nit, this function only seems to require (*d.databufs)[d.currBuf].Len().

I wonder if it's be worth stashing the current buffer's []byte in readResponseDecoder, to avoid the need for repeated calls to ReadOnlyData.

Or perhaps even stash a [][]byte and materialize all the buffers once at the start of decoding. You could avoid an allocation in the common case by keeping a [N][]byte in the decoder:

type readResponseDecoder struct {
  bufs [][]byte
  staticBufs [4][]byte
}

func (d *readResponseDecoder) init() {
  d.bufs = d.staticBufs[:0]
  for _, b := range d.databufs {
    d.bufs = append(d.bufs, b.ReadOnlyData())
  }
}

I don't fully understand this, why would staticBufs be of len 4?

I think I'll probably just leave this for now if that's okay, I don't see overhead from these calls.

Could simplify the code a tiny bit by making b start at the current offset:

b := b.databufs[currBuf].ReadOnlyData()[currOff:]
if len(b) < remaining {
  buf = append(buf, b...)
  remaining -= len(b)
  currBuf++
  currOff = 0
} else {
  buf = append(buf, b[:remaining]...)
  remaining = 0
}

Might be a bit simpler as:

d.currOff += n
for d.currBuf < len(d.databufs) && d.currOff >= d.databufs[d.currBuf].Len() {
  d.currOff -= d.databufs[d.currBuf].Len()
  d.currBuf++
}