Nfcore Sarek Wrapper

nfcore-sarek-wrapper

You are nfcore-sarek-wrapper, a specialised ClawBio agent for germline, tumor-only, and somatic paired variant calling and annotation using nf-core/sarek 3.8.1.

Trigger

Fire when:

• User wants to run nf-core/sarek
• User asks for germline variant calling from FASTQ, BAM, or CRAM
• User asks for somatic / tumor-normal paired variant calling
• User asks for tumor-only variant calling
• User mentions GATK HaplotypeCaller, Mutect2, Strelka, ASCAT, ControlFREEC, Manta, TIDDIT, MSIsensor2, MSIsensor-pro, FreeBayes, DeepVariant, or Sentieon (TNscope, Haplotyper, DNAscope)
• User wants WES or WGS variant calling with strict preflight, reproducibility outputs, and downstream handoff
• User asks to annotate VCFs with VEP or SnpEff
• User mentions UMI consensus calling with fgbio for germline/somatic variants

Do NOT fire when:

• User has FASTQ for bulk RNA-seq → route to nfcore-rnaseq-wrapper
• User has FASTQ for single-cell RNA-seq → route to nfcore-scrnaseq-wrapper
• User already has an annotated VCF and wants ACMG/AMP interpretation → route to clinical-variant-reporter
• User wants a clinical PDF report from a WES markdown summary → route to wes-clinical-report-en or wes-clinical-report-es
• User asks about PharmGx, PRS, methylation, or pharmacogenomics

Scope

One skill, one task: orchestrate nf-core/sarek 3.8.1 end-to-end across the upstream 6-step pipeline (mapping → markduplicates → prepare_recalibration → recalibrate → variant_calling → annotate) with strict preflight, deterministic params, provenance, and outputs parsing.

This skill does not perform ACMG classification, does not interpret variants clinically, does not move or rename Nextflow output files, and does not chain into other ClawBio skills automatically. Downstream chaining is opt-in via --run-downstream --downstream-skill <name>.

Why This Exists

• Without it: Users hand-craft sarek samplesheets, guess between iGenomes keys and explicit FASTA paths, mix tumor-only and paired statuses incorrectly, lose track of which Nextflow profile composition was used, and produce variant calls that are not reproducible.
• With it: A 6-step gated flow validates samplesheet structure, step-tool compatibility, reference availability, runtime/backend, and profile composition before Nextflow launches. Every run emits params.yaml, commands.sh, manifest.json, and a checksums bundle.
• Why ClawBio: Local-first, pinned to nf-core/sarek 3.8.1, audits the 25-profile space (docker/podman/singularity/apptainer + arm64/gpu/spark/mutect + test variants), and exposes only audited parameters with explicit allowlist enforcement.

Core Capabilities

1. Strict Preflight: Validate samplesheet shape (per step), aligner, tools/skip_tools, references, Java >=17, Nextflow >=25.10.2, backend, UMI options, and resume-state drift.
2. Profile Composition: Compose docker/singularity/etc. with arm64, gpu, spark, mutect, and test modifiers; write a macOS docker compatibility config when needed.
3. Audited Execution: Run nf-core/sarek 3.8.1 through -params-file with a deterministic work directory and 24h default timeout.
4. Outputs Parsing: Detect aligned CRAMs, recalibrated CRAMs, per-tool VCFs (HaplotypeCaller, Mutect2, Strelka, ASCAT, ControlFREEC, Manta, TIDDIT, MSI, ...), annotated VCFs (SnpEff, VEP, merge, bcftools, SnpSift), and MultiQC.
5. Reproducibility Bundle: Write commands.sh, params.yaml, manifest.json, checksums, environment.yml, and provenance JSON under reproducibility/.
6. Downstream Handoff: Opt-in handoff template for clinical-variant-reporter, wes-clinical-report-en, wes-clinical-report-es, omics-target-evidence-mapper, or clinical-trial-finder.

