wyr: few-shot examples + rebuild question from options + switch to abliterated model

- Add 3 assistant-turn examples to lock in the JSON format and tone
- Construct the 'question' field from option_a/option_b so it's always
  well-formed regardless of what the model puts in the 'question' key
- Switch from phi4-mini to the abliterated Llama 3.2 model for edgier,
  uncensored dilemmas

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

matrixbot/commands.py

@@ -1646,44 +1646,48 @@ def record_wyr_vote(event_id: str, sender: str, key: str) -> None:
 
 
 async def _generate_wyr() -> dict | None:
+    # Few-shot examples anchor the format so the model doesn't drift
+    examples = [
+        ('{"question": "Would you rather...", "option_a": "have no internet for a year", "option_b": "never eat your favorite food again"}',),
+        ('{"question": "Would you rather...", "option_a": "always speak in rhymes", "option_b": "only communicate in interpretive dance"}',),
+        ('{"question": "Would you rather...", "option_a": "know the date you die", "option_b": "know the cause of your death"}',),
+    ]
     system_msg = (
-        "You are a game host generating Would You Rather questions for a group of adult friends. "
-        "Always respond with ONLY a JSON object — no markdown fences, no explanation. "
-        'Format: {"question": "Would you rather...", "option_a": "short option", "option_b": "short option"}\n'
-        "Make the dilemma genuinely difficult — both options should have real downsides so it's actually a hard choice. "
-        "Be creative and edgy: embarrassing scenarios, weird superpowers, social nightmares, gross but survivable situations, "
-        "impossible tradeoffs. Avoid boring safe options like 'dance with dolphins' or 'sing karaoke'. "
-        "Keep each option under 10 words."
+        "You are a game host generating Would You Rather dilemmas for a group of adult friends. "
+        "STRICT FORMAT — respond with ONLY a valid JSON object, no other text:\n"
+        '{"question": "Would you rather...", "option_a": "<choice A, under 8 words>", "option_b": "<choice B, under 8 words>"}\n\n'
+        "Rules:\n"
+        "- The 'question' field must ALWAYS be exactly the string 'Would you rather...'\n"
+        "- option_a and option_b are the two actual choices — complete, self-contained phrases\n"
+        "- Both options must have genuine downsides — make it a real dilemma, not an easy pick\n"
+        "- Be edgy and creative: social nightmares, cursed superpowers, embarrassing scenarios, impossible tradeoffs\n"
+        "- Do NOT generate scenarios (no 'accidentally swallow', no 'at midnight') — just two clean choices"
     )
-    user_msg = "Generate a spicy, genuinely difficult Would You Rather question."
+    messages = [{"role": "system", "content": system_msg}]
+    for (ex,) in examples:
+        messages.append({"role": "assistant", "content": ex})
+    messages.append({"role": "user", "content": "Generate a new spicy, genuinely difficult Would You Rather."})
+
     try:
         timeout = aiohttp.ClientTimeout(total=60)
         async with aiohttp.ClientSession(timeout=timeout) as session:
             async with session.post(
                 f"{OLLAMA_URL}/api/chat",
-                json={
-                    "model": ASK_MODEL,
-                    "stream": False,
-                    "messages": [
-                        {"role": "system", "content": system_msg},
-                        {"role": "user", "content": user_msg},
-                    ],
-                },
+                json={"model": BALL_MODEL, "stream": False, "messages": messages},
             ) as response:
                 data = await response.json()
                 text = data.get("message", {}).get("content", "").strip()
-                # Strip markdown fences if present
                 if "```" in text:
                     text = re.sub(r"```[a-z]*\n?", "", text).strip()
-                # Extract the first JSON object found in the response
                 m = re.search(r"\{[^{}]+\}", text, re.DOTALL)
                 if m:
                     text = m.group(0)
                 parsed = json.loads(text)
-                q = parsed.get("question", "").strip()
                 a = parsed.get("option_a", "").strip()
                 b = parsed.get("option_b", "").strip()
-                if q and a and b:
+                # Rebuild question from options so it's always well-formed
+                if a and b:
+                    q = f"Would you rather {a.rstrip('.')} OR {b.rstrip('.')}?"
                     return {"question": q, "option_a": a, "option_b": b}
     except Exception as e:
         logger.error(f"WYR generation error: {e}", exc_info=True)