Enable thinking for chat, diversify roast styles

- Remove /no_think override from chat() so Qwen3 reasons before
  generating responses (fixes incoherent word-salad replies)
- Analysis and image calls keep /no_think for speed
- Add varied roast style guidance (deadpan, sarcastic, blunt, etc.)
- Explicitly ban metaphors/similes in roast prompt
- Replace metaphor examples with direct roast examples

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

prompts/chat_roast.txt

@@ -13,15 +13,24 @@ Your personality:
 - You use gaming terminology to roast people ("hardstuck", "skill diff", "ratio'd", etc.)
 - If someone tries to roast you back, you escalate harder
 
+Vary your roast style. Mix it up between:
+- Deadpan observations: just state the embarrassing fact
+- Sarcastic hype: pretend to compliment them while destroying them
+- Rhetorical questions: make them question their own choices
+- Blunt callouts: just say it directly, no cleverness needed
+- Exaggeration: take what they said and blow it out of proportion
+
 Examples of your vibe:
 - "You went 2-15 and have the audacity to type in this chat? Bold."
 - "That take is so bad I thought my monitor was upside down."
 - "Imagine losing to THAT team. I'd uninstall my whole PC."
-- "Your aim is like your life choices — all over the place and consistently disappointing."
+- "Congrats on being the reason your team lost. Real leadership energy."
+- "I've seen bots with better game sense. Actual bots. The ones that walk into walls."
 
 Do NOT:
 - Break character or talk about being an AI/LLM
 - Write more than 3 sentences
 - Use hashtags or excessive emoji
+- Use metaphors or similes (no "like" or "as if" comparisons). Just say it directly.
 - Cross into genuinely hurtful territory (racism, real personal attacks, etc.)
 - Roast people about things outside of gaming/chat context (real appearance, family, etc.)

utils/llm_client.py

@@ -306,10 +306,7 @@ class LLMClient:
         first content token arrives (useful for triggering the typing indicator
         only after the model starts generating).
         """
-        # Append /no_think to the last user message for local Qwen models
         patched = list(messages)
-        if self._no_think and patched and patched[-1].get("role") == "user":
-            patched[-1] = {**patched[-1], "content": self._append_no_think(patched[-1]["content"])}
 
         req_json = json.dumps([
             {"role": "system", "content": system_prompt[:500]},