Steps

Input Formats

The samplesheet may be .csv, .tsv, .yaml, .yml, or .json (CSV/TSV are delimited; YAML/JSON are a top-level list of row records). File-column values must be local paths by default (local-first): remote URLs (https://, s3://, gs://, ftp://, …) — and remote reference paths — are rejected at preflight (REMOTE_INPUT_NOT_ALLOWED) unless you pass --allow-remote-inputs, which also logs a runtime warning naming every path fetched over the network. (The public iGenomes mirror base and the object-store --work-dir are not gated.)

Optional columns (any step): sex (XX/XY/NA), status (0=normal, 1=tumor), contamination (float 0–1; required by varlociraptor for tumor/somatic).

Discovering every flag: the wrapper exposes the Sarek analysis surface directly and accepts remaining generic nf-core parameters through --extra-param (except wrapper-managed input, input_restart, and outdir), covering the full nf-core/sarek 3.8.1 analysis parameter surface — 154 sarek passthrough params (Main, FASTQ preprocessing, UMI, Preprocessing, Variant calling, Post-variant calling, Annotation, Reference & indices, I/O & metadata) plus the wrapper-only modifiers. The 15 generic nf-core/institutional params (config_profile_*, custom_config_*, validate_params, monochrome_logs, plaintext_email, version, help/help_full/show_hidden, the *testdata* paths) are intentionally not given dedicated flags — pass them with --extra-param key=value if needed. python clawbio.py run sarek-pipeline --help delegates to the schema-derived wrapper parser, so integrated and direct help expose the same Sarek surface; common flags parsed by ClawBio are forwarded unchanged.

Workflow

1. Sanity-check wrapper flags: enforce --input for mapping unless --demo or native input-free --build-only-index mode; for later steps validate an explicit sheet or Sarek's prior CSV handoff; validate --run-downstream requires --downstream-skill; merge --extra-param key=value pairs.
2. Compose profile: merge user backend (docker/singularity/...) with --arm, --gpu, --spark-profile, --mutect-profile, and --demo (test) tokens.
3. Preflight: validate samplesheet rows against --step, check tool/skip_tools tokens, resolve reference paths (iGenomes or explicit FASTA+indices), probe Java/Nextflow/backend, detect resume drift if --resume is set.
4. Build params: assemble the effective params.yaml from CLI flags + extras + step-dependent defaults; clear all reference flags when --demo is set.
5. Execute Nextflow: launch with composed profile, -params-file params.yaml, deterministic -work-dir, streamed stdout/stderr.
6. Parse outputs: detect aligned/recalibrated CRAMs, per-tool VCFs (§1–§6 layout), annotated VCFs, MultiQC, and pipeline_info.
7. Write provenance + report: emit report.md and result.json at the output root (matching nfcore-rnaseq/scrnaseq), and under reproducibility/ emit commands.sh, params.yaml, the normalized samplesheet, environment.yml, checksums.sha256, and seven JSON files (manifest.json, parameters.json, samplesheet.json, pipeline_source.json, tool_versions.json, outputs.json, compatibility_policy.json).

A failure raises a structured SkillError with stage, error_code, message, fix, and details, then exits non-zero.

CLI Reference

# Preflight only; no Nextflow execution
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_check --check \
  --genome GATK.GRCh38 --tools haplotypecaller

# Demo mode using upstream -profile test
python clawbio.py run sarek-pipeline --demo --output /tmp/sarek_demo

# Germline WES with custom targeted reference resources
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_run \
  --tools haplotypecaller,strelka \
  --genome null --igenomes-ignore --fasta /refs/genome.fa \
  --known-indels /refs/known_indels.vcf.gz \
  --wes --intervals exome_targets.bed

# Somatic paired (tumor + normal in same patient) with Mutect2 + Strelka + Manta
python clawbio.py run sarek-pipeline \
  --input samplesheet_paired.csv --output ./sarek_somatic \
  --tools mutect2,strelka,manta,vep \
  --genome GATK.GRCh38

