fix: sanitize note_updates before storing in sentiment pipeline

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

cogs/chat.py

@@ -171,6 +171,8 @@ class ChatCog(commands.Cog):
             # Update profile if warranted
             profile_update = result.get("profile_update")
             if profile_update:
+                # Sanitize before storing — strips any quoted toxic language
+                profile_update = await self.bot.llm.sanitize_notes(profile_update)
                 self.bot.drama_tracker.set_user_profile(user_id, profile_update)
                 self._dirty_users.add(user_id)
 

cogs/sentiment/__init__.py

@@ -469,13 +469,14 @@ class SentimentCog(commands.Cog):
 
         # Note update — route to memory system
         if note_update:
-            # Still update the legacy notes for backward compat with analysis prompt
-            self.bot.drama_tracker.update_user_notes(user_id, note_update)
+            # Sanitize before storing — strips any quoted toxic language
+            sanitized = await self.bot.llm.sanitize_notes(note_update)
+            self.bot.drama_tracker.update_user_notes(user_id, sanitized)
             self._dirty_users.add(user_id)
             # Also save as an expiring memory (7d default for passive observations)
             asyncio.create_task(self.bot.db.save_memory(
                 user_id=user_id,
-                memory=note_update[:500],
+                memory=sanitized[:500],
                 topics=db_topic_category or "general",
                 importance="medium",
                 expires_at=datetime.now(timezone.utc) + timedelta(days=7),

prompts/analysis.txt

@@ -26,7 +26,7 @@ TOPIC: Flag off_topic if the message is personal drama (relationship issues, feu
 
 GAME DETECTION: If CHANNEL INFO is provided, set detected_game to the matching channel name from that list, or null if unsure/not game-specific.
 
-USER NOTES: If provided, use to calibrate (e.g. if notes say "uses heavy profanity casually", profanity alone should score lower). Add a note_update only for genuinely new behavioral observations; null otherwise.
+USER NOTES: If provided, use to calibrate (e.g. if notes say "uses heavy profanity casually", profanity alone should score lower). Add a note_update only for genuinely new behavioral observations; null otherwise. NEVER quote or repeat toxic/offensive language in note_update — describe patterns abstractly (e.g. "directed a personal insult at another user", NOT "called someone a [slur]").
 
 RULE ENFORCEMENT: If SERVER RULES are provided, report clearly violated rule numbers in violated_rules. Only flag clear violations, not borderline.
 

utils/llm_client.py

@@ -86,7 +86,7 @@ ANALYSIS_TOOL = {
                 },
                 "note_update": {
                     "type": ["string", "null"],
-                    "description": "Brief new observation about this user's style/behavior for future reference, or null if nothing new.",
+                    "description": "Brief new observation about this user's style/behavior for future reference, or null if nothing new. NEVER quote toxic language — describe patterns abstractly (e.g. 'uses personal insults when frustrated').",
                 },
                 "detected_game": {
                     "type": ["string", "null"],
@@ -189,7 +189,7 @@ CONVERSATION_TOOL = {
                             },
                             "note_update": {
                                 "type": ["string", "null"],
-                                "description": "New observation about this user's pattern, or null.",
+                                "description": "New observation about this user's pattern, or null. NEVER quote toxic language — describe patterns abstractly.",
                             },
                             "detected_game": {
                                 "type": ["string", "null"],
@@ -977,6 +977,51 @@ class LLMClient:
             "profile_update": profile_update,
         }
 
+    async def sanitize_notes(self, notes: str) -> str:
+        """Rewrite user notes to remove any quoted toxic/offensive language.
+
+        Returns the sanitized notes string, or the original on failure.
+        """
+        if not notes or len(notes.strip()) == 0:
+            return notes
+
+        system_prompt = (
+            "Rewrite the following user behavior notes. Remove any quoted offensive language, "
+            "slurs, or profanity. Replace toxic quotes with abstract descriptions of the behavior "
+            "(e.g. 'directed a personal insult at another user' instead of quoting the insult). "
+            "Preserve all non-toxic observations, timestamps, and behavioral patterns exactly. "
+            "Return ONLY the rewritten notes, nothing else."
+        )
+        user_content = notes
+        if self._no_think:
+            user_content += "\n/no_think"
+
+        t0 = time.monotonic()
+        async with self._semaphore:
+            try:
+                temp_kwargs = {"temperature": 0.1} if self._supports_temperature else {}
+                response = await self._client.chat.completions.create(
+                    model=self.model,
+                    messages=[
+                        {"role": "system", "content": system_prompt},
+                        {"role": "user", "content": user_content},
+                    ],
+                    **temp_kwargs,
+                    max_completion_tokens=1024,
+                )
+                elapsed = int((time.monotonic() - t0) * 1000)
+                result = response.choices[0].message.content
+                if result and result.strip():
+                    self._log_llm("sanitize_notes", elapsed, True, notes[:300], result[:300])
+                    return result.strip()
+                self._log_llm("sanitize_notes", elapsed, False, notes[:300], error="Empty response")
+                return notes
+            except Exception as e:
+                elapsed = int((time.monotonic() - t0) * 1000)
+                logger.error("LLM sanitize_notes error: %s", e)
+                self._log_llm("sanitize_notes", elapsed, False, notes[:300], error=str(e))
+                return notes
+
     async def analyze_image(
         self,
         image_bytes: bytes,