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

Strip all standalone bracketed text from LLM responses

Browse Source 
The model paraphrases injected metadata in unpredictable ways, so
targeted regexes can't keep up. Replace them with a single rule: any
[bracketed block] on its own line gets removed, since real roasts
never use standalone brackets.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
AJ Isaacs
2026-02-23 15:24:18 -05:00
parent 942f5ddce7
commit 02b2870f2b
1 changed files with 9 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
cogs/chat.py
View File

@@ -204,16 +204,13 @@ class ChatCog(commands.Cog):
if typing_ctx:
await typing_ctx.__aexit__(None, None, None)
# Strip leaked metadata the LLM may echo back
# Strip leaked metadata the LLM may echo back.
# The LLM paraphrases/reformats injected context in unpredictable ways,
# so nuke any [bracketed block] that sits on its own line — real roasts
# don't use standalone bracketed text.
if response:
response = re.sub(r"\[Server context:[^\]]*\]\n?", "", response)
response = re.sub(r"\[Replying to bot's message:[^\]]*\]\n?", "", response)
response = re.sub(r"\[\w[\w ]* said:[^\]]*\]\n?", "", response)
response = re.sub(r"\[Notes about [^\]]*\]\n?", "", response)
response = re.sub(r"\[[^\]]*'s recent messages:[\s\S]*?\]\n?", "", response)
# Catch reformatted metadata (LLM drops prefix but keeps content)
response = re.sub(r"\[[^\]]*#[a-z-]+[^\]]*(?:drama score|offense)[^\]]*\]\n?", "", response, flags=re.IGNORECASE)
response = response.strip()
response = re.sub(r"^\s*\[[^\]]*\]\s*$", "", response, flags=re.MULTILINE)
response = re.sub(r"\n{2,}", "\n", response).strip()
if not response:
log_channel = discord.utils.get(message.guild.text_channels, name="bcs-log")
@@ -320,13 +317,10 @@ class ChatCog(commands.Cog):
recent_bot_replies=recent_bot_replies,
)
# Strip leaked metadata
# Strip leaked metadata (same catch-all as main chat path)
if response:
response = re.sub(r"\[Server context:[^\]]*\]\n?", "", response)
response = re.sub(r"\[.*?reacted to your message.*?\]\n?", "", response)
response = re.sub(r"\[Your message was:.*?\]\n?", "", response)
response = re.sub(r"\[[^\]]*#[a-z-]+[^\]]*(?:drama score|offense)[^\]]*\]\n?", "", response, flags=re.IGNORECASE)
response = response.strip()
response = re.sub(r"^\s*\[[^\]]*\]\s*$", "", response, flags=re.MULTILINE)
response = re.sub(r"\n{2,}", "\n", response).strip()
if not response:
return