Datalad Save

Skill: datalad-save

Record code changes, configuration edits, and non-run file modifications in a DataLad dataset. datalad save is the correct replacement for git add + git commit inside a DataLad dataset — it handles both git-tracked and annexed files correctly.

Steps

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

parent directory: ``bash ls .datalad/ 2>/dev/null || git rev-parse --show-toplevel ``

• Dataset found: continue.
• No dataset found (plain git repo or not tracked): fall back to standard

git add + git commit workflow and inform the user: > "No DataLad dataset detected — using plain git commit instead."

1. Run datalad status — show the user what has changed:

``bash datalad status `` Present the output. If nothing is modified or untracked, stop: > "Nothing to save — working tree is clean."

1. Determine save scope — based on $ARGUMENTS and conversation context:

• If specific paths are given (e.g., code/analysis.py configs/), save only those paths
• If no paths given, save all changes (datalad save without path arguments saves everything)
• Show the user which scope will be used and confirm if ambiguous

1. Get or confirm the commit message — read from $ARGUMENTS or ask:

> "What should the commit message be? (describe what changed and why)" Wait for the user's message. Never proceed with an empty or placeholder message.

1. Construct and show the save command:

`` datalad save -m "" [paths...] `` Show it to the user before executing.

1. Execute and confirm — run the command. Then run datalad status again and show

the output to confirm the working tree is clean. Report the commit hash.

Constraints

• Always require a meaningful -m message — never save with an empty string or

placeholder like "save" or "update".

• Never use git add or git commit inside a DataLad dataset — datalad save handles

both annexed and git-tracked files correctly; raw git commands can corrupt annex state.

• Never use git annex add manually — DataLad manages annex staging automatically.
• Always make the save scope explicit: tell the user whether all changes or specific

paths are being saved.

• Always run datalad status before saving to show the user what will be recorded.
• Always run datalad status after saving to confirm a clean working tree.
• If the user asks to save outputs produced by a script, redirect them to datalad run

instead — outputs should be recorded with the command that created them, not as standalone saves.

• Use datalad save --to-git to force a file into git tracking rather than annex

(useful for small config files, .gitignore fragments, etc. that should not be annexed).

• Use datalad save --version-tag to tag an analysis milestone after saving.

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.