| You get | A review record and a small example of complete, repair and pause decisions |
|---|---|
| Stays with you | Defining a passing result, approving another attempt and integrating any code |
| You need | A saved version history, failure evidence and a limit chosen before another attempt |
| Rough cost | Review time and any model or tool charges for actual attempts; this teaching script makes no model calls |
| Skip it if | The agent is offline rather than revising the same work |
1What it looks like handed off
You can see the last checked version, the failing requirement and how many repairs have been tried. When the same work returns, or the allowance is used up, the workflow pauses with that evidence instead of silently starting again.
A fresh community exchange about repeated review and revision prompted this guide. Our answer proposed a workflow safeguard; it did not establish that the poster installed it or that it solved the problem. The examples below are authored teaching material, not a client result or measured model performance.
2The job, end to end
- Write the exact requirement that is failing and the evidence that would establish a pass.
- Keep the current version and its review result. Distinguish a changed status label from changed work.
- Choose a repair allowance for the work item. Three is an example, not a recommended universal setting.
- Before another attempt, check whether the proposed work repeats a version already seen and whether the allowance remains.
- Record the next decision and repair count before launching permitted work. A restart must reload them, not reset them.
- Pause with the failed check, versions and remaining questions when the work repeats, the allowance runs out or the state cannot be trusted.
Try the fictional decision example
Download the Python teaching script and review sheet (CSV). The script uses only its five fictional cases. Inspect it first; with Python3 installed, run python3 decision-example.py to see the decisions. It makes no network request, model call or file change.
| Fictional observation | Decision and reason |
|---|---|
| Version A returns after A and B | Pause: the work product repeated |
| Three repairs used from an allowance of three | Pause: no allowance remains |
| A different version, one repair used from three | Permit one bounded repair and increment the proposed record |
| The saved record is missing | Pause: do not manufacture a fresh allowance |
| A trusted verification result is supplied as passed | Complete: only within that supplied check's scope |
The short labels A and B stand for work-product fingerprints. A real implementation must define what is compared; an identical draft does not prove every surrounding circumstance is identical. The script does not calculate fingerprints, inspect a document or verify its claims.
3What stays with you
A programmer must connect the decision to the workflow, bind each saved record to its exact work item, save state reliably and prevent concurrent workers from spending the same allowance. The download does none of that. It is not a built-in Hermes or OpenClaw setting, a plugin or a complete agent runner.
A model saying that it passed is not the trusted verification input. Use the actual test, document review or other evidence required for the job. Keep accounts, sending and spending permissions unchanged; a new repair is not permission for another consequential action.
4What happens when the check is wrong
A repeated draft can be legitimate after the source or requirement changes. Pause and have a person resolve that change, preserving the old history. Do not let the same model erase its counter or increase its own allowance merely because it wants another attempt.
If the computer or gateway is actually offline, start with the gateway troubleshooting guide instead. A transition guard does not repair a disconnected service.
5What it costs
Record model and tool charges along with time spent checking and correcting the work. A pause limits further attempts; it does not prove the work was useful or that money was saved. Try the fictional cases before connecting a rule to a real workflow.
6When not to bother
For one draft that you review yourself, a version note and an explicit stop may be enough. Do not build a workflow engine for a task you can finish manually. If you cannot define a passing result, resolve that first; a numeric repair limit will not define quality for you.
Updated : added related guides. Cost estimates and source-check dates retain their stated dates.
7Want this running on your own computer?
I'll set it up, and keep it running if you like. On a Mac mini, a Windows PC or Omarchy, with the agent you use or want: OpenClaw, Hermes, Claude Code, Codex or another. Approval settings depend on the agent, tools and configuration. Tell me about the job and I'll quote it. If a $20 subscription does it, I'll tell you. You keep the computer, the accounts and the setup.
Tyler Mayberry, who runs Cirlet (formerly Animas)
For independent insurance agencies nationwide, Cirlet offers AI and automation help. The first 30-minute conversation is free. Paid work and ongoing support are scoped and priced before they start.