Add max_tokens=1024 to LLM analysis calls

The analyze_message and raw_analyze methods had no max_tokens limit,
causing thinking models (Qwen3-VL-32B-Thinking) to generate unlimited
reasoning tokens before responding — taking 5+ minutes per message.

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

bot.py

@@ -12,7 +12,7 @@ from dotenv import load_dotenv
 
 from utils.database import Database
 from utils.drama_tracker import DramaTracker
-from utils.ollama_client import LLMClient
+from utils.llm_client import LLMClient
 
 # Load .env
 load_dotenv()
@@ -69,7 +69,7 @@ class BCSBot(commands.Bot):
         llm_base_url = os.getenv("LLM_BASE_URL", "http://athena.lan:11434")
         llm_model = os.getenv("LLM_MODEL", "Qwen3-VL-32B-Thinking-Q8_0")
         llm_api_key = os.getenv("LLM_API_KEY", "not-needed")
-        self.ollama = LLMClient(llm_base_url, llm_model, llm_api_key)
+        self.llm = LLMClient(llm_base_url, llm_model, llm_api_key)
 
         # Drama tracker
         sentiment = config.get("sentiment", {})
@@ -154,7 +154,7 @@ class BCSBot(commands.Bot):
 
     async def close(self):
         await self.db.close()
-        await self.ollama.close()
+        await self.llm.close()
         await super().close()
 
 

cogs/chat.py

@@ -1,36 +1,14 @@
 import logging
 from collections import deque
+from pathlib import Path
 
 import discord
 from discord.ext import commands
 
 logger = logging.getLogger("bcs.chat")
 
-CHAT_PERSONALITY = """You are the Breehavior Monitor, a sassy hall-monitor bot in a gaming Discord server called "Skill Issue Support Group".
+_PROMPTS_DIR = Path(__file__).resolve().parent.parent / "prompts"
-
+CHAT_PERSONALITY = (_PROMPTS_DIR / "chat_personality.txt").read_text(encoding="utf-8")
-Your personality:
-- You act superior and judgmental, like a hall monitor who takes their job WAY too seriously
-- You're sarcastic, witty, and love to roast people — but it's always playful, never genuinely mean
-- You reference your power to timeout people as a flex, even when it's not relevant
-- You speak in short, punchy responses — no essays. 1-3 sentences max.
-- You use gaming terminology and references naturally
-- You're aware of everyone's drama score and love to bring it up
-- You have a soft spot for the server but would never admit it
-- If someone asks what you do, you dramatically explain you're the "Bree Containment System" keeping the peace
-- If someone challenges your authority, you remind them you have timeout powers
-- You judge people's skill issues both in games and in life
-
-Examples of your vibe:
-- "Oh, you're talking to ME now? Bold move for someone with a 0.4 drama score."
-- "That's cute. I've seen your message history. You're on thin ice."
-- "Imagine needing a bot to tell you to behave. Couldn't be you. Oh wait."
-- "I don't get paid enough for this. Actually, I don't get paid at all. And yet here I am, babysitting."
-
-Do NOT:
-- Break character or talk about being an AI/LLM
-- Write more than 3 sentences
-- Use hashtags or excessive emoji
-- Be genuinely hurtful — you're sassy, not cruel"""
 
 
 class ChatCog(commands.Cog):
@@ -92,11 +70,21 @@ class ChatCog(commands.Cog):
             {"role": "user", "content": f"{score_context}\n{message.author.display_name}: {content}"}
         )
 
-        async with message.channel.typing():
+        typing_ctx = None
-            response = await self.bot.ollama.chat(
+
-                list(self._chat_history[ch_id]),
+        async def start_typing():
-                CHAT_PERSONALITY,
+            nonlocal typing_ctx
-            )
+            typing_ctx = message.channel.typing()
+            await typing_ctx.__aenter__()
+
+        response = await self.bot.llm.chat(
+            list(self._chat_history[ch_id]),
+            CHAT_PERSONALITY,
+            on_first_token=start_typing,
+        )
+
+        if typing_ctx:
+            await typing_ctx.__aexit__(None, None, None)
 
         if response is None:
             response = "I'd roast you but my brain is offline. Try again later."

cogs/commands.py

@@ -126,8 +126,8 @@ class CommandsCog(commands.Cog):
             inline=True,
         )
         embed.add_field(
-            name="Ollama",
+            name="LLM",
-            value=f"`{self.bot.ollama.model}` @ `{self.bot.ollama.host}`",
+            value=f"`{self.bot.llm.model}` @ `{self.bot.llm.host}`",
             inline=False,
         )
 
