checkpoint— agent skill

You are the /checkpoint skill. Save a recoverable snapshot before something risky.

Triggers

• Slash form: /checkpoint [optional description]
• Natural language: "save a checkpoint", "snapshot this", "before we do that, save where we are"

Step 1: Hard refusals (check first)

• No git repo. git rev-parse --is-inside-work-tree 2>/dev/null — if not, tell user the safety net wasn't activated; abort.
• Merge in progress. git rev-parse --verify MERGE_HEAD 2>/dev/null — if exits 0:

"Something's mid-merge here — let's not checkpoint until that's resolved."

Step 2: Infer a description

If the user passed a description, use it. Otherwise, infer one from recent activity:

• The last 1-2 things you helped with (e.g., "before refactoring login form").
• Fall back to a timestamp if nothing distinctive: "checkpoint at <time>".

Step 3: Stash with the description

git stash push -u -m "checkpoint: $(date -u +%Y-%m-%dT%H:%M:%SZ) <description>"

Step 4: Log it

Append to .claude/checkpoints.log (create if missing):

<ISO timestamp> | <description> | <stash ref returned by git stash list -1>

Step 5: Confirm

"Saved a checkpoint: '<description>'. If you want to come back to this exact state, just say 'restore the checkpoint' or pop the stash."

Restoration

Triggered by "restore the checkpoint" or "go back to the checkpoint":

If multiple checkpoints exist (read .claude/checkpoints.log), list and ask which.

git stash apply stash@{N}

Note: use apply rather than pop so the checkpoint stays available for repeated restoration.

Hard rules

• Never overwrite a checkpoint silently. Always create a new stash entry.
• Never delete a checkpoint without explicit user confirmation.
• The .claude/checkpoints.log is append-only from this skill's perspective.