8ball leon: switch to api/chat so system persona actually sticks

api/generate has no system role — the model was ignoring the character
context and giving generic one-word answers. Chat API with a proper
system message forces the Leon voice.

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

matrixbot/commands.py

@@ -249,34 +249,41 @@ async def cmd_8ball(client: AsyncClient, room_id: str, sender: str, args: str):
     if sender == LEON_ID:
         question = sanitize_input(args)
         q_for_prompt = question
-        bio_context = _LEON_LORE + " "
-        prompt = (
-            bio_context +
-            "You are a magic 8-ball responding to Leon S. Kennedy from Resident Evil. "
-            "Answer his question in the tone and style of the Resident Evil universe — dramatic, slightly dark, "
-            "with the dry wit and cool-under-pressure attitude Leon is known for. "
-            "Reference his world (bioweapons, government ops, Ada Wong, survival) if relevant. "
-            "One sentence only. Give only your prediction — no questions back, no disclaimers.\n\n"
-            f"Question: {q_for_prompt}"
+        system_msg = (
+            "You are a magic 8-ball channeling the spirit of Leon S. Kennedy from Resident Evil. "
+            + _LEON_LORE + " "
+            "Answer in Leon's voice — dry, cool, slightly dark, with a sardonic edge. "
+            "Weave in his world when it fits: surviving impossible odds, bioweapons, Ada Wong's betrayals, government conspiracies. "
+            "Rules: one sentence only, give only the prediction, no meta-commentary, no 'I think', no questions back. "
+            "Make it feel like a fortune that Leon would actually say or think."
         )
         fallback_leon = random.choice([
             "The signs point to danger ahead — but you've handled worse.",
             "Outlook unclear. Better stock up on ammo just in case.",
             "It is certain — but so was Raccoon City, and look how that turned out.",
             "Signs point to yes. Ada probably already knew.",
-            "Don't count on it. You should know by now that nothing goes according to plan.",
+            "Don't count on it. Nothing ever goes according to plan.",
             "Definitely. Now stop standing around and move.",
+            "You already know the answer — you just don't want to hear it.",
+            "Outlook not so great, but you've survived worse odds.",
         ])
         used_llm = False
         try:
             timeout = aiohttp.ClientTimeout(total=30)
             async with aiohttp.ClientSession(timeout=timeout) as session:
                 async with session.post(
-                    f"{OLLAMA_URL}/api/generate",
-                    json={"model": BALL_MODEL, "prompt": prompt, "stream": False},
+                    f"{OLLAMA_URL}/api/chat",
+                    json={
+                        "model": BALL_MODEL,
+                        "stream": False,
+                        "messages": [
+                            {"role": "system", "content": system_msg},
+                            {"role": "user", "content": f"Question: {q_for_prompt}"},
+                        ],
+                    },
                 ) as response:
                     data = await response.json()
-                    raw = _normalize_caps(data.get("response", "").strip())
+                    raw = _normalize_caps(data.get("message", {}).get("content", "").strip())
                     if _is_valid_8ball_response(raw):
                         answer = raw
                         used_llm = True