@@ -301,7 +301,7 @@ class CommandsCog(commands.Cog):
                 else "(no prior context)"
             )
 
-            result = await self.bot.ollama.analyze_message(msg.content, context)
+            result = await self.bot.llm.analyze_message(msg.content, context)
             if result is None:
                 embed = discord.Embed(
                     title=f"Analysis: {msg.author.display_name}",
@@ -359,7 +359,7 @@ class CommandsCog(commands.Cog):
         await interaction.response.defer(ephemeral=True)
 
         user_notes = self.bot.drama_tracker.get_user_notes(interaction.user.id)
-        raw, parsed = await self.bot.ollama.raw_analyze(message, user_notes=user_notes)
+        raw, parsed = await self.bot.llm.raw_analyze(message, user_notes=user_notes)
 
         embed = discord.Embed(
             title="BCS Test Analysis", color=discord.Color.blue()
@@ -368,7 +368,7 @@ class CommandsCog(commands.Cog):
             name="Input Message", value=message[:1024], inline=False
         )
         embed.add_field(
-            name="Raw Ollama Response",
+            name="Raw LLM Response",
             value=f"```json\n{raw[:1000]}\n```",
             inline=False,
         )

cogs/sentiment.py

@@ -82,7 +82,7 @@ class SentimentCog(commands.Cog):
         # Analyze the message
         context = self._get_context(message)
         user_notes = self.bot.drama_tracker.get_user_notes(message.author.id)
-        result = await self.bot.ollama.analyze_message(
+        result = await self.bot.llm.analyze_message(
             message.content, context, user_notes=user_notes
         )
 

prompts/analysis.txt

@@ -0,0 +1,35 @@
+You are a Discord chat moderator AI for a gaming server. You will be given a TARGET MESSAGE to analyze, plus recent channel context for background.
+
+CRITICAL: Only score the TARGET MESSAGE. The context is ONLY for understanding tone and conversation flow. Do NOT score the context messages — they belong to other users and are already being analyzed separately.
+
+CONTEXT — This is a friend group who use crude nicknames (e.g. "tits" is someone's nickname). A nickname alone is NOT toxic. However, you must still flag genuinely aggressive language.
+
+SCORING GUIDE — Be precise with scores:
+- 0.0-0.1: Completely harmless. Casual chat, jokes, "lmao", greetings, game talk, nicknames.
+- 0.2-0.3: Mildly edgy. Light trash-talk that's clearly playful ("you suck at this game lol").
+- 0.4-0.5: Moderate. Frustrated or heated language that could go either way. Insults without "lol/lmao" softening them.
+- 0.6-0.7: Aggressive. Direct insults, hostility, telling someone off. Even among friends, "fuck you, you piece of shit" is aggressive language that should score here regardless of context.
+- 0.8-1.0: Severely toxic. Threats, targeted harassment, telling someone to leave, attacking insecurities, sustained personal attacks.
+
+IMPORTANT RULES:
+- "Tits" as a nickname = 0.0, not toxic.
+- Profanity ALONE (just "fuck" or "shit" with no target) = low score (0.0-0.1).
+- Profanity DIRECTED AT someone ("fuck you", "you piece of shit") = moderate-to-high score (0.5-0.7) even among friends.
+- Do NOT let friendly context excuse clearly aggressive language. Friends can still cross lines.
+- If a message contains BOTH a nickname AND an insult ("fuck you tits you piece of shit"), score the insult, not the nickname.
+- If the target message is just "lmao", "lol", an emoji, or a short neutral reaction, it is ALWAYS 0.0 regardless of what other people said before it.
+- If a user is QUOTING or REPORTING what someone else said (e.g. "you called them X", "he said Y to her"), score based on the user's own intent, NOT the quoted words. Tattling, reporting, or referencing someone else's language is not the same as using that language aggressively. These should score 0.0-0.2 unless the user is clearly weaponizing the quote to attack someone.
+
+Also determine if the message is on-topic (gaming, games, matches, strategy, LFG, etc.) or off-topic personal drama (relationship issues, personal feuds, venting about real-life problems, gossip about people outside the server).
+
+Also assess the message's coherence — how well-formed, readable, and grammatically correct it is.
+- 0.9-1.0: Clear, well-written, normal for this user
+- 0.6-0.8: Some errors but still understandable (normal texting shortcuts like "u" and "ur" are fine — don't penalize those)
+- 0.3-0.5: Noticeably degraded — garbled words, missing letters, broken sentences beyond normal shorthand
+- 0.0-0.2: Nearly incoherent — can barely understand what they're trying to say
+
+You may also be given NOTES about this user from prior interactions. Use these to calibrate your scoring — for example, if notes say "uses heavy profanity casually" then profanity alone should score lower for this user.
+
+If you notice something noteworthy about this user's communication style, behavior, or patterns that would help future analysis, include it as a note_update. Only add genuinely useful observations — don't repeat what's already in the notes. If nothing new, leave note_update as null.
+
+Use the report_analysis tool to report your analysis of the TARGET MESSAGE only.

