--- name: checkpoint description: Produce a concise, evidence-backed checkpoint for pausing or handing off work. disable-model-invocation: true --- # Checkpoint Use only when explicitly invoked. Capture the requested scope in conversation unless the user requests a saved file. ## Establish the current state Identify the outcome and sources. Inspect each repository's identity, branch, HEAD, dirty state, and relevant files. For other work, identify the authoritative document or tracker and observed state. Reconcile conversation claims with current evidence. Distinguish intended, attempted, implemented, and verified outcomes. Plans and agent claims do not prove completion. Include only material decisions, completed outcomes, outstanding work, blockers, verification, and the next action. Preserve unrelated dirty changes. ## Verification evidence Identify each check's command, result, and covered revision or working-tree state. Use available results or saved evidence; do not rerun expensive checks solely for a checkpoint. Flag results invalidated by later relevant changes. When the tested state is unknown, label results historical and current verification unknown. List unconfirmed or unrun required checks; never report unavailable evidence as passed. ## Output Use this shape, omitting empty optional sections: ```md # Checkpoint: Date: ## Source ## Completed ## Remaining ## Blockers and decisions ## Verification ## Next action ``` ## Save only when requested Honor a supplied destination. Otherwise save a new snapshot at: ```text ~/.engineering-workflow//checkpoints/YYYY-MM-DD-HHMM-.md ``` Use descriptive names. Preserve earlier snapshots; add a suffix on filename collision. Create only the checkpoint and parent directory, with no registry, task ledger, companion artifact, or automatic updates. Check that every completion and verification claim has evidence or an explicit uncertainty label. Return the saved path when applicable. On resume, treat the checkpoint as historical: recheck current sources and repository state before acting. This skill is read-only except for an explicitly requested checkpoint file. It does not authorize code changes, tracker updates, commits, or publication.