I don't think this answer is well understood yet - or at least, not published. We know from the Jaenisch lab papers that single-stranded oligos having 60 nt of homology on either side of the target are functional to direct homology-directed repair (HDR). It's not clear that double-stranded molecules would be better; in fact, they might be targets for CRISPR cleavage too, depending on the exact design of the target & editing design. (Single-stranded DNAs are presumably not recognized by the Cas9 as targets for cleavage, especially if they correspond to the same strand of the protospacer as they can't base pair with the guide RNA).
So, ssDNA oligos are fine for inserting loxP sites or peptide tags as we can order "ultramers" of up to 200 nt+ from IDT, etc. but this won't be practical for inserting longer cassettes like GFP reporters. Here I think the jury is still out on the length of homology arms needed. ~ 1 kb arms seem to be the ballpark for published accounts of reporter insertion, via CRISPR. Bigger is probably better, but what's minimally necessary for efficiency?