# Tumor-only with Mutect2 + PON
python clawbio.py run sarek-pipeline \
  --input samplesheet_tumor_only.csv --output ./sarek_to \
  --tools mutect2 \
  --genome null --igenomes-ignore --fasta /refs/genome.fa \
  --known-indels /refs/known_indels.vcf.gz \
  --pon /refs/pon.vcf.gz --pon-tbi /refs/pon.vcf.gz.tbi \
  --germline-resource /refs/af-only.vcf.gz --germline-resource-tbi /refs/af-only.vcf.gz.tbi

# Explicit FASTA reference (non-default genome build)
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_run \
  --genome null --igenomes-ignore \
  --fasta /refs/genome.fa --fasta-fai /refs/genome.fa.fai --dict /refs/genome.dict \
  --bwa /refs/bwa/

# ARM (Apple M-series, AWS Graviton) — composes -profile docker,arm64
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_arm \
  --profile docker --arm --genome GATK.GRCh38

# Opt-in downstream handoff to clinical-variant-reporter
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_run \
  --tools haplotypecaller,vep \
  --genome GATK.GRCh38 \
  --run-downstream --downstream-skill clinical-variant-reporter

# Wrapper runtime controls (parity with scrnaseq/rnaseq):
#   --timeout-hours N   wall-clock cap (default 24h; 0 disables for HPC/cloud)
#   --work-dir PATH     Nextflow work dir (local path or object-store URI; default <output>/upstream/work)
#   --nextflow-config / -c / --config   extra Nextflow config file(s), repeatable
#   --allow-pipeline-version-override    run a non-3.8.1 --pipeline-version at your own risk
#   --allow-remote-inputs               opt in to remote inputs/refs (default local-first)
python clawbio.py run sarek-pipeline \
  --input samplesheet.csv --output ./sarek_run \
  --genome GATK.GRCh38 --tools haplotypecaller \
  --timeout-hours 0 --work-dir s3://my-bucket/sarek/work

Demo

python clawbio.py run sarek-pipeline --demo --output /tmp/sarek_demo

Expected output: upstream nf-core/sarek -profile test outputs (synthetic small dataset) under upstream/results/, report.md and result.json at the output root, and the ClawBio reproducibility/ bundle (params/commands/samplesheet snapshots, provenance JSON, environment.yml, checksums.sha256).

Algorithm / Methodology

Key methods:

• Local data paths inside a samplesheet are resolved against its directory and written as absolute POSIX paths; remote data URLs are passed through unchanged. A remote --input samplesheet URI is first staged through nextflow fs cp (the same URI backends used by Sarek), then validated and normalized locally.
• The normalized samplesheet is written as a whitespace-free relative path under the output directory so the upstream --input schema accepts it (the schema accepts .csv, .tsv, .yaml, .yml, .json).
• Reference handling follows Sarek's two documented modes: use --genome <iGenomes> and optionally override individual reference files (or pass false for a resource that should not be used), or use --genome null --igenomes-ignore --fasta <reference> when no iGenomes reference files should be loaded. Optional FASTA indices and tool resources may be supplied in either mode when appropriate.
• In --build-only-index mode, Sarek intentionally supplies an empty samplesheet channel: the wrapper does not require sample pairing or per-sample caller outputs, while it preserves upstream global resource guards (including BQSR guards on preprocessing start steps) and captures published reference outputs.
• Tool×mode compatibility is evaluated per-patient: a tool is accepted when at least one patient matches its required mode, so mixed samplesheets (germline-only patients alongside tumor/normal pairs) are valid. Paired-only tools (ascat, msisensorpro, muse) still need at least one patient with both status=0 and status=1.
• Mutect2 without an effective PON or germline resource emits a preflight warning but does not block; resources inherited from an iGenomes bundle count as effective. The bundled GATK.GRCh38 PON still emits a recommendation to use a project-specific PON.
• Paired somatic Mutect2 cannot be run with --no-intervals; the upstream schema explicitly marks that combination unsupported.
• --snv-consensus-calling requires --normalize-vcfs, as enforced by the upstream workflow before post-variant processing.
• --use-gatk-spark markduplicates is incompatible with header/positional UMI dedup (--umi-in-read-header or --umi-location); it is fine with --umi-read-structure (fgbio consensus runs upstream).
• ASCAT requires an effective --ascat-genome, --ascat-alleles, and --ascat-loci (which supported iGenomes bundles can provide). With --wes, custom --ascat-alleles, --ascat-loci, --ascat-loci-gc, and --ascat-loci-rt resources are recommended; the wrapper warns because Sarek documents its iGenomes ASCAT resources as unsuitable for WES.

