One of the interesting things we learned while testing the new OpenAI Provider Adapter harness on ARC-AGI-3 is that even if you set reasoning effort to "none" in the API request, reasoning tokens are still used during context-window compaction. To be clear, this is different from saying the model is reasoning about which game action to take. Those tokens appear to reflect the internal work of deciding what context, state, and prior reasoning to retain during compaction. Another in-the-weeds…