| name | raincloud-fetch |
| description | Run only the fetch stage (download raw bytes) for the given slugs. Use when the user wants to prime the cache, debug fetch logic, or download upstream bytes without running the full pipeline. |
| argument-hint | <slug>... |
| disable-model-invocation | true |
| allowed-tools | Bash(python -m scripts.pipeline.fetch *) |
Run the fetch-only entrypoint:
python -m scripts.pipeline.fetch $ARGUMENTS
Behavior:
- Writes to
outputs/raw_downloads/<slug>/ (unversioned cache — same upstream bytes serve any schema_version).
- Idempotent: skips when local file matches
expected_bytes / expected_sha256 from the spec. Delete outputs/raw_downloads/<slug>/ to force re-fetch.
- Sibling slugs sharing a URL (GloVe sizes, OSM Germany kinds) are deduped via hardlink.
- With no arg, fetches every dataset in the manifest — almost certainly not what the user wants. Confirm before invoking that way.
For Kaggle entries with requires_interactive_accept: true, a 403 surfaces as an error pointing at the URL the user must click through in a browser (signed into Kaggle) before retrying. See SKILLS.md "Adding a Kaggle dataset gated behind ToS acceptance" and /raincloud-add-kaggle-tos.
Use /raincloud-build instead when you want the full pipeline, not just the download step. Useful standalone for cache priming or debugging fetch logic.
