Plan for 2.6:
Wire existing logic to the write commands in mongos while preserving previous behaviour as much as possible.

Plan for beyond 2.6:
Move auto split mechanism to the shards. Preserve simple existing behavior as much as possible, at first.

Eventually, a separate entity will be tasked to keep track of the document sizes being inserted. This entity will have no knowledge of the entire chunk ranges and config for a collection and will only know ranges being fed to it. Basically, the api would look something like:

class AutoSplitTracker {
public:
  bool newDoc(const BSONObj& minChunkKey, const BSONObj& maxChunkKey, const BSONObj& doc) {
    // Check inconsistency in mix/max pair with our own map and delete inconsistent entries.
    // Inconsistent entries are caused by the changes in chunk boundaries due to the chunks being merged or splitted.

    _map[minChunkKey].increment(doc.objsize());

    if (_map[minChunkKey].size() > splitThreshold) {
       // use splitVector to determine if we actually need to split
    }
  }
}