diff --git a/agent_v0/server_v1/api_stream.py b/agent_v0/server_v1/api_stream.py
index e3be254e3..2d794cb61 100644
--- a/agent_v0/server_v1/api_stream.py
+++ b/agent_v0/server_v1/api_stream.py
@@ -1166,26 +1166,37 @@ async def get_next_action(session_id: str, machine_id: str = "default"):
     with _replay_lock:
         queue = _replay_queues.get(session_id, [])
 
-        if not queue:
-            # Seul le lookup machine_replay_target est conservé (sûr : mapping explicite
-            # créé lors du POST /replay). Le cross-session stealing a été supprimé
-            # car il causait des race conditions entre agents.
-            if machine_id != "default":
-                target_sid = _machine_replay_target.get(machine_id)
-                if target_sid and target_sid != session_id:
-                    target_queue = _replay_queues.get(target_sid, [])
-                    if target_queue:
-                        logger.info(
-                            f"Replay machine-target: {machine_id} -> "
-                            f"transfert queue {target_sid} -> {session_id}"
-                        )
-                        queue = target_queue
-                        _replay_queues[session_id] = target_queue
-                        del _replay_queues[target_sid]
-                        for state in _replay_states.values():
-                            if state["session_id"] == target_sid and state["status"] == "running":
-                                state["session_id"] = session_id
-                        _machine_replay_target[machine_id] = session_id
+        if not queue and machine_id != "default":
+            # Lookup 1 : machine_replay_target (mapping explicite POST /replay)
+            target_sid = _machine_replay_target.get(machine_id)
+            if target_sid and target_sid != session_id:
+                target_queue = _replay_queues.get(target_sid, [])
+                if target_queue:
+                    queue = target_queue
+                    _replay_queues[session_id] = target_queue
+                    del _replay_queues[target_sid]
+                    for state in _replay_states.values():
+                        if state["session_id"] == target_sid and state["status"] == "running":
+                            state["session_id"] = session_id
+                    _machine_replay_target[machine_id] = session_id
+                    logger.info(f"Replay machine-target: {machine_id} -> {target_sid} -> {session_id}")
+
+            # Lookup 2 : chercher dans les replay_states actifs pour cette machine
+            if not queue:
+                for state in _replay_states.values():
+                    if (state.get("machine_id") == machine_id
+                            and state["status"] == "running"
+                            and state["session_id"] != session_id):
+                        other_sid = state["session_id"]
+                        other_queue = _replay_queues.get(other_sid, [])
+                        if other_queue:
+                            queue = other_queue
+                            _replay_queues[session_id] = other_queue
+                            del _replay_queues[other_sid]
+                            state["session_id"] = session_id
+                            _machine_replay_target[machine_id] = session_id
+                            logger.info(f"Replay machine-state: {machine_id} -> {other_sid} -> {session_id}")
+                            break
 
         if not queue:
             return {"action": None, "session_id": session_id, "machine_id": machine_id}
@@ -1603,11 +1614,12 @@ async def resolve_target(request: ResolveTargetRequest):
         tmp_path = tmp.name
 
     try:
-        # Lancer la résolution visuelle dans le thread GPU
+        # Lancer la résolution visuelle dans un thread SÉPARÉ (pas le GPU executor
+        # qui peut être saturé par le SessionWorker). Le template matching est CPU-only.
         import asyncio
         loop = asyncio.get_event_loop()
         result = await loop.run_in_executor(
-            _gpu_executor,
+            None,  # ThreadPool par défaut (pas _gpu_executor)
             _resolve_target_sync,
             tmp_path,
             request.target_spec,
diff --git a/visual_workflow_builder/backend/api_v3/dag_execute.py b/visual_workflow_builder/backend/api_v3/dag_execute.py
index 464efa348..3e4c41957 100644
--- a/visual_workflow_builder/backend/api_v3/dag_execute.py
+++ b/visual_workflow_builder/backend/api_v3/dag_execute.py
@@ -969,11 +969,26 @@ def execute_windows():
             if vwb_type in ('keyboard_shortcut', 'hotkey') and 'keys' in params:
                 action['keys'] = params['keys']
 
+    # Injecter le machine_id pour le ciblage multi-machine
+    # Chercher la première machine Windows connectée si pas spécifié
+    if 'machine_id' not in data or not data.get('machine_id'):
+        try:
+            machines_resp = req.get('http://localhost:5005/api/v1/traces/stream/machines', timeout=3)
+            if machines_resp.ok:
+                machines = machines_resp.json().get('machines', [])
+                for m in machines:
+                    mid = m.get('machine_id', '')
+                    if mid and mid != 'default' and 'windows' in mid.lower():
+                        data['machine_id'] = mid
+                        break
+        except Exception:
+            pass
+
     try:
         resp = req.post(
             'http://localhost:5005/api/v1/traces/stream/replay/raw',
             json=data,
-            timeout=30,  # Augmenté car le template matching peut prendre du temps
+            timeout=30,
         )
         return jsonify(resp.json()), resp.status_code
     except req.ConnectionError: