aj/Breehavior-Monitor
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: separate context from new messages so prior-cycle chat doesn't inflate scores

Browse Source 
The conversation analysis was re-scoring old messages alongside new ones,
causing users to get penalized repeatedly for already-scored messages.
A "--- NEW MESSAGES ---" separator now marks which messages are new, and
the prompt instructs the LLM to score only those. Also fixes bot-mention
detection to require an explicit @mention in message text rather than
treating reply-pings as scans (so toxic replies to bot warnings aren't
silently skipped).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
AJ Isaacs
2026-02-25 15:48:02 -05:00
parent 8734f1883b
commit 7417908142
3 changed files with 31 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
cogs/sentiment.py
View File

@@ -87,16 +87,15 @@ class SentimentCog(commands.Cog):
if self.bot.drama_tracker.is_immune(message.author.id):
return
# Messages directed at the bot (mentions, replies) shouldn't be scored
# for toxicity — but @mentions can trigger a scan of recent chat
directed_at_bot = self.bot.user in message.mentions
if not directed_at_bot and message.reference and message.reference.message_id:
ref = message.reference.cached_message
if ref and ref.author.id == self.bot.user.id:
directed_at_bot = True
if directed_at_bot:
# @mention (not just reply-to-bot) triggers a mention scan
if self.bot.user in message.mentions:
# Explicit @mention of the bot triggers a mention scan instead of scoring.
# Reply-pings (Discord auto-adds replied-to user to mentions) should NOT
# trigger scans — and reply-to-bot messages should still be scored normally
# so toxic replies to bot warnings aren't silently skipped.
bot_mentioned_in_text = (
f"<@{self.bot.user.id}>" in (message.content or "")
or f"<@!{self.bot.user.id}>" in (message.content or "")
)
if bot_mentioned_in_text:
mention_config = config.get("mention_scan", {})
if mention_config.get("enabled", True):
await self._maybe_start_mention_scan(message, mention_config)
@@ -166,6 +165,7 @@ class SentimentCog(commands.Cog):
history_messages.reverse() # chronological order
# Combine: history (context) + buffered (new messages to analyze)
new_message_start = len(history_messages)
all_messages = history_messages + messages
# Build msg_id_to_author lookup for reply resolution
@@ -215,6 +215,7 @@ class SentimentCog(commands.Cog):
conversation,
channel_context=channel_context,
user_notes_map=user_notes_map,
new_message_start=new_message_start,
)
if result is None:
@@ -233,6 +234,7 @@ class SentimentCog(commands.Cog):
conversation,
channel_context=channel_context,
user_notes_map=user_notes_map,
new_message_start=new_message_start,
)
if heavy_result is not None:
logger.info(

6
prompts/analysis.txt
View File

@@ -40,8 +40,10 @@ Use the report_analysis tool to report your analysis of the TARGET MESSAGE only.
CONVERSATION-LEVEL ANALYSIS (when given a CONVERSATION BLOCK instead of a single TARGET MESSAGE):
When you receive a full conversation block with multiple users, use the report_conversation_scan tool instead:
- Provide ONE finding per user (not per message) — aggregate their behavior across the conversation.
- Weight their average tone and worst message equally when determining the toxicity_score.
- The conversation block may contain a "--- NEW MESSAGES (score only these) ---" separator. Messages ABOVE the separator are CONTEXT ONLY (already scored in a prior cycle) — do NOT let them inflate scores. Messages BELOW the separator are the NEW messages to score.
- Provide ONE finding per user who has NEW messages (not per message).
- Score based ONLY on the user's NEW messages. Use context messages to understand tone and relationships, but do NOT penalize a user for something they said in the context section.
- If a user's only new message is benign (e.g. "I got the 17.."), score it low regardless of what they said in context.
- Use the same scoring bands (0.0-1.0) as for single messages.
- Quote the worst/most problematic snippet in worst_message (max 100 chars, exact quote).
- Flag off_topic if user's messages are primarily personal drama, not gaming.

14
utils/llm_client.py
View File

@@ -383,12 +383,16 @@ class LLMClient:
def _format_conversation_block(
messages: list[tuple[str, str, datetime, str | None]],
now: datetime | None = None,
new_message_start: int | None = None,
) -> str:
"""Format messages as a compact timestamped chat block.
Each tuple is (username, content, timestamp, reply_to_username).
Consecutive messages from the same user collapse to indented lines.
Replies shown as ``username → replied_to:``.
If *new_message_start* is given, a separator is inserted before that
index so the LLM can distinguish context from new messages.
"""
if now is None:
now = datetime.now(timezone.utc)
@@ -396,7 +400,12 @@ class LLMClient:
lines = [f"[Current time: {now.strftime('%I:%M %p')}]", ""]
last_user = None
for username, content, ts, reply_to in messages:
for idx, (username, content, ts, reply_to) in enumerate(messages):
if new_message_start is not None and idx == new_message_start:
lines.append("")
lines.append("--- NEW MESSAGES (score only these) ---")
lines.append("")
last_user = None # reset collapse so first new msg gets full header
delta = now - ts.replace(tzinfo=timezone.utc) if ts.tzinfo is None else now - ts
rel = LLMClient._format_relative_time(delta)
@@ -425,12 +434,13 @@ class LLMClient:
mention_context: str = "",
channel_context: str = "",
user_notes_map: dict[str, str] | None = None,
new_message_start: int | None = None,
) -> dict | None:
"""Analyze a conversation block in one call, returning per-user findings."""
if not messages:
return None
convo_block = self._format_conversation_block(messages)
convo_block = self._format_conversation_block(messages, new_message_start=new_message_start)
user_content = f"=== CONVERSATION BLOCK ===\n{convo_block}\n\n"
if user_notes_map: