Normally, a snapshot read outside a multi-document transaction will target multiple shards, and mongos selects a timestamp for readConcern.atClusterTime.

If only a single shard is targeted, snapshot timestamp selection can be deferred to the mongod. Thus, sending readConcern: {level: "snapshot"} is sufficient. Implement this optimization in mongos. Capture the atClusterTime returned by the shard and store it on the cursor (for find/aggregate) and return it to the client.