Datalad Log

Skill: datalad-log

Browse the recorded history of a DataLad dataset: list run commits, inspect what a specific run recorded, and trace which command produced a given output file.

Steps

Browsing run history

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

``bash ls .datalad/ 2>/dev/null `` If not in a dataset, stop and inform the user.

1. List recorded runs — show the run history with:

``bash datalad log --oneline ` Or to show only datalad run commits (excludes saves, checkpoints, manual commits): `bash git log --oneline --grep="datalad run" `` Present the output as a numbered list showing commit SHA and message.

1. Inspect a specific run — when the user wants detail on a particular commit:

``bash datalad rerun --report `` This shows the run record (command, inputs, outputs, message) without re-executing. Report:

• The original command
• Declared inputs (-i)
• Declared outputs (-o)
• Commit message
• Timestamp

Tracing output provenance

When the user asks "what command produced ``":

1. Find commits that touched the file:

``bash git log --oneline -- ``

1. Check each commit for a run record:

``bash datalad rerun --report ` If the commit is a datalad run` commit, this shows the full run metadata. If it is a plain save or checkpoint, report that instead.

1. Report the full provenance chain — if the file was produced by a run, show:

• Which commit recorded it
• The command that produced it
• The declared inputs to that command

Identifying checkpoint commits

The auto-checkpoint hook creates commits with messages like:

[datalad] checkpoint 2026-03-12T14:05:22Z: code/analysis.py outputs/result.csv

These are not run records — they are auto-saves. They will appear in git log and datalad log output. To show only datalad run commits (excludes checkpoints and saves):

git log --oneline --grep="\[datalad run\]"

Note: using --invert-grep with multiple --grep flags uses OR logic and will not correctly exclude checkpoints — use the single --grep="\[datalad run\]" form instead.

Constraints

• Never re-execute a run when the user only wants to inspect it — always use

datalad rerun --report for inspection, not datalad rerun alone.

• Always distinguish between datalad run commits (provenance records) and plain save

commits or checkpoint commits in your report.

• If datalad log is not available (older DataLad), fall back to

git log --oneline --grep="datalad run" and explain the fallback.

• For file-level differences between two commits, direct the user to /datalad-diff.

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.