Check out the branch, run the code locally, read it, ask claude code or codex or whatever harness to explain things you don't understand. Ask it to diagram the code boundaries and modules and model the system so you can visually understand it.
use AI to write smoke tests without access to the code i.e. first describe something a correct implementation should do with only the knowledge of the interface not the implementation, and run the tests to see if it behaves the way you think it should.
Would recommend checking out coderabbit change stack. it essentially does exactly what you're looking for
Cloudflare has a good post on the Ai or review system. We modeled ours after it. Much of it comes down to two things
1. Removing the mechanical parts from agent responsibility in your CI workflow
2. Having good instructions and subagents / team / loop
Check out the branch, run the code locally, read it, ask claude code or codex or whatever harness to explain things you don't understand. Ask it to diagram the code boundaries and modules and model the system so you can visually understand it.
use AI to write smoke tests without access to the code i.e. first describe something a correct implementation should do with only the knowledge of the interface not the implementation, and run the tests to see if it behaves the way you think it should.