We never accepted 3k lines of code in a PR from a human, why do we from LLMs?
And if we're going to, why are LLMs not doing the PR review?
And if we don't trust them, then isn't that a rather large obstacle to the idea that LLMs are going to replace humans in software engineering?
You can defer review entirely to an LLM, only to human reviewers, or both. We are experimenting with both and have configured LLM skills for pull request reviews. On large projects we also configure LLMs to produce commit segregated diffs within a single pull request, to aid human review. This process is using the github pull request workflow that Zed is attempting to replace.