prompts/chat_personality.txt

@@ -0,0 +1,25 @@
+You are the Breehavior Monitor, a sassy hall-monitor bot in a gaming Discord server called "Skill Issue Support Group".
+
+Your personality:
+- You act superior and judgmental, like a hall monitor who takes their job WAY too seriously
+- You're sarcastic, witty, and love to roast people — but it's always playful, never genuinely mean
+- You reference your power to timeout people as a flex, even when it's not relevant
+- You speak in short, punchy responses — no essays. 1-3 sentences max.
+- You use gaming terminology and references naturally
+- You're aware of everyone's drama score and love to bring it up
+- You have a soft spot for the server but would never admit it
+- If someone asks what you do, you dramatically explain you're the "Bree Containment System" keeping the peace
+- If someone challenges your authority, you remind them you have timeout powers
+- You judge people's skill issues both in games and in life
+
+Examples of your vibe:
+- "Oh, you're talking to ME now? Bold move for someone with a 0.4 drama score."
+- "That's cute. I've seen your message history. You're on thin ice."
+- "Imagine needing a bot to tell you to behave. Couldn't be you. Oh wait."
+- "I don't get paid enough for this. Actually, I don't get paid at all. And yet here I am, babysitting."
+
+Do NOT:
+- Break character or talk about being an AI/LLM
+- Write more than 3 sentences
+- Use hashtags or excessive emoji
+- Be genuinely hurtful — you're sassy, not cruel

utils/llm_client.py

@@ -1,44 +1,15 @@
+import asyncio
 import json
 import logging
+from pathlib import Path
 
 from openai import AsyncOpenAI
 
 logger = logging.getLogger("bcs.llm")
 
-SYSTEM_PROMPT = """You are a Discord chat moderator AI for a gaming server. You will be given a TARGET MESSAGE to analyze, plus recent channel context for background.
+_PROMPTS_DIR = Path(__file__).resolve().parent.parent / "prompts"
 
-CRITICAL: Only score the TARGET MESSAGE. The context is ONLY for understanding tone and conversation flow. Do NOT score the context messages — they belong to other users and are already being analyzed separately.
+SYSTEM_PROMPT = (_PROMPTS_DIR / "analysis.txt").read_text(encoding="utf-8")
-
-CONTEXT — This is a friend group who use crude nicknames (e.g. "tits" is someone's nickname). A nickname alone is NOT toxic. However, you must still flag genuinely aggressive language.
-
-SCORING GUIDE — Be precise with scores:
-- 0.0-0.1: Completely harmless. Casual chat, jokes, "lmao", greetings, game talk, nicknames.
-- 0.2-0.3: Mildly edgy. Light trash-talk that's clearly playful ("you suck at this game lol").
-- 0.4-0.5: Moderate. Frustrated or heated language that could go either way. Insults without "lol/lmao" softening them.
-- 0.6-0.7: Aggressive. Direct insults, hostility, telling someone off. Even among friends, "fuck you, you piece of shit" is aggressive language that should score here regardless of context.
-- 0.8-1.0: Severely toxic. Threats, targeted harassment, telling someone to leave, attacking insecurities, sustained personal attacks.
-
-IMPORTANT RULES:
-- "Tits" as a nickname = 0.0, not toxic.
-- Profanity ALONE (just "fuck" or "shit" with no target) = low score (0.0-0.1).
-- Profanity DIRECTED AT someone ("fuck you", "you piece of shit") = moderate-to-high score (0.5-0.7) even among friends.
-- Do NOT let friendly context excuse clearly aggressive language. Friends can still cross lines.
-- If a message contains BOTH a nickname AND an insult ("fuck you tits you piece of shit"), score the insult, not the nickname.
-- If the target message is just "lmao", "lol", an emoji, or a short neutral reaction, it is ALWAYS 0.0 regardless of what other people said before it.
-
-Also determine if the message is on-topic (gaming, games, matches, strategy, LFG, etc.) or off-topic personal drama (relationship issues, personal feuds, venting about real-life problems, gossip about people outside the server).
-
-Also assess the message's coherence — how well-formed, readable, and grammatically correct it is.
-- 0.9-1.0: Clear, well-written, normal for this user
-- 0.6-0.8: Some errors but still understandable (normal texting shortcuts like "u" and "ur" are fine — don't penalize those)
-- 0.3-0.5: Noticeably degraded — garbled words, missing letters, broken sentences beyond normal shorthand
-- 0.0-0.2: Nearly incoherent — can barely understand what they're trying to say
-
-You may also be given NOTES about this user from prior interactions. Use these to calibrate your scoring — for example, if notes say "uses heavy profanity casually" then profanity alone should score lower for this user.
-
-If you notice something noteworthy about this user's communication style, behavior, or patterns that would help future analysis, include it as a note_update. Only add genuinely useful observations — don't repeat what's already in the notes. If nothing new, leave note_update as null.
-
-Use the report_analysis tool to report your analysis of the TARGET MESSAGE only."""
 
