feat: grounding sur image fenêtre au lieu du full screen

Utilise shot_XXXX_window.png (capture fenêtre active) au lieu du
full screen pour le grounding VLM. Image plus petite, ciblée,
sans bruit (taskbar, autres fenêtres).

Coordonnées fenêtre converties en coordonnées écran via window_rect.
window_capture (rect, window_size, click_relative) ajouté au target_spec.

Résultat : 50% → 80% de précision sur la session VM (16/20 clics).

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

agent_v0/server_v1/api_stream.py

@@ -4548,21 +4548,35 @@ def _resolve_by_grounding(
     else:
         return None
 
-    # Redimensionner le screenshot (800px de large pour le VLM)
+    # Utiliser la capture fenêtre si disponible (plus ciblée, moins de bruit)
+    # Sinon fallback sur le full screen
+    window_capture = target_spec.get("window_capture", {})
+    window_rect = window_capture.get("rect")  # [x1, y1, x2, y2] écran
+
     try:
         from PIL import Image as PILImage
-        img = PILImage.open(screenshot_path)
+        from pathlib import Path
+
+        # Chercher le screenshot fenêtre (_window.png)
+        full_path = Path(screenshot_path)
+        win_path = full_path.parent / full_path.name.replace("_full.png", "_window.png")
+
+        if win_path.is_file() and window_rect:
+            img = PILImage.open(str(win_path))
+            using_window = True
+            logger.debug("Grounding : image fenêtre %s (%dx%d)", win_path.name, *img.size)
+        else:
+            img = PILImage.open(screenshot_path)
+            using_window = False
+
         orig_w, orig_h = img.size
-        target_w = 800
-        ratio = target_w / orig_w
-        img_small = img.resize((target_w, int(orig_h * ratio)))
-        small_w, small_h = img_small.size
+        small_w, small_h = orig_w, orig_h  # pas de redimensionnement
 
         buf = io.BytesIO()
-        img_small.save(buf, format="JPEG", quality=75)
+        img.save(buf, format="JPEG", quality=80)
         shot_b64 = base64.b64encode(buf.getvalue()).decode()
     except Exception as e:
-        logger.warning("Grounding : erreur redimensionnement — %s", e)
+        logger.warning("Grounding : erreur chargement image — %s", e)
         return None
 
     # Prompt natif Qwen2.5-VL — format bbox_2d (le seul fiable)
@@ -4723,10 +4737,25 @@ def _resolve_by_grounding(
         logger.info("Grounding : coordonnées hors bornes (%.3f, %.3f)", x_pct, y_pct)
         return None
 
-    logger.info(
-        "Grounding OK [%s] : '%s' → (%.4f, %.4f) en %.1fs",
-        _grounding_model, description[:50], x_pct, y_pct, elapsed,
-    )
+    # Convertir coordonnées fenêtre → coordonnées écran
+    if using_window and window_rect:
+        win_x1, win_y1, win_x2, win_y2 = window_rect
+        win_w = win_x2 - win_x1
+        win_h = win_y2 - win_y1
+        # x_pct/y_pct sont relatifs à la fenêtre, convertir en relatif à l'écran
+        abs_x = win_x1 + x_pct * win_w
+        abs_y = win_y1 + y_pct * win_h
+        x_pct = abs_x / screen_width
+        y_pct = abs_y / screen_height
+        logger.info(
+            "Grounding OK [%s/window] : '%s' → (%.4f, %.4f) en %.1fs",
+            _grounding_model, description[:50], x_pct, y_pct, elapsed,
+        )
+    else:
+        logger.info(
+            "Grounding OK [%s/full] : '%s' → (%.4f, %.4f) en %.1fs",
+            _grounding_model, description[:50], x_pct, y_pct, elapsed,
+        )
 
     return {
         "resolved": True,