So is the proposal:
* Store files in content-addressed storage (which may be well-known paths on S3)
* Store Directory ASCII-protos in git (and maybe clone them to a well-known path on S3)
* Have a couple of scripts for:
o Looking up digest of a (name, version, binary_name) tuple
o Fetching a file by digest
o Building things in a Docker image or something
o Uploading files to S3
?