fix: anchor match sur screenshot entier + proximité élément SomEngine

Le template matching du crop anchor contre les régions YOLO échouait
car l'anchor (150x150) est plus grand que les éléments détectés.
Maintenant : match sur le screenshot entier → centre du match →
élément SomEngine le plus proche (max 100px).

Fonctionne pour les icônes mais limité par la taille du crop
(150x150 de barre de titre matche à plusieurs endroits).

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

agent_v0/server_v1/api_stream.py

@@ -3532,7 +3532,8 @@ def _resolve_by_som(
     # Pour les icônes sans texte : comparer le crop de référence contre
     # chaque région YOLO détectée par SomEngine.
     anchor_b64 = target_spec.get("anchor_image_base64", "")
-    if anchor_b64 and not anchor_label:
+    by_text = target_spec.get("by_text", "").strip()
+    if anchor_b64 and (not anchor_label or not by_text):
         try:
             import cv2
             import numpy as np
@@ -3546,51 +3547,51 @@ def _resolve_by_som(
             screenshot_cv = cv2.imread(screenshot_path, cv2.IMREAD_GRAYSCALE)
 
             if anc_img is not None and screenshot_cv is not None:
-                best_elem = None
+                # Template matching de l'anchor sur le SCREENSHOT ENTIER
-                best_score = 0.0
+                # (pas sur les régions individuelles — l'anchor est souvent plus grand)
                 anc_h, anc_w = anc_img.shape[:2]
+                if screenshot_cv.shape[0] >= anc_h and screenshot_cv.shape[1] >= anc_w:
+                    res = cv2.matchTemplate(screenshot_cv, anc_img, cv2.TM_CCOEFF_NORMED)
+                    _, max_score, _, max_loc = cv2.minMaxLoc(res)
 
-                for elem in som_result.elements:
+                    if max_score >= 0.5:
-                    x1, y1, x2, y2 = elem.bbox
+                        # Centre du match
-                    # Agrandir la zone de 20% pour tolérer les différences
+                        match_cx = max_loc[0] + anc_w // 2
-                    margin_x = int((x2 - x1) * 0.2)
+                        match_cy = max_loc[1] + anc_h // 2
-                    margin_y = int((y2 - y1) * 0.2)
-                    rx1 = max(0, x1 - margin_x)
-                    ry1 = max(0, y1 - margin_y)
-                    rx2 = min(screenshot_cv.shape[1], x2 + margin_x)
-                    ry2 = min(screenshot_cv.shape[0], y2 + margin_y)
-                    region = screenshot_cv[ry1:ry2, rx1:rx2]
 
-                    if region.shape[0] < anc_h or region.shape[1] < anc_w:
+                        # Trouver l'élément SomEngine le plus proche du centre du match
-                        continue
+                        best_elem = None
+                        best_dist = float("inf")
+                        for elem in som_result.elements:
+                            cx, cy = elem.center
+                            dist = ((match_cx - cx) ** 2 + (match_cy - cy) ** 2) ** 0.5
+                            if dist < best_dist:
+                                best_dist = dist
+                                best_elem = elem
 
-                    res = cv2.matchTemplate(region, anc_img, cv2.TM_CCOEFF_NORMED)
+                        if best_elem and best_dist < 100:  # Max 100px de distance
-                    _, score, _, _ = cv2.minMaxLoc(res)
+                            elapsed = time.time() - t0
-                    if score > best_score:
+                            cx_norm, cy_norm = best_elem.center_norm
-                        best_score = score
+                            logger.info(
-                        best_elem = elem
+                                "SoM resolve ANCHOR : match crop score=%.3f → "
-
+                                "elem '#%d %s' (dist=%.0fpx) → (%.4f, %.4f) en %.1fs",
-                if best_elem and best_score >= 0.6:
+                                max_score, best_elem.id, best_elem.label,
-                    elapsed = time.time() - t0
+                                best_dist, cx_norm, cy_norm, elapsed,
-                    cx_norm, cy_norm = best_elem.center_norm
+                            )
-                    logger.info(
+                            return {
-                        "SoM resolve ANCHOR : match crop '#%d' score=%.3f → (%.4f, %.4f) en %.1fs",
+                                "resolved": True,
-                        best_elem.id, best_score, cx_norm, cy_norm, elapsed,
+                                "method": "som_anchor_match",
-                    )
+                                "x_pct": round(cx_norm, 6),
-                    return {
+                                "y_pct": round(cy_norm, 6),
-                        "resolved": True,
+                                "matched_element": {
-                        "method": "som_anchor_match",
+                                    "label": best_elem.label or f"icon #{best_elem.id}",
-                        "x_pct": round(cx_norm, 6),
+                                    "type": best_elem.source,
-                        "y_pct": round(cy_norm, 6),
+                                    "role": "som_anchor_match",
-                        "matched_element": {
+                                    "confidence": max_score,
-                            "label": best_elem.label or f"icon #{best_elem.id}",
+                                    "som_id": best_elem.id,
-                            "type": best_elem.source,
+                                },
-                            "role": "som_anchor_match",
+                                "score": max_score,
-                            "confidence": best_score,
+                            }
-                            "som_id": best_elem.id,
-                        },
-                        "score": best_score,
-                    }
         except ImportError:
             pass
         except Exception as e:

agent_v0/server_v1/stream_processor.py

@@ -502,14 +502,24 @@ def _vlm_identify_element(anchor_b64: str, window_title: str = "") -> str:
                 if raw.lower().startswith(prefix.lower()):
                     raw = raw[len(prefix):]
                     break
+            # Rejeter les réponses qui sont du bavardage, pas un label
+            reject_patterns = (
+                "several", "multiple", "various", "image",
+                "I can", "there are", "there is", "elements",
+                "the following", "here are",
+            )
+            if any(p in raw.lower()[:30] for p in reject_patterns):
+                logger.debug("VLM identify : réponse bavarde rejetée (raw='%s')", raw[:60])
+                return ""
+
             # Prendre les 5 premiers mots utiles
             words = raw.split()[:5]
             label = " ".join(words).strip('",.\' ').rstrip(".")
-            if label and 2 <= len(label) <= 60:
+            if label and 2 <= len(label) <= 40:
                 logger.info("VLM identify element : '%s'", label)
                 return label
             else:
-                logger.debug("VLM identify : label trop court ou vide après nettoyage (raw='%s')", raw[:80])
+                logger.debug("VLM identify : label trop court/long après nettoyage (raw='%s')", raw[:80])
     except Exception as e:
         logger.debug("VLM identify element échoué : %s", e)