Bio Workflow Methods Docwriter

Write Methods text and run documentation from real workflow artifacts. Every command, version, and parameter comes from evidence; nothing is inferred.

Instructions

1. Collect the evidence package listed in reference/evidence-checklist.md: launch command, configs, engine version, tool versions or container digests, input checksums or accessions, reference and database versions, QC reports, and the final output list.
2. Build a draft run_manifest.yaml with the extractor for the engine:
   • Nextflow: extract_nextflow_run.py reads trace.txt and the work/ task folders. --tool-versions is a JSON mapping of process name to {"tool": ..., "version": ...}; --outputs-manifest lists one final output path per line.
   • Snakemake: extract_snakemake_run.py reads a JSONL file with one record per job (rule, status, tool, version, command, inputs, outputs). Assemble it from --printshellcmds logs and the version records; see fixtures/snakemake/jobs.jsonl.
   • CWL: extract_cwl_run.py reads one normalized JSON record (run metadata plus steps and outputs), not a raw CWLProv folder; see fixtures/cwl/provenance.json.
3. Fill the remaining fields from evidence only. While drafting, mark a missing value NOT CAPTURED; the final manifest must not contain it. If the value cannot be recovered, record the gap under limitations and do not call the manifest reproducible.
4. Validate the manifest against schemas/run-manifest.schema.json with validate_run_manifest.py.
5. Draft METHODS.md from templates/methods_report.md: workflow summary at the top, then inputs, exact rerun command, environment, steps as executed, QC, outputs, and limitations.
6. For manuscript Methods prose, hand the manifest to /scientific-writing. Keep script names, internal flags, and runtime bookkeeping in Methods or the supplement, not in Results.

Resolve the installed skill once:

METHODS_SKILL="${METHODS_SKILL:-$HOME/.agents/skills/bio-workflow-methods-docwriter}"

Each script declares its dependencies inline (PEP 723), so uv run --script needs no separate environment.

Quick Reference

Input Requirements

• Workflow artifacts (Nextflow/Snakemake/CWL logs and configs)
• Tool version records or container digests
• QC reports and output manifests

Output

• METHODS.md (agent-authored; workflow summary first, then detailed steps)
• run_manifest.yaml (machine-readable run manifest)
• Artifact contract: schemas/run-manifest.schema.json
• Paper-summary schema: schemas/bio-paper-schema.yaml

Quality Gates

• No invented commands, versions, or parameters
• Every step has inputs, outputs, and versions captured
• Commands were sourced from task scripts, not environment-bearing wrappers, and contain no credentials
• No NOT CAPTURED, UNKNOWN, or TBD placeholder remains in a required field
• Workflow summary appears at top of METHODS.md
• The final manifest passes validate_run_manifest.py

Examples

Example 1: Nextflow run to validated manifest

METHODS_SKILL="${METHODS_SKILL:-$HOME/.agents/skills/bio-workflow-methods-docwriter}"
uv run --script "$METHODS_SKILL/scripts/extract_nextflow_run.py" \
  --trace trace.txt --workdir work --out run_manifest.yaml \
  --run-id rnaseq-2026-02-02 --pipeline-name nf-core/rnaseq --pipeline-version 3.14.0 \
  --commit-sha <sha> --engine-version <nextflow -version> \
  --launch-command "<exact launch command>" \
  --tool-versions tool-versions.json --outputs-manifest outputs.txt
uv run --script "$METHODS_SKILL/scripts/validate_run_manifest.py" run_manifest.yaml

Troubleshooting

Issue: Missing tool versions in logs Solution: Use NOT CAPTURED only while assembling a draft. The final validator rejects it; recover the version from provenance or report the missing evidence in limitations without claiming a reproducible manifest.