Replacing Pull Requests with Delta

10 points by caius

Splitting a big diff across a stack of branches can make it easier to navigate, but the decisions behind the code still need review. Smaller diffs don't supply that context. A reviewer may feed your diff into another agent to help understand it, but that agent has to piece together decisions you already worked through.

Some thoughts:

For large and complex plan, you can write a document first and have people review that in a PR.
The commit message/PR description can be used for high-level decisions (and/or you can use named notes a la GHC).
You can write code comments for non-obvious low-level decisions.
For decisions that are specific to a particular time/things you're not sure about whether they belong in code comments, you can leave an inline comment on the PR for explanation.

Expecting reviewers to trawl through threads of slop breaks the interface of review & the responsibility on the author that comes with it.

Sure, making collaboration have less friction is a good idea. But collaborative development/pair programming and code review are separate things.

stig

Seconded. I have observed several times that an agent loop that includes yes of a "code-review" skill, and iterate until no more findings, produced a PR that: when using the same "code review" skill on the PR in a fresh session, with the same model, had several "major" review findings. When I queried the implementing agent about this, it told me that it was because it told the review agents to ignore those findings.