In WDL 1.1, String has an "obvious and unambiguous" coercion to File:

https://github.com/openwdl/wdl/blob/main/versions/1.1/SPEC.md#type-coercion

What is this conversion, exactly?

For example, a natural way to implement a scatter is to have each copy of the body execute in parallel on a different machine. So if you do something like:

workflow Example {

    input {
        String filename = "/etc/hostname"
    }

    File firstFile = filename
    String firstContent = read_string(firstFile)
    scatter(iteration in range(10000)) {
        File file = filename
        String content = read_string(file)
    }

    output {
        String first = firstContent
        Array[String] others = content
    }

}

What is supposed to be true about the others array? Are the values all the same? Are they all the same as first? For that matter, is firstalways the same thing you would see if you ran cat /etc/hostname instead of executing the workflow?

When working with functions that take either File or String, the spec says that Strings are interpreted "relative to the current working directory of the task"; presumably that would apply to coercions as well. But here we're not in a task, and a workflow as a distributed thing doesn't seem to me to have a natural working directory.

In the Toil workflow runner's WDL interpreter, I really want to be allowed to move workflow code to any machine at any line, or even between lines, so I can ship fragments of it to the cluster when they're ready to run, just like the tasks. But that seems to be breaking some of the GATK workflows which assume that, at least at top-level scope, a workflow can always read the same filesystem its caller sees. So I'm trying to work out how legitimate that assumption is, and whether there are guarantees that go further than that.