feat(perception): implement Vision-Critic validation gate to block LLM hallucinations via cropped screenshot validation

GramAddict/core/perception/intent_resolver.py

@@ -19,7 +19,7 @@ class IntentResolver:
     """
 
     def resolve(
-        self, intent_description: str, candidates: List[SpatialNode], screen_height: int = 2400
+        self, intent_description: str, candidates: List[SpatialNode], screen_height: int = 2400, device=None
     ) -> Optional[SpatialNode]:
         """
         Finds the best matching node for a given intent autonomously.
@@ -121,10 +121,86 @@ class IntentResolver:
             data = json.loads(res)
             idx = data.get("selected_index")
             if idx is not None and 0 <= idx < len(filtered_candidates):
-                return filtered_candidates[idx]
+                proposed_node = filtered_candidates[idx]
+                if device and not self._visual_critic(intent_description, proposed_node, device):
+                    return None
+                return proposed_node
         except Exception as e:
             import logging
 
             logging.getLogger(__name__).warning(f"⚠️ [IntentResolver] VLM resolution failed ({e}).")
 
         return None
+
+    def _visual_critic(self, intent_description: str, node: SpatialNode, device) -> bool:
+        """
+        Validates the proposed node visually to prevent blind hallucinations.
+        Crops the UI snippet of the bounding box and asks the VLM to confirm it matches the intent.
+        """
+        import logging
+        from GramAddict.core.llm_provider import query_telepathic_llm
+        from GramAddict.core.config import Config
+        import base64
+        from io import BytesIO
+
+        logger = logging.getLogger(__name__)
+
+        try:
+            # 1. Crop the node's visual representation from the screen with padding
+            pad = 20
+            img = device.deviceV2.screenshot()
+            width, height = img.size
+            
+            x1 = max(0, node.x1 - pad)
+            y1 = max(0, node.y1 - pad)
+            x2 = min(width, node.x2 + pad)
+            y2 = min(height, node.y2 + pad)
+            
+            cropped = img.crop((x1, y1, x2, y2))
+            
+            # Skip validation if the area is bizarrely small
+            if (x2 - x1) < 10 or (y2 - y1) < 10:
+                return True
+            
+            buffered = BytesIO()
+            cropped.save(buffered, format="JPEG")
+            img_b64 = base64.b64encode(buffered.getvalue()).decode('utf-8')
+            
+            cfg = Config()
+            model = getattr(cfg.args, "ai_telepathic_model", "llava:latest")
+            url = getattr(cfg.args, "ai_telepathic_url", "http://localhost:11434/api/generate")
+
+            prompt = (
+                f"You are a strict UI Visual Validator.\n"
+                f"Look at this cropped image of a UI element.\n"
+                f"Does this element visually look like the correct target for the user intent: '{intent_description}'?\n"
+                f"Look at icons, text, and general appearance. Be highly skeptical.\n\n"
+                f"Reply ONLY with a valid JSON object strictly matching this schema:\n"
+                f'{{"valid": true}} or {{"valid": false}}'
+            )
+
+            logger.info(f"⚖️ [Visual Critic] Validating proposal (id='{node.resource_id}', text='{node.text}') for intent '{intent_description}'...")
+            res = query_telepathic_llm(
+                model=model,
+                url=url,
+                system_prompt="Strict visual JSON critic.",
+                user_prompt=prompt,
+                use_local_edge=True,
+                images_b64=[img_b64]
+            )
+            
+            import json
+            data = json.loads(res)
+            is_valid = data.get("valid", True)
+            
+            if is_valid:
+                logger.info("✅ [Visual Critic] VLM confirmed the visual target.")
+            else:
+                logger.warning(f"❌ [Visual Critic] VLM REJECTED the target (Hallucination blocked).")
+                
+            return is_valid
+            
+        except Exception as e:
+            logger.warning(f"⚠️ [Visual Critic] Validation failed, falling back to accept: {e}")
+            return True
+