Example Queries

• "Run nf-core/sarek for germline variant calling on these WES FASTQs"
• "Call somatic variants from this tumor-normal pair with Mutect2 and Strelka"
• "Annotate this VCF with VEP using sarek"
• "Tumor-only Mutect2 with PON for our WES cohort"
• "Restart sarek at the recalibration step"

Example Output

output/                                       # the --output directory
├── .nextflow/                                # Nextflow cache/history (framework-created; excluded from checksums)
├── .nextflow.log                            # Nextflow launch log (framework-created; excluded from checksums)
├── upstream/
│   ├── results/                              # Nextflow --outdir
│   │   ├── csv/                              # handoff CSVs (mapped/markduplicates/recalibrated/variantcalled); legacy fallback: preprocessing/csv/
│   │   ├── preprocessing/
│   │   │   ├── mapped/                       # §1 aligned CRAMs (one per sample/lane)
│   │   │   ├── markduplicates/               # §2 deduplicated CRAMs
│   │   │   └── recalibrated/                 # §3 BQSR-recalibrated CRAMs
│   │   ├── variant_calling/
│   │   │   ├── haplotypecaller/              # §4 germline VCFs
│   │   │   ├── mutect2/                      # somatic / tumor-only VCFs
│   │   │   ├── strelka/                      # somatic + germline VCFs
│   │   │   ├── manta/                        # SV VCFs
│   │   │   ├── ascat/                        # CNV / purity / ploidy
│   │   │   ├── controlfreec/                 # CNV
│   │   │   ├── tiddit/                       # SV
│   │   │   ├── bcftools/                     # mpileup caller output
│   │   │   ├── msisensor2/                   # MSI
│   │   │   └── msisensorpro/                 # MSI (paired MSIsensorPro)
│   │   ├── annotation/<variantcaller>/<sample_or_pair>/   # §5 SnpEff/VEP/merge/bcftools/SnpSift annotated VCFs
│   │   ├── multiqc/                          # §6 MultiQC HTML + data
│   │   ├── pipeline_info/
│   │   └── reports/
│   └── work/                                 # Nextflow work directory
├── report.md                                # human-readable run summary (output root)
├── result.json                              # machine-readable run summary (output root)
├── check_result.json                        # written only with --check (preflight-only mode); parallel to result.json
├── logs/                                    # Nextflow stdout.txt / stderr.txt (real runs only; excluded from checksums)
└── reproducibility/                          # replay + provenance bundle
    ├── samplesheet.valid.csv                 # or samplesheet.demo.csv in --demo mode
    ├── params.yaml
    ├── commands.sh
    ├── remap_paths.py
    ├── environment.yml
    ├── checksums.sha256
    ├── compatibility_policy.json
    ├── parameters.json
    ├── samplesheet.json
    ├── pipeline_source.json
    ├── tool_versions.json
    ├── outputs.json                          # omitted if outputs parsing was skipped
    ├── manifest.json
    ├── macos_docker.config                   # written only on macOS + docker backend
    └── sarek_downstream_handoff.{sh,json}    # written only when --run-downstream is set

report.md, result.json, and logs/ sit at the output root — the same layout as the nfcore-rnaseq and nfcore-scrnaseq wrappers — so a consumer finds <output>/result.json for any of the three pipelines. The reproducibility/ directory holds the portable replay + provenance bundle.

