Workflow examples

May 30, 2026 · View on GitHub

Workflows are curated multi-skill recipes for the trust/evidence path. They are not publishable skills. Day-to-day use still starts with each skill's SKILL.md.

Run any workflow:

make run-workflow \
  WORKFLOW=examples/workflows/<file>.yaml \
  WORKFLOW_INPUT=<path-to-input> \
  WORKFLOW_OUT=runs/<run_id>

Output includes per-step evidence packs, workflow_summary.json, and workflow_run_record.md. Steps halt on failure.

Workflow 1: CT DICOM to segmentation evidence

File: ct_dicom_to_segmentation_evidence.yaml

Path:

DICOM series
  -> dicom_series_to_volume (metadata + geometry preflight, DICOM-to-NIfTI)
  -> nv_segment_ct (trusted)
  -> ct_segmentation_quality_v1
  -> workflow / trust summary
StepSkillTrusted verifierHandoff
convertdicom_series_to_volume—${input} DICOM directory → output.path NIfTI
segmentnv_segment_ctct_segmentation_quality_v1${convert.output.path}

Run:

make run-workflow-ct-seg \
  WORKFLOW_INPUT=/path/to/dicom_series \
  WORKFLOW_OUT=runs/ct_dicom_seg_evidence

Or:

make run-workflow \
  WORKFLOW=examples/workflows/ct_dicom_to_segmentation_evidence.yaml \
  WORKFLOW_INPUT=skills/dicom-series-to-volume/fixtures/clean_axial \
  WORKFLOW_OUT=runs/ct_dicom_seg_evidence

What it proves: orientation/spacing/geometry gates on conversion; segmentation sanity checks; paired anatomy plausibility (and optional Dice when ground truth is recorded). Halts before segmentation when convert fails (e.g. LR-flipped orientation fixture flipped_lr).

Not the primary skill path. To run one tool on your data without evidence packs, follow skills/dicom-series-to-volume/SKILL.md and skills/nv-segment-ct/SKILL.md in sequence.

A1: DICOM preflight gate

File: dicom_preflight_gate.yaml

GPU-free pass/warn/fail preflight. See examples/README.md.

Other examples