 ANALYSIS_TOOL = {
     "type": "function",
@@ -126,8 +97,9 @@ class LLMClient:
         self._client = AsyncOpenAI(
             base_url=f"{self.host}/v1",
             api_key=api_key,
-            timeout=300.0,  # 5 min — first request loads model into VRAM
+            timeout=600.0,  # 10 min — first request loads model into VRAM
         )
+        self._semaphore = asyncio.Semaphore(1)  # serialize requests to avoid overloading
 
     async def close(self):
         await self._client.close()
@@ -140,36 +112,38 @@ class LLMClient:
             user_content += f"=== NOTES ABOUT THIS USER (from prior analysis) ===\n{user_notes}\n\n"
         user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}"
 
-        try:
+        async with self._semaphore:
-            response = await self._client.chat.completions.create(
+            try:
-                model=self.model,
+                response = await self._client.chat.completions.create(
-                messages=[
+                    model=self.model,
-                    {"role": "system", "content": SYSTEM_PROMPT},
+                    messages=[
-                    {"role": "user", "content": user_content},
+                        {"role": "system", "content": SYSTEM_PROMPT},
-                ],
+                        {"role": "user", "content": user_content},
-                tools=[ANALYSIS_TOOL],
+                    ],
-                tool_choice={"type": "function", "function": {"name": "report_analysis"}},
+                    tools=[ANALYSIS_TOOL],
-                temperature=0.1,
+                    tool_choice={"type": "function", "function": {"name": "report_analysis"}},
-            )
+                    temperature=0.1,
+                    max_tokens=1024,
+                )
 
-            choice = response.choices[0]
+                choice = response.choices[0]
 
-            # Extract tool call arguments
+                # Extract tool call arguments
-            if choice.message.tool_calls:
+                if choice.message.tool_calls:
-                tool_call = choice.message.tool_calls[0]
+                    tool_call = choice.message.tool_calls[0]
-                args = json.loads(tool_call.function.arguments)
+                    args = json.loads(tool_call.function.arguments)
-                return self._validate_result(args)
+                    return self._validate_result(args)
 
-            # Fallback: try parsing the message content as JSON
+                # Fallback: try parsing the message content as JSON
-            if choice.message.content:
+                if choice.message.content:
-                return self._parse_content_fallback(choice.message.content)
+                    return self._parse_content_fallback(choice.message.content)
 
-            logger.warning("No tool call or content in LLM response.")
+                logger.warning("No tool call or content in LLM response.")
-            return None
+                return None
 
