Strip Tatlock <think> reasoning before TTS and history

Live smoke test showed /v1/chat/completions replies open with the
Steward's <think> block (Open WebUI convention) — without stripping,
the device would speak the internal monologue aloud.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

gateway/src/desklock_gateway/tatlock.py

@@ -1,9 +1,19 @@
 """Client for the Tatlock butler's OpenAI-compatible chat API."""
 
+import re
+
 import httpx
 
 from .config import settings
 
+# Tatlock prefixes replies with <think> reasoning (Open WebUI convention).
+# Strip it: it must never reach TTS or the on-screen reply text.
+_THINK_RE = re.compile(r"<think>.*?</think>\s*", re.DOTALL)
+
+
+def strip_reasoning(text: str) -> str:
+    return _THINK_RE.sub("", text).strip()
+
 
 class TatlockClient:
     def __init__(self, base_url: str | None = None) -> None:
@@ -25,7 +35,7 @@ class TatlockClient:
             },
         )
         response.raise_for_status()
-        reply = response.json()["choices"][0]["message"]["content"]
+        reply = strip_reasoning(response.json()["choices"][0]["message"]["content"])
         self._history.append({"role": "assistant", "content": reply})
         return reply