Cloud Swarms and Async AI Code Reviews
Explore how cloud-based ultrareview offloads AI code reviews to a multi-agent swarm, handling branches and PRs asynchronously without clogging your local terminal. The episode also covers recent fixes for merge-base edge cases, smarter PR parsing, and tighter billing safeguards for premium reviews.
Show Notes
- Ultrareview for Claude Code: Multi-Agent Cloud Code Review ...: https://claudedirectory.org/blog/ultrareview-claude-code-guide
Chapter 1
Cloud-Hosted Swarms: /ultrareview Unlocks Async Code Reviews
Lachlan Reed
So, I- I was out in the shed the other night, right? Working on the carburetor for the old DT175, and... and I had this massive local test suite running on my laptop. Just... sitting there, fan screaming like a gecko, totally locking up my terminal. And it got me thinking, why are we- we still doing this to ourselves with local AI code reviews when we've got the cloud?
James Turner
Exactly! It is that local bottleneck. When you run /code-review locally, it completely eats up your terminal session and your interactive context window. But /ultrareview... that's a whole different animal. It runs a multi-agent cloud review of your current branch or a GitHub PR, offloading all that heavy lifting to Anthropic's server infrastructure.
Lachlan Reed
Yeah, nah, it's brilliant. It kicks off a whole swarm, mate. Like, parallel agents for security, performance, correctness... but- but look, it used to have some real, uh, rough edges. Like, if you tried to run /ultrareview on a brand new repo or an isolated branch with no common git merge base... bam. Hard crash. Refused to run.
James Turner
Oh, I remember that error. It was so frustrating. But they just resolved that in the latest patch! Now, instead of throwing a fit, if there's no merge base, it actually offers to review all tracked files. It just falls back gracefully.
Lachlan Reed
Tracked files! That's- that's a massive win. Saves you having to, like, fake a commit or merge main in just to get a quick sanity check. And... and the branch parsing too. You don't have to get the formatting absolutely spot-on anymore. It- it takes human formats, like, uh, #123, or even just raw copy-pasted PR URLs.
James Turner
Yeah, PR 123, #123, whatever you throw at it. And if you make a typo on a branch name, it'll actually search the remote origin and suggest the closest match. But wait, did you see the security fix? There was this really wild edge case with the /clear command.
Lachlan Reed
Wait, /clear? What- what was happening there?
James Turner
So, if you ran /clear in an interactive session, it would sneakily bypass the cloud billing confirmation on any subsequent /ultrareview runs. So you could end up with unexpected billing spikes because it didn't ask you "Are you sure?" the next time.
Lachlan Reed
Crikey. That's a- a nasty one. Glad they patched that up. I don't want a rogue swarm running up a massive bill while I'm off making a cuppa.
James Turner
Definitely not. Especially now that the underlying engine has transitioned to claude-opus-4-8. The findings quality is way higher now, but you definitely want to keep an eye on those API costs.
Lachlan Reed
Yeah, claude-opus-4-8 is a beast. But that's also why /ultrareview is metered completely separately from your standard interactive limits. And get this--you can't have, like, automated subagents or loops trigger it. They locked that down so you don't get some runaway recursive loop eating your wallet.
James Turner
Which makes total sense. You want that human-in-the-loop confirmation before you launch a premium cloud swarm. It's perfect for running right before you open a PR, or just pointing it directly at a PR URL to let it run asynchronously.
Lachlan Reed
Spot on. No more waiting around the shed for local terminal spinners. Alright, I'm off to actually finish that carburetor. Talk soon, mate.
James Turner
Later, Lachlan.