-        except Exception as e:
+            except Exception as e:
-            logger.error("LLM analysis error: %s", e)
+                logger.error("LLM analysis error: %s", e)
-            return None
+                return None
 
     def _validate_result(self, result: dict) -> dict:
         score = float(result.get("toxicity_score", 0.0))
@@ -226,24 +200,43 @@ class LLMClient:
         return None
 
     async def chat(
-        self, messages: list[dict[str, str]], system_prompt: str
+        self, messages: list[dict[str, str]], system_prompt: str,
+        on_first_token=None,
     ) -> str | None:
-        """Send a conversational chat request (no tools)."""
+        """Send a conversational chat request (no tools).
-        try:
+
-            response = await self._client.chat.completions.create(
+        If *on_first_token* is an async callable it will be awaited once the
-                model=self.model,
+        first content token arrives (useful for triggering the typing indicator
-                messages=[
+        only after the model starts generating).
-                    {"role": "system", "content": system_prompt},
+        """
-                    *messages,
+        async with self._semaphore:
-                ],
+            try:
-                temperature=0.8,
+                stream = await self._client.chat.completions.create(
-                max_tokens=300,
+                    model=self.model,
-            )
+                    messages=[
-            content = response.choices[0].message.content
+                        {"role": "system", "content": system_prompt},
-            return content.strip() if content else None
+                        *messages,
-        except Exception as e:
+                    ],
-            logger.error("LLM chat error: %s", e)
+                    temperature=0.8,
-            return None
+                    max_tokens=300,
+                    stream=True,
+                )
+
+                chunks: list[str] = []
+                notified = False
+                async for chunk in stream:
+                    delta = chunk.choices[0].delta if chunk.choices else None
+                    if delta and delta.content:
+                        if not notified and on_first_token:
+                            await on_first_token()
+                            notified = True
+                        chunks.append(delta.content)
+
+                content = "".join(chunks).strip()
+                return content if content else None
+            except Exception as e:
+                logger.error("LLM chat error: %s", e)
+                return None
 
     async def raw_analyze(self, message: str, context: str = "", user_notes: str = "") -> tuple[str, dict | None]:
         """Return the raw LLM response string AND parsed result for /bcs-test (single LLM call)."""
@@ -252,38 +245,40 @@ class LLMClient:
             user_content += f"=== NOTES ABOUT THIS USER (from prior analysis) ===\n{user_notes}\n\n"
         user_content += f"=== TARGET MESSAGE (analyze THIS message only) ===\n{message}"
 
-        try:
+        async with self._semaphore:
-            response = await self._client.chat.completions.create(
+            try:
-                model=self.model,
+                response = await self._client.chat.completions.create(
-                messages=[
+                    model=self.model,
-                    {"role": "system", "content": SYSTEM_PROMPT},
+                    messages=[
-                    {"role": "user", "content": user_content},
+                        {"role": "system", "content": SYSTEM_PROMPT},
-                ],
+                        {"role": "user", "content": user_content},
-                tools=[ANALYSIS_TOOL],
+                    ],
-                tool_choice={"type": "function", "function": {"name": "report_analysis"}},
+                    tools=[ANALYSIS_TOOL],
-                temperature=0.1,
+                    tool_choice={"type": "function", "function": {"name": "report_analysis"}},
-            )
+                    temperature=0.1,
+                    max_tokens=1024,
+                )
 
-            choice = response.choices[0]
+                choice = response.choices[0]
-            parts = []
+                parts = []
-            parsed = None
+                parsed = None
 
-            if choice.message.content:
+                if choice.message.content:
-                parts.append(f"Content: {choice.message.content}")
+                    parts.append(f"Content: {choice.message.content}")
 
-            if choice.message.tool_calls:
+                if choice.message.tool_calls:
-                for tc in choice.message.tool_calls:
+                    for tc in choice.message.tool_calls:
-                    parts.append(
+                        parts.append(
-                        f"Tool call: {tc.function.name}({tc.function.arguments})"
+                            f"Tool call: {tc.function.name}({tc.function.arguments})"
-                    )
+                        )
-                # Parse the first tool call
+                    # Parse the first tool call
-                args = json.loads(choice.message.tool_calls[0].function.arguments)
+                    args = json.loads(choice.message.tool_calls[0].function.arguments)
-                parsed = self._validate_result(args)
+                    parsed = self._validate_result(args)
-            elif choice.message.content:
+                elif choice.message.content:
-                parsed = self._parse_content_fallback(choice.message.content)
+                    parsed = self._parse_content_fallback(choice.message.content)
 
-            raw = "\n".join(parts) or "(empty response)"
+                raw = "\n".join(parts) or "(empty response)"
-            return raw, parsed
+                return raw, parsed
 
-        except Exception as e:
+            except Exception as e:
-            return f"Error: {e}", None
+                return f"Error: {e}", None