We currently have these three top-level APIs:

func LoadSchema(name string, r io.Reader) (*schema.TypeSystem, error)
func LoadSchemaBytes(src []byte) (*schema.TypeSystem, error)
func LoadSchemaFile(path string) (*schema.TypeSystem, error)

It would be nice to support loading multiple schemas into a single TypeSystem, such as loading the selector schema alongside a custom schema that uses the selector type.

One can do this directly via the schema/dsl and schema/dmt packages, but it would require finding those APIs and figuring out how to string them together, so it's not very nice.

Another option is to concatenate the schema files and load them via LoadSchemaBytes. Certainly doable, but it's still hacky, and means that position information will be mangled.

Going with the flow of go:embed and io/fs, I think the following would be nice:

func LoadSchemaFS(fsys fs.FS) (*schema.TypeSystem, error)

Then, for instance:

//go:embed schemas/*.ipldsch
var schemaFS fs.FS

[...] = LoadSchemaFS(schemaFS)

It would walk the entire filesystem, so it would work with directories too, e.g. //go:embed schemas. One could still only parse subsets of directory trees via APIs like https://pkg.go.dev/io/fs#Sub or https://github.com/josharian/filterfs.

cc @rvagg @hannahhoward per recent graphsync discussions