checkpoint

A checkpoint distinguishes intended, attempted, implemented, and verified outcomes. It compares conversation claims with current repository and check evidence, so the person resuming can see what still needs proof.

Install the skill with the command above. Checkpoint runs only when you invoke it: /checkpoint in Claude or $checkpoint in Codex. Use it before pausing work, handing it off, or recording what your checks prove.

The default result stays in conversation. Add a save request for a file, for example /checkpoint save it, I'll resume tomorrow. You can supply a destination; otherwise a new Markdown snapshot goes under ~/.engineering-workflow/<project-or-subject>/checkpoints/. Earlier snapshots stay intact.

The agent checks the repository, branch, current commit, and uncommitted changes. Plans and earlier agent claims do not prove completion.

Each verification result names the command, result, and commit or working-tree state it covered. The agent reuses available evidence instead of rerunning expensive checks just for the checkpoint. It flags results made stale by later changes. A required check that never ran remains unrun, never passed.

SKILL.md has the snapshot format and saving rules.

A checkpoint with the applicable parts:

• Source: work item, repository, branch, commit, and relevant dirty state.
• Completed outcomes with evidence, and remaining work.
• Blockers and decisions needed to resume.
• Verification: what passed, what it covered, and what remains unverified.
• Next action: a concrete starting point.

On resume, the agent rechecks live sources before acting; the checkpoint is historical. Apart from a requested snapshot, checkpoint changes no files and does not commit or publish.