feat(qw2): hook LoopDetector dans api_stream + extension replay_state

replay_state enrichi de _screenshot_history (5 dernières images PIL) et
_action_history (5 dernières signatures action).

report_action_result :
- met à jour les deux anneaux après chaque action
- évalue le LoopDetector (singleton lazy avec _clip_embedder serveur)
- si detected → bascule paused_need_help avec pause_reason="loop_detected"
  et bus event lea:loop_detected (signal + evidence)

Tous les chemins d'erreur (embedder absent, OOM, exception) loggent et
laissent le replay continuer — aucun blocage par la couche détection.

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_x = int(popup_coords["x_pct"] * width) + mon_offset_x
-                            real_y = int(popup_coords["y_pct"] * height)
+                            real_y = int(popup_coords["y_pct"] * height) + mon_offset_y
                             self._click((real_x, real_y), "left")
                             time.sleep(1.0)
                             print(f"    [OBSERVER] Popup fermée — reprise du flow normal")
@@ -718,8 +733,8 @@ class ActionExecutorV1:
                         self.notifier.replay_target_not_found(target_desc)
                         return result
 
-                real_x = int(x_pct * width)
+                real_x = int(x_pct * width) + mon_offset_x
-                real_y = int(y_pct * height)
+                real_y = int(y_pct * height) + mon_offset_y
                 button = action.get("button", "left")
                 mode = "VISUAL" if result.get("visual_resolved") else "COORD"
                 print(
@@ -781,8 +796,8 @@ class ActionExecutorV1:
                     print(f"    [TYPE] raw_keys disponibles ({len(raw_keys)} events) — replay exact")
                 # Cliquer sur le champ avant de taper (si coordonnees disponibles)
                 if x_pct > 0 and y_pct > 0:
-                    real_x = int(x_pct * width)
+                    real_x = int(x_pct * width) + mon_offset_x
-                    real_y = int(y_pct * height)
+                    real_y = int(y_pct * height) + mon_offset_y
                     print(f"    [TYPE] Clic prealable sur ({real_x}, {real_y})")
                     self._click((real_x, real_y), "left")
                     time.sleep(0.3)
@@ -808,8 +823,8 @@ class ActionExecutorV1:
                 logger.info(f"Replay key_combo : {keys} (raw_keys={'oui' if raw_keys else 'non'})")
 
             elif action_type == "scroll":
-                real_x = int(x_pct * width) if x_pct > 0 else int(0.5 * width)
+                real_x = (int(x_pct * width) if x_pct > 0 else int(0.5 * width)) + mon_offset_x
-                real_y = int(y_pct * height) if y_pct > 0 else int(0.5 * height)
+                real_y = (int(y_pct * height) if y_pct > 0 else int(0.5 * height)) + mon_offset_y
                 delta = action.get("delta", -3)
                 print(f"    [SCROLL] delta={delta} a ({real_x}, {real_y})")
                 self.mouse.position = (real_x, real_y)
@@ -1386,6 +1401,16 @@ Example: x_pct=0.50, y_pct=0.30"""
             data = resp.json()
             action = data.get("action")
             if action is None:
+                # pause_for_human : afficher le message de décision à l'utilisateur
+                if data.get("replay_paused") and data.get("pause_message"):
+                    msg = data["pause_message"]
+                    print(f"[PAUSE] {msg}")
+                    logger.info(f"Replay en pause — message : {msg}")
+                    self.notifier.notify(
+                        title="Léa — Validation requise",
+                        message=msg[:250],
+                        timeout=30,
+                    )
                 return False
 
         except (requests.exceptions.ConnectionError, requests.exceptions.Timeout) as e:

agent_v0/deploy/windows_client/agent_v1/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
@@ -2914,6 +2950,12 @@ 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 +3159,51 @@ 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 +3942,82 @@ 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,

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 +1,88 @@
+# 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.
+"""
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+
+@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
+
+    # 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")
+
+    # 3. Fallback composite (backward compat — comportement actuel mss.monitors[0])
+    return _COMPOSITE_FALLBACK

agent_v0/server_v1/replay_engine.py

@@ -1381,6 +1381,9 @@ 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)
     }
 
 

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():
+def _capture_screen(monitor_idx=None):
-    """Capture l'écran principal et retourne (PIL.Image, width, height)."""
+    """Capture l'écran et retourne (PIL.Image, width, height, offset_x, offset_y).
-    try:
-        import mss
-        from PIL import Image as PILImage
 
+    Args:
+        monitor_idx: Index logique 0..N-1 du monitor à capturer (cf. screeninfo).
+            Si None : capture composite (mss.monitors[0]) — comportement legacy.
+
+    Returns:
+        (image, w, h, offset_x, offset_y). offset = (0,0) en mode composite.
+    """
+    try:
         with mss.mss() as sct:
-            monitor = sct.monitors[0]
+            if monitor_idx is None:
+                # Comportement actuel : composite tous écrans
+                monitor = sct.monitors[0]
+                offset_x, offset_y = 0, 0
+            else:
+                # mss skip monitors[0] (composite). Index logique 0 → mss.monitors[1].
+                mss_idx = int(monitor_idx) + 1
+                if mss_idx >= len(sct.monitors):
+                    logger.warning(
+                        "mss.monitors[%d] hors limites (n=%d) — fallback composite",
+                        mss_idx, len(sct.monitors),
+                    )
+                    monitor = sct.monitors[0]
+                    offset_x, offset_y = 0, 0
+                else:
+                    monitor = sct.monitors[mss_idx]
+                    offset_x = int(monitor.get("left", 0))
+                    offset_y = int(monitor.get("top", 0))
+
             screenshot = sct.grab(monitor)
             screen = PILImage.frombytes('RGB', screenshot.size, screenshot.bgra, 'raw', 'BGRX')
-            return screen, monitor['width'], monitor['height']
+            return screen, monitor['width'], monitor['height'], offset_x, offset_y
     except Exception as e:
         logger.debug(f"Capture écran échouée: {e}")
-        return None, 0, 0
+        return None, 0, 0, 0, 0
 
 
 def _grounding_ocr(target_text: str, anchor_bbox: Optional[Dict] = None) -> Optional[Dict[str, Any]]:
@@ -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

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/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"