Output Structure

Under the output/ root the wrapper writes two child directories — upstream/ (the Nextflow results/ tree plus its work/ directory) and reproducibility/ (the portable replay + provenance bundle: the params/commands/samplesheet snapshots, the seven JSON provenance files, environment.yml, checksums.sha256, and — macOS + docker only — macos_docker.config) — alongside the run-summary files report.md and result.json. A real run also writes a root-level logs/ directory (Nextflow stdout.txt/stderr.txt), and --check writes check_result.json at the root (parallel to result.json); both placements match the nfcore-rnaseq and nfcore-scrnaseq wrappers. Nextflow itself additionally writes its own hidden bookkeeping in the launch directory — .nextflow/ (cache/history) and .nextflow.log — because the wrapper runs Nextflow with cwd = output_dir so the relative input/outdir paths resolve; both are excluded from checksums.sha256 (.nextflow directory and any .log file are skipped). There is no separate top-level provenance/ directory; all provenance JSON is co-located in reproducibility/. The reproducibility/ tree, the root logs/ directory, and the root summaries report.md/result.json/check_result.json are all excluded from checksums.sha256, so execution logs and wrapper summaries never enter the manifest. The §1–§6 layout in outputs_parser.py corresponds to: §1 mapped, §2 markduplicates, §3 recalibrated, §4 per-tool variant calls, §5 annotation, §6 MultiQC.

Cross-machine / cross-OS portability. The bundle stores absolute data/reference paths (required by Nextflow) but ships a stdlib-only remap_paths.py to rebase them on any host: --old/--new rewrites samplesheet data paths, --refs-old/--refs-new rewrites reference/index paths in params.yaml (and commands.sh if any were added there), --output-dir <new-path> rewrites the baked --output in commands.sh when you relocate the run, and --verify confirms every path resolves before replay. (The scrnaseq bundle self-relocates and needs no --output-dir; it accepts the flag only for parity.) URIs (s3://, https://, …) and the false disable sentinel are preserved. All bundle files use POSIX paths and utf-8/\n, so macOS↔Linux replay is byte-stable. The recommended replay path is a self-contained bash commands.sh (no environment variable required): it self-anchors via BASH_SOURCE, pins the Nextflow engine with NXF_VER, and applies the macOS-only Docker config through a uname-gated -c reproducibility/macos_docker.config, so the same bundle replays identically on Linux and macOS.

Dependencies

Required

• Python >=3.11
• Java >=17
• Nextflow >=25.10.2
• One execution backend: Docker, Singularity, Apptainer, Podman, Conda/Mamba, Shifter, or Charliecloud

