Hermes: disable Qwen3 thinking mode, add repetition_penalty

thinking mode caused the model to loop on uncertain items (e.g. weapon
slot identification). disable_thinking=False + repetition_penalty=1.15
gives clean, non-repetitive responses in ~5s.

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

tools/hermes.py

@@ -133,6 +133,8 @@ def _ask(messages: list, img_b64: str) -> str:
         "model": HERMES_MODEL,
         "max_tokens": MAX_TOKENS,
         "temperature": 0.3,
+        "repetition_penalty": 1.15,
+        "chat_template_kwargs": {"enable_thinking": False},
         "messages": [{"role": "system", "content": _SYSTEM}] + send_messages,
     }).encode()