diff --git a/pipes/WDL/tasks/tasks_assembly.wdl b/pipes/WDL/tasks/tasks_assembly.wdl
index c0c797905..7e2e971bb 100644
--- a/pipes/WDL/tasks/tasks_assembly.wdl
+++ b/pipes/WDL/tasks/tasks_assembly.wdl
@@ -131,14 +131,23 @@ task select_references {
       --loglevel=DEBUG
 
     # create basename-only version of ref_clusters output file
+    # create tar-bundles of ref_clusters fastas, since Cromwell doesn't delocalize files in a Array[Array[File]] = read_tsv
     python3 <<CODE
-    import os.path
+    import os, os.path, shutil, tarfile
+    os.mkdir("clusters")
     with open("~{contigs_basename}.ref_clusters.tsv", 'r') as inf:
       with open("~{contigs_basename}.ref_clusters.basenames.tsv", 'w') as outf:
         for line in inf:
           fnames = line.strip().split('\t')
-          fnames = list([f[:-6] if f.endswith('.fasta') else f for f in map(os.path.basename, fnames)])
-          outf.write('\t'.join(fnames) + '\n')
+          assert fnames
+          basefnames = list([f[:-6] if f.endswith('.fasta') else f for f in map(os.path.basename, fnames)])
+          outf.write('\t'.join(basefnames) + '\n')
+
+          with tarfile.open(os.path.join("clusters", basefnames[0] + "." + str(len(basefnames))  + ".tar.gz"), "w:gz") as tarball:
+            for f in fnames:
+              shutil.copy(f, ".")
+              tarball.add(os.path.basename(f))
+              os.unlink(os.path.basename(f))
     CODE
 
     # create top-hits output files
@@ -147,7 +156,7 @@ task select_references {
   >>>
 
   output {
-    Array[Array[File]]   matched_reference_clusters_fastas = read_tsv("~{contigs_basename}.ref_clusters.tsv")
+    Array[File]          matched_reference_clusters_fastas_tars = glob("clusters/*.tar.gz")
     Array[Array[String]] matched_reference_clusters_basenames = read_tsv("~{contigs_basename}.ref_clusters.basenames.tsv")
     Array[String]        top_matches_per_cluster_basenames = read_lines("TOP_FASTAS_BASENAMES")
     Array[File]          top_matches_per_cluster_fastas = read_lines("TOP_FASTAS")
@@ -279,6 +288,7 @@ task scaffold {
         cut -f 3 "~{sample_name}.refs_skani_dist.full.tsv" | tail +2 | head -1 > SKANI_ANI
         cut -f 4 "~{sample_name}.refs_skani_dist.full.tsv" | tail +2 | head -1 > SKANI_REF_AF
         cut -f 5 "~{sample_name}.refs_skani_dist.full.tsv" | tail +2 | head -1 > SKANI_CONTIGS_AF
+        basename "$CHOSEN_REF_FASTA" .fasta > CHOSEN_REF_BASENAME
 
         assembly.py order_and_orient \
           "~{contigs_fasta}" \
@@ -345,6 +355,7 @@ task scaffold {
         Int    reference_num_segments_required       = read_int("reference_num_segments_required")
         Int    reference_length                      = read_int("reference_length")
         Array[String] scaffolding_chosen_ref_names   = read_lines("~{sample_name}.scaffolding_chosen_refs.txt")
+        String scaffolding_chosen_ref_basename       = read_string("CHOSEN_REF_BASENAME")
         File   scaffolding_chosen_ref                = "~{sample_name}.scaffolding_chosen_ref.fasta"
         File   scaffolding_stats                     = "~{sample_name}.refs_skani_dist.full.tsv"
         File   scaffolding_alt_contigs               = "~{sample_name}.scaffolding_alt_contigs.fasta"
diff --git a/pipes/WDL/tasks/tasks_utils.wdl b/pipes/WDL/tasks/tasks_utils.wdl
index 18383b5c1..e715ad251 100644
--- a/pipes/WDL/tasks/tasks_utils.wdl
+++ b/pipes/WDL/tasks/tasks_utils.wdl
@@ -148,6 +148,35 @@ task sed {
     }
 }
 