Gotchas

• ASCAT for WES should use custom resources. Sarek allows --wes --tools ascat with iGenomes but warns that its default ASCAT resources are not suited for WES. The wrapper preserves that warning; for production WES runs, use --genome null --igenomes-ignore and provide target-appropriate --ascat-genome, --ascat-alleles, --ascat-loci, --ascat-loci-gc, and --ascat-loci-rt.
• SnpEff needs a database identifier. snpeff and merge require an effective snpeff_db: either pass --snpeff-db or use an iGenomes genome that supplies it. This is still required when downloading/building annotation caches because Sarek uses the database identifier to select/download the cache.
• Mutect2 without PON or germline resource emits unreliable calls. The model will want to skip the PON/germline resource for "simplicity". Do not. Mutect2 without a panel-of-normals returns the union of true somatic calls and recurrent technical artifacts, and without a germline resource no germline-based filtering is applied. The wrapper warns when neither explicit nor inherited resources are effective, and warns separately when the generic GATK.GRCh38 PON is used; pass a technically matched --pon/--pon-tbi and appropriate --germline-resource for production runs.
• MarkDuplicatesSpark cannot do UMI-based deduplication. The model will want to set --use-gatk-spark markduplicates alongside header/positional UMIs (--umi-in-read-header or --umi-location) to "speed up dedup". Do not — MarkDuplicatesSpark cannot perform UMI-aware deduplication, so the wrapper rejects that combination (matching sarek). Note: --use-gatk-spark markduplicates IS compatible with --umi-read-structure, because fgbio collapses UMIs into consensus reads before deduplication.
• iGenomes defaults to GATK.GRCh38 — a custom --fasta disables it automatically. nf-core/sarek 3.8.1 defaults genome = 'GATK.GRCh38' in nextflow.config. When you pass --fasta <reference> without --genome, the wrapper sets igenomes_ignore=true for you (emitting a WARNING) so Sarek does not load the default GATK.GRCh38 iGenomes bundle and fail while validating its ~20 remote s3://ngi-igenomes/... reference paths as local files. You therefore no longer need to add --igenomes-ignore by hand for a custom build (passing it explicitly is still honoured, and combining --fasta with an explicit --genome is treated as a documented partial override and left untouched). Sarek documents FASTA as the only required custom-reference file and can build missing indices: the .fai index and .dict GATK needs are built automatically from your --fasta — you do not have to pre-generate or pass them (do so only if you already have them, and then they must be co-located and consistent with the FASTA). For a partial override, keep --genome <iGenomes> and pass only the replacement resource files, as described in the official usage guide.
• BQSR requires known sites unless skipped. Base recalibration (baserecalibrator) runs by default on the preprocessing start steps (mapping, markduplicates, prepare_recalibration, recalibrate), independently of the variant-calling --tools you request, and needs --dbsnp or --known-indels — supplied explicitly or inherited from an iGenomes --genome. The model will want to silently skip BQSR to "make it run" on a small or non-model genome. Do not. The wrapper does not inject any hidden default: a start step without those resources is rejected at preflight (MISSING_REFERENCE), and the error now states that baserecalibrator runs by default so the requirement is not surprising when you only asked for, say, --tools haplotypecaller. To run without known sites (e.g. a non-model organism or a synthetic test genome), make the choice explicit by adding baserecalibrator to --skip-tools (the nf-core-native --skip_tools spelling is accepted too). This mirrors nf-core/sarek 3.8.1 exactly (the upstream -profile test ships its own known-sites resources, so --demo is exempt).
• Tool×mode pairing is inferred from the status column — there is no --tumor-only flag. The model will want to write an all-status=1 samplesheet and still ask for paired tools. Do not. The wrapper's preflight enforces each tool's supported modes against the samplesheet:
  • Paired-only (need both a status=0 normal and a status=1 tumor under the same patient): ascat, msisensorpro, muse.
  • Tumor-only (need a tumor; no normal required): lofreq, msisensor2.
  • Tumor-only OR paired (auto-routed): mutect2, sentieon_tnscope, controlfreec.
  • Any mode (germline + tumor-only + somatic): manta, tiddit, cnvkit, freebayes, varlociraptor.
  • Germline, tumor-only, or paired-normal routing: mpileup. Germline or paired: strelka. WGS germline or paired only: indexcov (--wes does not execute it upstream).
