checkpoint rewind

The problem

AI agents running long, multi-step tasks are prone to "hallucination debt." One bad tool call or corrupted file write often goes unnoticed, leading the agent to compound errors for dozens of steps until the entire session is unsalvageable.

What it does

• Enforces a strict checkpoint-before-action discipline for risky operations like file writes, deployments, or payments.
• Maintains a rolling buffer of the last five "known-good" state snapshots including verified goals and assumptions.
• Detects early-warning signs of session corruption, such as contradictory tool outputs or unverified memory claims.
• Executes a standardized rewind protocol to roll back to the last clean state and force a strategy change on retry.
• Switches to damage containment for external actions that cannot be undone, like sent emails or API transactions.

Why this beats prompting it yourself

Generic prompts often lead agents to apologize or retry the exact same failed logic until they hit a limit. This skill implements a rigorous state-machine logic that treats recovery as a first-class feature, ensuring the agent actually changes its approach after a rollback rather than repeating mistakes.

Use cases

• Managing complex refactors where a single broken dependency could break the build.
• Executing multi-stage data migrations that require verification at every milestone.
• Orchestrating autonomous sub-agents where handoffs must be snapshotted to prevent context loss.
• Running long-horizon research tasks that exceed standard context windows.

Known limitations

Cannot physically undo external side effects like sent emails or processed payments. Requires the agent to have internal note-taking or scratchpad capabilities to store checkpoint cards.