I'm kind of imagining a rule

fetch_single_coordinate(coord: MavenCoord) -> Digest

, except I just tell the engine to act as though that rule has executed and give it both the input and the output

Using REAPI’s notion of an Action: In the context of “batching” test runs, I’ve thought about “synthetic Actions” where you would run

pytest

in a single process execution to execute a batch of tests all at once and then unpack the Action->ActionResult resulting from that process execution by creating N synthetic Action->ActionResult mappings for each test. You would do that by creating a synthetic command-line for the fake process executions. It would match how Pants would setup a run of a single test. Then later if Pants tried to execute a single test, by constructing the command-line for a single test, it would see that the single test already ran even though it was run in a single batch run of

pytest

.

Interesting, yeah. I don't think this would work for testing, but for external dependency resolution I could also imagine taking a full index of the remote repo as a prior input. Then even with relatively frequent cache invalidation, you could at least be sure that your resolution and fetching actions are operating effectively in-memory, and even "inefficient" fetches aren't actually touching the network

Yes 100%, I think this wants to be deconstructed into "resolve to versions" as one process and "fetch a version" as another