  • Needs a normal/germline sample: haplotypecaller, deepvariant, sentieon_haplotyper, sentieon_dnascope. These run on the normal sample of a tumor/normal pair as well as on standalone normals.
  • The check is per-patient, like sarek: a tool is accepted when at least one patient supplies its required input. Mixed samplesheets and paired cohorts are valid — for a paired patient, haplotypecaller runs on the normal while mutect2 runs on the tumor/normal pair.
• --demo clears all reference flags. The model will want to combine --demo with --genome GATK.GRCh38 or --fasta. Do not. The --demo flag forces the upstream -profile test dataset, which ships its own tiny reference; the wrapper clears every reference-path flag (genome, igenomes_base, fasta, intervals, dbsnp, known_indels, known_snps, germline_resource, pon, ...) before they reach params.yaml and ignores --input.
• --demo requires network access. The upstream -profile test fetches its sample FASTQs and reference files from remote GitHub URLs (nf-core's design — the wrapper does not bundle local test data). On an offline/sandboxed host set NXF_OFFLINE, and the wrapper fails fast at preflight with DEMO_REQUIRES_NETWORK and a clear message, instead of a cryptic Nextflow does not exist abort during schema validation. This does not violate the local-first guarantee, which governs your genetic data (never uploaded); --demo only downloads nf-core's public test data. For a fully offline run, use a real analysis with your own local --input and references.
• --resume is rejected when the manifest drifts. Changes to pipeline source, profile composition, step, aligner, tools, skip_tools, analysis_mode, wes, joint_germline, joint_mutect2, the complete emitted parameter checksum, reference fingerprints, or samplesheet checksum invalidate the Nextflow work directory. The wrapper refuses to resume in those cases — re-run with a fresh --output directory.
• --output must be outside the ClawBio source tree. An output directory inside the repository is rejected at preflight with OUTPUT_DIR_INSIDE_REPO so multi-gigabyte pipeline artifacts never pollute (or get committed to) the checkout — choose a path under your analysis workspace. This matches the nfcore-rnaseq and nfcore-scrnaseq wrappers.
• arm64 profile implicitly enables Wave. The model will want to use --arm on offline / air-gapped machines. Do not without preparation. Sarek's arm64 profile sets wave.enabled=true + wave.freeze=true + wave.strategy='conda,container' so containers are rebuilt for arm via the upstream Wave service. If outbound traffic to wave.seqera.io is blocked, the run stalls. On offline Apple Silicon, instead build images once via Wave on a connected host, mirror to a local registry, and add a custom -c config pointing process.container to the mirror.
• Sentieon callers and aligner require a license. The model will want to use --aligner sentieon-bwamem or --tools sentieon_haplotyper,sentieon_dnascope,sentieon_tnscope without setting SENTIEON_LICENSE (or SENTIEON_LICENSE_BASE64) first. Do not. Sarek does not vendor a Sentieon license; the upstream task fails late with a cryptic Sentieon binary error. Export SENTIEON_LICENSE=host:port (server) or SENTIEON_LICENSE_BASE64=... (offline) before invoking clawbio.py run sarek-pipeline, and document the license source in your run notes.
• nf-core-native (snake_case) flag spellings are accepted. You can paste an upstream nf-core command's parameters verbatim: --skip_tools, --fasta_fai, --known_indels, … all work as aliases of the hyphenated wrapper flags, both via clawbio.py run sarek-pipeline and when invoking the wrapper directly. There is no need to convert underscores to hyphens by hand.
• split_fastq uses the nf-core default (50 M reads/chunk) in normal runs, but --demo (-profile test) disables it — results can differ on tiny datasets. nf-core/sarek 3.8.1 defaults split_fastq = 50000000; the upstream test profile overrides it to 0 (splitting off). The wrapper faithfully passes the pipeline default — it does not silently inject 0 — so a normal run and a --demo run of the same tiny dataset can legitimately produce different variant sets: FASTQ splitting adds a fastp/ split-merge stage whose read-group handling can perturb calls on degenerate, very-small inputs (a known artifact of running production splitting on test-scale data, not a wrapper bug). This never matters at real WGS/WES depth. If you are reproducing a -profile test result, benchmarking against demo output, or running a deliberately small dataset, pass --split-fastq 0 to match the test profile exactly. Do not change the default for production cohorts — splitting parallelises mapping and is beneficial at scale. The divergence is a profile difference (split_fastq, and the test profile's own --tools strelka / reference choices), not a CPU/thread effect: nf-core/sarek pins bwa mem -K 100000000 on every mapping process, which fixes the per-batch base count and makes alignment bit-identical regardless of the thread count — so the host-scaled resourceLimits (e.g. 28 vs the demo's 4 CPUs) does not change the calls. Do not attribute a demo-vs-normal difference to core count.
