Loading page
About this tool
Review an AI agent action plan locally for irreversible steps, external side effects, and missing backup, approval, checkpoint, or rollback details.
The Agent Undo Plan Validator reads a proposed AI agent plan — as a JSON action array or as one step per line — and labels every step reversible, recoverable or irreversible, then checks the plan for the four safeguards that would make an undo real: backup, checkpoint, confirmation and rollback. Crucially it grades each safeguard twice, as merely mentioned or as actually detailed, so 'we will take a backup first' is treated differently from a line that says where the backup goes and how it is restored. It is for anyone reviewing an agent plan before granting execution, and it never runs, sends or verifies anything — the analysis is local keyword classification over the text you paste.
Open Agent Undo Plan Validator on AltFTool — it loads instantly in your browser.
Paste the plan into the Action plan box as JSON with an actions key, or as numbered or line-separated steps; Load JSON sample fills an example.
The review updates as you type, tagging every action Reversible, Recoverable or Irreversible and quoting the matched phrase behind it.
Check Backup, Checkpoint, Confirmation and Rollback as Detailed, Too vague or Not found, then press Download report for agent-undo-plan-validation.txt.
A safeguard counts as detailed only when the plan names a location, method, commit, version or approver alongside it — a bare promise to 'restore if needed' is reported as vague, not present.
Sending a message, publishing content and moving money are classified irreversible outright, because no rollback step recalls a delivered email or a completed payment.
Every finding cites the matched wording and the rule it hit, so you can tell a genuine destructive step from a false positive on the word 'remove' in a sentence.
Two things. Either it matches an inherently irreversible external effect — communication, publication, or money movement — or it contains a destructive phrase such as delete, purge, drop table, truncate, force-push or factory reset without a detailed backup-or-checkpoint and a detailed restoration path in the plan. Add both of those in specific terms and the same step downgrades to recoverable.
Four: Backup, Checkpoint, Confirmation and Rollback. Which ones are required depends on the step — a state change needs a checkpoint and a rollback, an external side effect needs confirmation and rollback, and a destructive step needs all four. Each is reported as detailed, mentioned but vague, or not found.
Either JSON — a bare array, or an object with an actions, steps, tasks, operations or plan key — or plain text with one step per line. Markdown bullets, numbered lists, checkboxes, headings and code fences are stripped automatically, so you can paste a runbook straight from a document.
No. This is keyword-based static analysis: it cannot see runtime behaviour, hidden instructions, or what a custom tool name really does, and it never checks that the backups and approvals described in the plan actually exist. A reversible label means a rollback path looks plausible in the wording, so verify that path yourself before granting execution.