Highlights New datamaestro.helpers module: pluggable helpers contributed by third-party packages through the datamaestro.helpers entry-point group. HFResolver Protocol — redirect HF Hub repos to a local directory (e.g. an HPC cluster's shared model/dataset mirror). Built-in _EnvHFResolver driven by colon-separated env vars: DATAMAESTRO_HF_MODELS_CACHE (model repos: // with config.json) DATAMAESTRO_HF_DATASETS_CACHE (dataset repos: //) Always registered — no plugin required for the env-driven flow. HFDownloader and HFSnapshotDownloader consult get_helpers("hf_resolver") before reaching the network. No symlinks, no copies — the mirror directory is exposed as the resource path directly. Use case On HPC clusters with a pre-mirrored HF cache (e.g. Jean-Zay's $DSDIR/HuggingFace_Models / $DSDIR/HuggingFace): export DATAMAESTRO_HF_MODELS_CACHE=$DSDIR/HuggingFace_Models export DATAMAESTRO_HF_DATASETS_CACHE=$DSDIR/HuggingFace — HF lookups are served from the shared mirror with zero user-quota cost. Distinctive INFO log lines fire on every hit. Tests Adds 13 tests covering the env resolver (single & colon-separated roots, missing-config-json case, dataset lookup, Protocol compliance), plugin layering via entry points, and short-circuit behaviour of both downloaders.