Allow GPU duration hook to receive tools

app.py

@@ -63,7 +63,12 @@ def _bounded_output_tokens(value: float) -> int:
     return max(1, min(requested, MAX_NEW_TOKENS))
 
 
-def _gpu_duration(_: str, __: float, max_new_tokens: float) -> int:
+def _gpu_duration(
+    _: str,
+    __: float,
+    max_new_tokens: float,
+    *_tools: object,
+) -> int:
     """Reserve only the GPU time that the requested output can plausibly use."""
     output_tokens = _bounded_output_tokens(max_new_tokens)
     return min(120, max(30, 20 + math.ceil(output_tokens * 0.08)))