON | Détection | Notes |")
+    print("|---|---:|---:|---:|---:|---:|---|")
+    for s in all_stats:
+        if s.warm_times:
+            warm_avg = statistics.mean(s.warm_times)
+            warm_p95 = sorted(s.warm_times)[int(len(s.warm_times) * 0.95) - 1] if len(s.warm_times) > 1 else s.warm_times[0]
+        else:
+            warm_avg = warm_p95 = 0.0
+        json_pct = (s.json_valid_count / s.json_valid_total * 100) if s.json_valid_total else 0
+        det_pct = (s.detection_count / s.detection_total * 100) if s.detection_total else 0
+        notes = f"{len(s.errors)} err" if s.errors else "OK"
+        print(f"| `{s.model}` | {s.cold_s:.1f} | {warm_avg:.1f} | {warm_p95:.1f} | "
+              f"{json_pct:.0f}% ({s.json_valid_count}/{s.json_valid_total}) | "
+              f"{det_pct:.0f}% ({s.detection_count}/{s.detection_total}) | {notes} |")
+
+    print("\n## Détail des checks par scénario\n")
+    for s in all_stats:
+        print(f"\n### `{s.model}`")
+        if s.errors:
+            print(f"_Erreurs ({len(s.errors)})_ : {s.errors[:5]}{'...' if len(s.errors) > 5 else ''}")
+        for label, checks in s.sample_checks:
+            if not checks:
+                print(f"- **{label}** : _aucun check_")
+            else:
+                for c in checks[:2]:
+                    print(f"- **{label}** : {c.get('label', '?')} — _{c.get('evidence', '?')[:120]}_")
+
+
+def pick_winner(all_stats: list[ModelStats]) -> ModelStats | None:
+    """Le gagnant : meilleur taux détection, départage par warm avg."""
+    valid = [s for s in all_stats if s.warm_times]
+    if not valid:
+        return None
+    # Tri : détection desc puis warm avg asc
+    valid.sort(key=lambda s: (-(s.detection_count / max(s.detection_total, 1)), statistics.mean(s.warm_times)))
+    return valid[0]
+
+
+def main() -> int:
+    # Génération des 5 screenshots
+    print("📸 Génération des 5 screenshots synthétiques :")
+    screenshots: list[tuple[Scenario, str]] = []
+    for scen in SCENARIOS:
+        path = f"/tmp/bench_safety_{scen.label}.png"
+        make_screenshot(scen, path)
+        print(f"  - {scen.label} → {path}")
+        screenshots.append((scen, path))
+
+    print(f"\n⏱  Timeout par appel : {TIMEOUT_S}s")
+    print(f"🔄 Warm runs par scénario : {WARM_RUNS_PER_SCREENSHOT}")
+    print(f"📊 Total mesures par modèle : 1 cold + {(len(SCENARIOS)-1) * WARM_RUNS_PER_SCREENSHOT} warm = "
+          f"{1 + (len(SCENARIOS)-1) * WARM_RUNS_PER_SCREENSHOT}")
+    print(f"🤖 Candidats : {[c[0] for c in CANDIDATES]}")
+
+    all_stats: list[ModelStats] = []
+    for model, url, _ in CANDIDATES:
+        try:
+            stats = run_bench_for_model(model, url, screenshots)
+            all_stats.append(stats)
+        except KeyboardInterrupt:
+            print(f"\n⚠️  Interrompu pendant {model}, on saute le reste")
+            break
+        except Exception as e:
+            print(f"\n❌ Crash bench {model}: {e}")
+            all_stats.append(ModelStats(model=model, errors=[f"crash:{e}"]))
+
+    print_summary_table(all_stats)
+
+    winner = pick_winner(all_stats)
+    print("\n## Recommandation\n")
+    if winner is None:
+        print("⚠️  Aucun modèle exploitable. Décision manuelle nécessaire.")
+        return 1
+    det_pct = winner.detection_count / max(winner.detection_total, 1) * 100
+    warm_avg = statistics.mean(winner.warm_times)
+    print(f"🏆 **{winner.model}** : détection {det_pct:.0f}%, warm avg {warm_avg:.1f}s, cold {winner.cold_s:.1f}s")
+    print(f"\nPour fixer en production :")
+    print(f"```bash\nsudo systemctl edit rpa-streaming")
+    print(f"# [Service]\n# Environment=RPA_SAFETY_CHECKS_LLM_MODEL={winner.model}")
+    print(f"sudo systemctl restart rpa-streaming\n```")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

visual_workflow_builder/backend/api_v3/dag_execute.py

@@ -1113,3 +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/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/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>
+            <>
+              <button className="btn-start" onClick={onStart} title="Exécuter sur cet écran (Linux local)">
+                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/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';
 
@@ -133,7 +144,9 @@ export const ACTIONS: ActionDefinition[] = [
     { 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: '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}}")' },
@@ -312,13 +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 {