Google Batch: How to write to an input directory?

[VasLem]

I have trouble understanding what the right and most elegant way is to write to an input directory in Google Cloud.
Example of a process that computes the reference sizes:

process compute_sizes {
    container 'biopython/biopython'

    input:
        path reference_dir
        val reference_fname

    script:
    """
    #!/usr/bin/env python3
    from Bio import SeqIO
    with open("${reference_dir}/${reference_fname}.sizes", "w") as out_handle:
        with open("${reference_dir}/${reference_fname}", "r") as handle:
            for record in SeqIO.parse(handle, "fasta"):
                out_handle.write(f"{record.id}:{len(record.seq)}")
    """
}

Output working directory contents are (gsutil ls -r gs://BUCKET/workdir):

gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/:
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.begin
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.err
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.out
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.run
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.sh
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.command.trace
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/.exitcode

gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/GC_hg38_ref/:
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/GC_hg38_ref/
gs://BUCKET/workdir/78/1ae8a7bd928b63132b03db4b70c905/GC_hg38_ref/Homo_sapiens_assembly38.fasta.sizes

The original reference directory does not get updated:

gsutil ls gs://BUCKET/reference/GC_hg38_ref/Homo_sapiens_assembly38.fasta.sizes
CommandException: One or more URLs matched no objects.

Adding:

    output:
        path "${reference_dir}/*.sizes", includeInputs: false, optional: true

to the process definition does not do anything.

The only way to manage doing this is by publishing to the input directory, however I cannot pass the input directory to the publishDir directive as is as when I am using path variable, the input directory is actually the mounted softlink and not the actual GS location, so the only way is to pass an extra argument to the process with val instead of path like so:

process compute_sizes {
    container 'biopython/biopython'
    publishDir $REFERENCE_DIRECTORY_AGAIN , mode: "copy", saveAs: {file(it).getName()}

    input:
        path reference_dir
        val reference_fname
        val REFERENCE_DIRECTORY_AGAIN
    output:
        path "${reference_dir}/*.sizes", includeInputs: false
    script:
    """
    #!/usr/bin/env python3
    from Bio import SeqIO
    with open("${reference_dir}/${reference_fname}.sizes", "w") as out_handle:
        with open("${reference_dir}/${reference_fname}", "r") as handle:
            for record in SeqIO.parse(handle, "fasta"):
                out_handle.write(f"{record.id}:{len(record.seq)}")
    """
}

Maybe it is my OCD, but this way is ugly, unmaintainable, and the programmer should have experience to use it. There is no documentation about this bug and workaround. So I have to ask fellow developers, what am I missing here? Thank you all for your answers in advance!