My current workflow: I use GitHub web to look at my commits and leave inline comments on the lines. Then I have Claude Code fetch all the comments and apply the changes or answer my questions.

I don't always have multiple commitsn: sometimes it's just one big commit that I then ask the AI to split, and usually after a full review. I wouldn't say it's a common use case, but it's mine.

To give you an idea of how much I try to optimize this part of my work, I installed Stylus (a Chrome extension) to change GitHub's syntax highlighting colors, so I'm glad you've integrated something similar natively.

That said, with my big commits of several hundred or even thousands of lines, your Stage tool and the hosted version are unusable compared to GitHub web. I think improving performance should be a priority, probably through virtualization (windowing).

Another issue: I never open PRs. As I mentioned, I comment directly on my commits on a branch. PRs make sense for a team workflow, but I work solo.

I've been spending a lot of my energy lately on how to run eng teams where we:

1. Maximize long-term shipping velocity

2. Maximize quality (whatever that means)

3. Maintain minimal complexity

4. Are intentional about which skills we let atrophy, which we keep sharp, and which new ones we have to build

5. Make juniors more capable, not just more productive

These are always in tension.

I've been thinking about instituting some sort of socratic method during planning and review plus spaced interval testing to ensure both the humans and AI coding agents understand and find some max of the factors above.

EDIT: I should mention that I think the idea is cool. We're in a new age where reviewing large amounts of unfamiliar code has become a larger problem than it was previously.

A normal git diff gets messy once the agent changes several files for different reasons. Grouping the change into “chapters” seems like the right idea.

Do you infer those chapters only from the diff, or can you also use the agent’s original plan/task history?

If this tool was in the terminal I'd use it.