+task tar_extract {
+    meta {
+        description: "Extract a tar file"
+    }
+    input {
+        File   tar_file
+        Int    disk_size = 375
+        String tar_opts = "-z"
+    }
+    command <<<
+        mkdir -p unpack
+        cd unpack
+        tar -xv ~{tar_opts} -f "~{tar_file}"
+    >>>
+    runtime {
+        docker: "quay.io/broadinstitute/viral-baseimage:0.2.0"
+        memory: "2 GB"
+        cpu:    1
+        disks:  "local-disk " + disk_size + " LOCAL"
+        disk: disk_size + " GB" # TES
+        dx_instance_type: "mem1_ssd1_v2_x2"
+        maxRetries: 2
+        preemptible: true
+    }
+    output {
+        Array[File] files = glob("unpack/*")
+    }
+}
+
 task fasta_to_ids {
     meta {
         description: "Return the headers only from a fasta file"
diff --git a/pipes/WDL/workflows/scaffold_and_refine_multitaxa.wdl b/pipes/WDL/workflows/scaffold_and_refine_multitaxa.wdl
index 7defe3335..4112f609b 100644
--- a/pipes/WDL/workflows/scaffold_and_refine_multitaxa.wdl
+++ b/pipes/WDL/workflows/scaffold_and_refine_multitaxa.wdl
@@ -9,7 +9,7 @@ import "assemble_refbased.wdl" as assemble_refbased
 
 workflow scaffold_and_refine_multitaxa {
     meta {
-        description: "Scaffold de novo contigs against a set of possible references and subsequently polish with reads."
+        description: "Scaffold de novo contigs against a set of possible references and subsequently polish with reads. This workflow accepts a very large set of input reference genomes. It will subset the reference genomes to those with ANI hits to the provided contigs/MAGs and cluster the reference hits by any ANI similarity to each other. It will choose the top reference from each cluster and produce one assembly for each cluster. This is intended to allow for the presence of multiple diverse viral taxa (coinfections) while forcing a choice of the best assembly from groups of related reference genomes."
         author: "Broad Viral Genomics"
         email:  "viral-ngs@broadinstitute.org"
         allowNestedInputs: true
@@ -51,14 +51,19 @@ workflow scaffold_and_refine_multitaxa {
 
     # assemble and produce stats for every reference cluster
     Array[String] assembly_header = ["entity:assembly_id", "assembly_name", "sample_id", "sample_name", "taxid", "tax_name", "assembly_fasta", "aligned_only_reads_bam", "coverage_plot", "assembly_length", "assembly_length_unambiguous", "reads_aligned", "mean_coverage", "percent_reference_covered", "scaffolding_num_segments_recovered", "reference_num_segments_required", "reference_length", "reference_accessions", "skani_num_ref_clusters", "skani_this_cluster_num_refs", "skani_dist_tsv", "scaffolding_ani", "scaffolding_pct_ref_cov", "intermediate_gapfill_fasta", "assembly_preimpute_length_unambiguous", "replicate_concordant_sites", "replicate_discordant_snps", "replicate_discordant_indels", "replicate_discordant_vcf", "isnvsFile", "aligned_bam", "coverage_tsv", "read_pairs_aligned", "bases_aligned", "coverage_genbank", "assembly_method", "sample"]
-    scatter(ref_cluster in select_references.matched_reference_clusters_fastas) {
+    scatter(ref_cluster_tar in select_references.matched_reference_clusters_fastas_tars) {
+
+        call utils.tar_extract {
+            input:
+                tar_file = ref_cluster_tar
+        }
 
         # assemble (scaffold-and-refine) genome against this reference cluster
         call assembly.scaffold {
             input:
                 reads_bam = reads_unmapped_bam,
                 contigs_fasta = contigs_fasta,
-                reference_genome_fasta = ref_cluster,
+                reference_genome_fasta = tar_extract.files,
                 min_length_fraction = 0,
                 min_unambig = 0,
                 allow_incomplete_output = true
@@ -120,8 +125,8 @@ workflow scaffold_and_refine_multitaxa {
             "reference_length" :            scaffold.reference_length,
             "reference_accessions" :        tax_lookup.map["accessions"],
 
-            "skani_num_ref_clusters" :      length(select_references.matched_reference_clusters_basenames),
-            "skani_this_cluster_num_refs" : length(ref_cluster),
+            "skani_num_ref_clusters" :      length(select_references.matched_reference_clusters_fastas_tars),
+            "skani_this_cluster_num_refs" : length(tar_extract.files),
             "skani_dist_tsv" :              scaffold.scaffolding_stats,
             "scaffolding_ani" :             scaffold.skani_ani,
             "scaffolding_pct_ref_cov" :     scaffold.skani_ref_aligned_frac,
@@ -164,7 +169,7 @@ workflow scaffold_and_refine_multitaxa {
         String assembly_method                             = "viral-ngs/scaffold_and_refine_multitaxa"
 
         #String assembly_top_taxon_id               = select_references.top_matches_per_cluster_basenames[0]
-        Int    skani_num_ref_clusters              = length(select_references.matched_reference_clusters_basenames)
+        Int    skani_num_ref_clusters              = length(select_references.matched_reference_clusters_fastas_tars)
         File   skani_contigs_to_refs_dist_tsv      = select_references.skani_dist_full_tsv
 
         Array[String] assembly_all_taxids          = taxid