fix: improve semantic overlap validation, strip task prefixes, and refine fallback prompts

app.py

@@ -133,15 +133,41 @@ def append_history(state, displayed_task: str, raw_task: str) -> None:
 def is_semantic_repeat(new_task: str, history: list) -> bool:
     if not history:
         return False
-    new_words = set(new_task.lower().split())
+        
+    stop_words = {"i", "need", "to", "a", "the", "an", "and", "or", "but", "with", "for", "of", "on", "at", "by", "start", "with"}
+    
+    def get_stems(task_str):
+        words = task_str.lower().split()
+        stems = []
+        for w in words:
+            w = w.strip(".,!?\"'();:")
+            if w in stop_words or not w:
+                continue
+            # Simple suffix stemming
+            if len(w) > 4:
+                if w.endswith("ing"):
+                    w = w[:-3]
+                elif w.endswith("ies"):
+                    w = w[:-3] + "y"
+                elif w.endswith("es"):
+                    w = w[:-2]
+                elif w.endswith("ed"):
+                    w = w[:-2]
+                elif w.endswith("s") and not w.endswith("ss"):
+                    w = w[:-1]
+            stems.append(w)
+        return set(stems)
+
+    new_words = get_stems(new_task)
     if not new_words:
         return False
+        
     for past in history[-3:]:
-        past_words = set(past.lower().split())
+        past_words = get_stems(past)
         if not past_words:
             continue
         overlap = len(new_words & past_words) / max(len(new_words | past_words), 1)
-        if overlap > 0.7:  # 70% word overlap = semantic repeat
+        if overlap > 0.6:  # 60% overlap on key word stems = semantic repeat
             return True
     return False
 
@@ -159,6 +185,21 @@ def clean_and_validate_task(raw_output: str, history: list, skipped_history: lis
     if not cleaned:
         return "", False
         
+    # Strip conversational prefixes
+    prefixes_to_strip = [
+        "i need to ", "i should ", "i have to ", "let's ", "we should ",
+        "please ", "you need to ", "you should "
+    ]
+    cleaned_lower = cleaned.lower()
+    for prefix in prefixes_to_strip:
+        if cleaned_lower.startswith(prefix):
+            cleaned = cleaned[len(prefix):]
+            cleaned_lower = cleaned.lower()
+            
+    # Capitalize the first letter again
+    if cleaned:
+        cleaned = cleaned[0].upper() + cleaned[1:]
+        
     # Check for prompt leakage and scaffolding words
     invalid_keywords = [
         "completed tasks", "failures", "system", "user", "assistant", 
@@ -251,14 +292,17 @@ async def generate_atomic_task(goal: str, previous_failures: int, history: list
             
             user_content = f"Goal: {goal}\nCompleted Tasks: {history_str}\nFailures: {previous_failures}\n"
             if last_skipped:
-                user_content += f"CRITICAL: The user explicitly skipped '{last_skipped}'. Do NOT output it again. Generate an ALTERNATIVE physical starting action.\n"
+                user_content += (
+                    f"CRITICAL: The user explicitly skipped '{last_skipped}'. Do NOT output it again. "
+                    f"Stay on the current sub-topic or context of the last step, but generate an ALTERNATIVE physical starting action."
+                )
             elif last_task:
-                user_content += f"CRITICAL: Do NOT output '{last_task}'. Output the strictly NEXT new physical step.\n"
+                user_content += f"CRITICAL: Do NOT output '{last_task}'. Stay on the same sub-topic or context, but output the strictly NEXT new physical step.\n"
             else:
                 user_content += "Output the NEXT step.\n"
             
             messages = [
-                {"role": "system", "content": "You are a cognitive pacemaker. Focus strictly on the physical goal and break it down into an extremely tiny, atomic physical action under 8 words. Ignore mental blocks or feelings of being stuck and output ONLY the single physical action step."},
+                {"role": "system", "content": "You are a cognitive pacemaker. Focus strictly on the physical goal and break it down into an extremely tiny, atomic physical action under 8 words. Stay contextually coherent with the user's completed tasks, do NOT jump to unrelated sub-tasks, and output ONLY the single physical action step."},
                 {"role": "user", "content": user_content}
             ]
             fallback_res = await run_minicpm(