Fix ACTION6 error: check available_actions, graceful fallback to CNN

app.py

@@ -521,7 +521,17 @@ class FluidAgent:
             'candidates': [(signal, answer, confidence)] if answer is not None else []
         }
         
-        if answer is not None and confidence > 0.40 and target_cell is not None:
+        # Check what actions are actually available
+        avail_ids = set()
+        if available_actions:
+            avail_ids = set(int(a.value if hasattr(a, 'value') else a) 
+                           for a in available_actions)
+        else:
+            avail_ids = set(range(1, 7))
+        
+        # If we have a strong analytic answer and ACTION6 exists, use it
+        if (answer is not None and confidence > 0.40 and 
+            target_cell is not None and 6 in avail_ids):
             # Im found hypothesis, Re found cell → ACTION6
             r, c, _ = target_cell
             H, W = grid.shape
@@ -535,15 +545,19 @@ class FluidAgent:
             
             chosen_id = 6
         else:
-            # CNN fallback
+            # CNN fallback (or analytic without click action)
+            if answer is not None and confidence > 0.40:
+                # We have strong answer but no ACTION6 - pick best alternative
+                meta['source'] = 'analytic_fallback'
+                meta['note'] = f"Confidence {confidence:.2f} but ACTION6 not available"
+            else:
+                meta['source'] = 'cnn'
+            
             with torch.no_grad():
                 logits = self.model(feat.unsqueeze(0)).squeeze(0)
-                avail = list(range(1, 7))
-                if available_actions:
-                    avail = [int(a.value if hasattr(a, 'value') else a)
-                             for a in available_actions if
-                             int(a.value if hasattr(a, 'value') else a) <= 6]
-                indices = [m-1 for m in avail if 1 <= m <= 6]
+                
+                # Mask to only available actions
+                indices = [m-1 for m in avail_ids if 1 <= m <= 6]
                 masked = torch.full((6,), float('-inf'))
                 for i in indices:
                     masked[i] = logits[i]
@@ -555,8 +569,11 @@ class FluidAgent:
                 cnn_action_idx = np.random.choice(6, p=probs)
             
             chosen_id = cnn_action_idx + 1
-            meta['source'] = 'cnn'
             meta['probs'] = probs.tolist()
+            
+            # Make sure chosen action is actually available
+            if chosen_id not in avail_ids and avail_ids:
+                chosen_id = list(avail_ids)[0]
         
         self.prev_feat = feat
         self.prev_action = chosen_id - 1
@@ -672,6 +689,11 @@ def _run_agent(game_id, api_key, max_steps):
             
             action, meta = _agent.choose(grid, avail, levels=levels, state=state)
             
+            # Add available actions to meta for debugging
+            if avail:
+                meta['available_actions'] = [int(a.value if hasattr(a, 'value') else a) 
+                                            for a in avail]
+            
             diff = (grid != prev_grid) if prev_grid is not None else None
             prev_grid = grid.copy()
             
@@ -783,9 +805,15 @@ def pull_frame():
     reasoning = meta.get('reasoning', [])
     hyp_text = '\n'.join(reasoning[:2]) if reasoning else 'none'
     
+    avail_actions = meta.get('available_actions', [])
+    avail_str = f"Available: {avail_actions}" if avail_actions else ""
+    
+    source_label = {'analytic': 'Analytic', 'analytic_fallback': 'Analytic (no click)', 'cnn': 'CNN'}
+    
     _latest['status'] = (
-        f"{source_emoji} **{'Analytic' if source == 'analytic' else 'CNN'}** | "
-        f"Step {step} | Action {action}\n\n{hyp_text}"
+        f"{source_emoji} **{source_label.get(source, source)}** | "
+        f"Step {step} | Action {action}\n\n"
+        f"{hyp_text}\n\n{avail_str}"
     )
     
     return (_latest['grid_img'], _latest['hyp_img'], _latest['cand_img'], _latest['status'])