Author here. Good question! /review is great for a quick pass on something you're still writing. I use it daily.
There are things it doesn't do well though. it can't take into account different priors. Built in reviews run on the same model family that wrote the code, so a subagent gets you fresh context and the same blind spots. Paranoia runs the review on another vendor's CLI. Here, Codex and Claude. For example, in my own work the same vendor declared converged and the other came back ten minutes later with two fatals and twelve majors on the same doc.
There's memory. Built in reviews are stateless, so round six happily re-reports round one's bug in a new spelling and nobody's any the wiser. Here the reviewer registers the defect class as a regex, the server re-runs it every round with git grep, and a computed trailer won't call it done while it still matches.
It's more expensive to set up, though, and require a subscription to both ChatGPT and Claude to get the most out of it. Not worth it for small changes, worth it for critical changes that are hard to walk back. For a fully automated harness, assuming you have max subs on both sides, it's perfect for making sure the agent is getting things right.
Why not just use the built in /review functions?
Author here. Good question! /review is great for a quick pass on something you're still writing. I use it daily.
There are things it doesn't do well though. it can't take into account different priors. Built in reviews run on the same model family that wrote the code, so a subagent gets you fresh context and the same blind spots. Paranoia runs the review on another vendor's CLI. Here, Codex and Claude. For example, in my own work the same vendor declared converged and the other came back ten minutes later with two fatals and twelve majors on the same doc.
There's memory. Built in reviews are stateless, so round six happily re-reports round one's bug in a new spelling and nobody's any the wiser. Here the reviewer registers the defect class as a regex, the server re-runs it every round with git grep, and a computed trailer won't call it done while it still matches.
It's more expensive to set up, though, and require a subscription to both ChatGPT and Claude to get the most out of it. Not worth it for small changes, worth it for critical changes that are hard to walk back. For a fully automated harness, assuming you have max subs on both sides, it's perfect for making sure the agent is getting things right.