Datalad Get

Skill: datalad-get

Retrieve annexed file content or initialize absent subdatasets. datalad get fetches content from configured siblings (remotes) and makes files accessible locally.

Steps

1. Verify DataLad context — check for .datalad/ in the current directory or a

parent: ``bash ls .datalad/ 2>/dev/null || git rev-parse --show-toplevel ` If no dataset is found, inform the user and stop — datalad get` only works inside a DataLad dataset.

1. Identify targets — read paths, globs, or subdataset handles from $ARGUMENTS.

If not specified, ask the user what they want to retrieve. For large datasets with many files, mention --jobs N (e.g., --jobs 4) to retrieve content in parallel.

Determine the nature of each target:

• File or directory: retrieve annexed content (datalad get )
• Subdataset path: may need -n (handle only) or full content (-r)

1. Determine retrieval mode — ask or infer:

• Full content (default): downloads file data so files are accessible
• Handle only (-n): clones the subdataset without downloading file content —

useful to inspect structure before committing to a large download

• Recursive (-r): retrieves content in all subdatasets under the path

1. Construct command:

• Single file or directory:

`` datalad get ``

• Recursive (all subdatasets + files):

`` datalad get -r ``

• Subdataset handle only (no file content):

`` datalad get -n `` Show the command before executing.

1. Execute and confirm — run the command. After completion, verify content is

accessible: ``bash ls -lh `` Report whether files are now readable (regular files, not broken symlinks).

Reference

Load ${CLAUDE_PLUGIN_ROOT}/../references/subdataset-patterns.md for the absent/present subdataset model, the -n handle-only pattern, and the neuroimaging nested layout.

Load ${CLAUDE_PLUGIN_ROOT}/../references/global-options.md when the user asks about --on-failure behavior during recursive gets, wants JSON output, needs to override a config value, or asks why a get is failing (suggest -l debug).

Constraints

• datalad get is read-only with respect to dataset metadata — it never modifies

tracked state or creates new commits.

• Distinguish "content missing" (needs datalad get) from "file untracked" (needs

datalad save) — do not confuse the two.

• Never modify .gitmodules directly — if a subdataset is absent, use datalad get -n

to initialize it, not manual git submodule commands.

• Always show the command before executing.
• If retrieval fails due to no configured sibling, suggest running datalad siblings

to check available remotes and datalad siblings enable -s if a special remote needs to be activated.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: neuroanalytics
• Source: neuroanalytics/data-science-harness
• License: MIT

Install and usage instructions live in the source repository linked above.