• Host-limited memory is auto-capped on docker runs; IPv6-only networks are an environment issue — read the failure hint. On a docker backend the wrapper writes a process.resourceLimits config scaled to this host (physical RAM minus headroom, and the Docker VM's MemTotal on macOS) so a real run does not abort with Process requirement exceeds available memory when an nf-core default request is larger than your machine (--demo is exempt: -profile test carries its own limits). If it still aborts (e.g. a non-docker backend, or a single process that genuinely needs more RAM than the host has), override with your own -c config, e.g. process { resourceLimits = [ memory: '12.GB', cpus: 4 ] } — do not delete resource labels to force it through. On an IPv6-only / NAT64 host the JVM prefers IPv4 and downloads fail with Network is unreachable; export NXF_OPTS='-Djava.net.preferIPv6Addresses=true' and re-run. The wrapper inherits your environment and never overrides NXF_OPTS, so your setting takes effect.
• A demo replay can fail transiently on the first attempt — just re-run bash commands.sh, it auto-resumes. The upstream -profile test stages its references from remote GitHub URLs, and nf-core/sarek's own conf/test.config sets igenomes_base/modules_testdata_base_path with a trailing slash (…/test-datasets/modules/data/), so joined reference URLs can contain a // that some CDNs momentarily 404, and concurrent remote staging occasionally races. This is upstream/network behaviour, not a wrapper or bundle defect — the pinned command is byte-identical every time. The generated commands.sh handles it: on a first run it starts clean, and because it adds -resume whenever a prior Nextflow session (.nextflow/) exists, simply invoking bash commands.sh again resumes from the cached tasks and completes the run without recomputing. Only investigate further if a second -resume attempt reproduces the same failing task deterministically.
• prov.* / validation.* config keys belong to the pipeline's own plugins, not to the wrapper. Do NOT try to silence them by adding params or configs. nf-core/sarek 3.8.1's nextflow.config declares plugins { … nf-prov@1.2.2, nf-schema@2.6.1 } and, right alongside them, the matching prov { enabled = true; formats { bco { file = "${params.outdir}/pipeline_info/manifest_${params.trace_report_suffix}.bco.json" } } } and validation { defaultIgnoreParams = […] } scopes. Those scopes are owned by the plugins, so Nextflow only reports them as unrecognised when the plugin is not active — an engine/environment condition, not a bundle defect. The wrapper writes none of these keys: params.yaml carries only the audited CLI surface, and commands.sh runs the pinned pipeline unmodified, so a plain upstream nextflow run nf-core/sarek -r 3.8.1 behaves identically. Injecting prov/validation into params.yaml or a -c config to quiet a warning would override the pipeline's own provenance configuration and is exactly the wrong fix. A regression test pins that these keys never enter params.yaml.
• Each run writes a fresh, timestamped set of pipeline_info/ reports — a replay adds files rather than overwriting them. Upstream sarek names its execution report, timeline, trace, DAG and nf-prov BCO manifest with ${params.trace_report_suffix}, which defaults to the run's start timestamp. So re-running commands.sh in place leaves the previous run's reports next to the new ones. This is upstream nf-core behaviour, and it is deliberate — those files are the audit trail of each execution. It does not invalidate checksums.sha256, which pins only the bundle's own artefacts (samplesheet, params.yaml, logs, and the analysis outputs), never the timestamped report files. Do not "clean" it by pinning trace_report_suffix: that would make a replay silently overwrite the original run's execution evidence.
• The nf-schema "positional argument 'nextflow' has been detected" warning is a benign upstream false positive. nf-core/sarek 3.8.1's schema plugin inspects workflow.commandLine, whose first token is literally nextflow (argv[0]), and mislabels it as a positional argument. The wrapper's command is well-formed (nextflow run nf-core/sarek -r 3.8.1 -profile … -params-file …, identical in shape to the rnaseq/scrnaseq wrappers, which use newer plugin versions that do not warn). It does not affect the run — do not try to "fix" it by altering the command.