Require the API key on /v1 and default thinking off

gateway.py

@@ -23,6 +23,11 @@ MAX_QUEUE_SIZE = int(os.getenv("MAX_QUEUE_SIZE", "32"))
 REQUEST_TIMEOUT = float(os.getenv("REQUEST_TIMEOUT_SECONDS", "900"))
 API_KEY = os.getenv("API_KEY", "").strip()
 LLAMA_API_KEY = os.getenv("LLAMA_API_KEY", "").strip()
+# Spark X2.5 is a reasoning model: with thinking on it routinely spends the
+# whole max_tokens budget reasoning and returns an empty answer, which breaks
+# agent callers. Default it off; a caller can still pass
+# chat_template_kwargs={"enable_thinking": true} explicitly.
+DEFAULT_ENABLE_THINKING = os.getenv("DEFAULT_ENABLE_THINKING", "false").lower() == "true"
 ENGLISH_GUARD = (
     "Respond entirely in English unless the user explicitly requests another "
     "language. Preserve code, identifiers, and quoted source material exactly."
@@ -94,6 +99,13 @@ def require_api_key(authorization: str | None = Header(default=None)) -> None:
         raise HTTPException(status_code=401, detail="Invalid API key")
 
 
+def apply_defaults(payload: dict[str, Any]) -> dict[str, Any]:
+    kwargs = dict(payload.get("chat_template_kwargs") or {})
+    kwargs.setdefault("enable_thinking", DEFAULT_ENABLE_THINKING)
+    payload = {**payload, "chat_template_kwargs": kwargs}
+    return ensure_language_guard(payload)
+
+
 def ensure_language_guard(payload: dict[str, Any]) -> dict[str, Any]:
     if DEFAULT_LANGUAGE.lower() != "en":
         return payload
@@ -139,7 +151,7 @@ async def execute_completion(payload: dict[str, Any]) -> dict[str, Any]:
         async with httpx.AsyncClient(timeout=REQUEST_TIMEOUT) as client:
             response = await client.post(
                 f"{LLAMA_URL}/v1/chat/completions",
-                json=ensure_language_guard(payload),
+                json=apply_defaults(payload),
                 headers=llama_auth_headers(),
             )
             response.raise_for_status()
@@ -265,7 +277,7 @@ async def chat_completions(request: Request):
         except httpx.HTTPStatusError as exc:
             raise HTTPException(exc.response.status_code, exc.response.text) from exc
 
-    payload = ensure_language_guard(payload)
+    payload = apply_defaults(payload)
 
     async def stream_response():
         await wait_for_llama()