Extended thinking

Model feature enabling deeper reasoning via dedicated "thinking tokens" processed before the visible response. Activated with `--thinking`.