Ultrareview: Bug Hunting in the Cloud

Shipped alongside Claude Opus 4.7, /ultrareview is the heavyweight cousin of the /review command we already covered. Where /review is a quick sanity check you run between keystrokes, /ultrareview is what you fire up before pushing that big refactor to the main branch – the one that touched half your DPR, three packages, and the persistence layer.

What It Actually Does

Type /ultrareview at the prompt and Claude Code does something unusual: it stops reviewing the code on your machine. Instead, it packages your branch (or pulls a PR directly from GitHub) and ships it off to a sandbox running on Claude Code’s web infrastructure. Inside that sandbox, a small fleet of reviewer agents goes to work in parallel. Each one pokes at the diff from a different angle – one looks for memory leaks and dangling references, another hunts exception-handling holes, another complains about range checks or unfreed objects, and so on.

The trick that separates it from every other automated reviewer is verification. When one agent claims “you forgot to FreeAndNil that TStringList on line 214,” a second agent has to reproduce the finding independently before it reaches you. Bugs that cannot be reproduced get dropped on the floor. What lands in your terminal five to ten minutes later is a short list of issues that already survived a skeptical second pair of eyes. For a Delphi project, where a single missed FreeAndNil or a dangling pointer will happily corrupt memory in production for weeks before anyone notices, that verification step matters.

How to Run It

From any git repository – which, yes, means your Delphi project must live under git – you simply type:

/ultrareview

With no arguments, it reviews the diff between your current branch and the default branch, uncommitted changes included. That is usually what you want: you have been hacking on a feature branch, the .pas files are a mess of half-saved work, and you want someone to look at the whole thing before you merge.

To target a pull request instead, pass its number:

/ultrareview 1234

In PR mode the sandbox pulls the code straight from GitHub. This is also the fallback when your local repository is too big to bundle and upload – not uncommon for a Delphi project that drags along c:\Projects\LightSaber\, half of c:\Projects-3rd_Packages\, and the usual pile of compiled DCUs.

Before the review launches, Claude Code shows a confirmation screen with the scope, your remaining free runs, and the estimated cost. You confirm, and the review keeps running in the background. Close the terminal, switch to the IDE, go compile something else – when the verified findings are ready, they show up as a notification in your session. Each finding includes a file path and line number and a short explanation, and you can hand it straight back to Claude with a “fix this” prompt.

Running reviews are visible through /tasks, where you can also stop one that is no longer useful. Stopping discards the partial findings; there is no peeking at the agents while they are still arguing among themselves.

/review vs /ultrareview

Think of /review as the Error Insight squiggles the Delphi IDE shows while you type, and /ultrareview as a proper code review from a grumpy senior colleague. You use the first one all day long. You call the second one in only when the code is almost ready to ship.

The Catch

Ultrareview is a premium feature and it bills against extra usage, not against your plan’s included quota. Pro and Max subscribers get three free runs – a one-time allotment, not a monthly refill. After that, each review costs roughly five to twenty dollars depending on the size of the change, and “extra usage” has to be enabled on your account before Claude Code will even launch a paid run.

It also has a few hard limits worth knowing about. You must be signed in with a Claude.ai account; an API-key-only login will not do. It is unavailable on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, and organizations with Zero Data Retention turned on cannot use it at all. These are the direct consequence of the fact that your source really does leave your machine and run in Anthropic’s cloud. If you are working on a closed-source commercial Delphi product, talk to whoever owns the NDA before you type the command.

One more limitation that bites Delphi programmers in particular: /ultrareview needs a git repository. The older habit of keeping a Delphi project as a loose folder on c:\Projects\ – no git, just zip backups every few days – will not work here. Neither will a SVN or Mercurial checkout. If you want to use this command on, say, c:\Projects\ProjectAssist\, you first need to git init the folder, commit a baseline, and do your changes on a branch. Not a bad habit to pick up anyway.

When It Earns Its Keep

Not every change deserves an ultrareview. A one-line typo fix in a label caption does not. A branch that touches six units, rewrites how your application saves its INI file, and changes the shape of a few public properties almost certainly does. The value is highest exactly where human reviewers are weakest: large diffs where the eye glazes over and subtle bugs slip through because a single reviewer, reading once, cannot hold the whole thing in their head. A forgotten FreeAndNil in a rarely-hit branch, an exception silently swallowed three levels down, an off-by-one in a for i:= 0 to High(Arr) that only trips when the array is empty – these are the kind of bugs a verified multi-agent pass is good at catching, and the kind a tired human reviewer on a Friday afternoon is good at missing.

Used that way – sparingly, before merging something you actually care about – the three free runs alone can pay for themselves in bugs caught. And the ones you pay for afterwards are cheap compared to the bug reports you would otherwise get from users two weeks after release.