@badlogicgames If you are using a commit-based workflow you may find value in having the review strip concept from Reviewable. I added to my own version of pi-diff-review. The tool remembers what commit has been reviewed and defaults to showing the new changes.
@badlogicgames I was hoping this could be solved at the harness layer. Codex is really hates to pair program. I sometimes have more success by letting codex commit, make my own changes in a commit, then hand it back to codex.
@badlogicgames I have opinions. 1. Make the diffs editable (so you can add in-file review comments or fix things), 2. Mark files as reviewed (at that diff). IMO this is the best of Jane Street-style code review and Reviewable incremental review.
Great to see OpenAI demo how they use Codex internally. But at review time, it still feels a bit like Homer’s bird at the keyboard. Human-in-the-loop review matters, and I'm sure they do it. https://t.co/RpjQQcxxFL
@mitsuhiko I appreciate you posting the video. I struggled with a good review workflow.
Now I try to do almost everything in a single terminal split into three panes (agent | diff view | shell).
Wrote up my workflow https://t.co/MHsS2JBFUJ
@simonw Running multiple coding agents used to be too much cognitive load with window juggling and no flow.
Now: one Warp terminal (split into agent | diff | terminal panes), each git worktree in its own tab, plus one Cursor window to switch worktrees.
@CptKrrk@carsNbikelane I think both intersections (mystic and Everett) are particularly tricky. Left turns can be pretty fast 15-20mph so many times the car will not need to slow down to take the turn. Thus you cannot anticipate the left hook.
@joshm@arcinternet I use it but also enable it a lot by accident. The default keybinding clashes with forward-delete binding which I use a lot when writing in textfields. I guess I should change the shortcut. I need at least C-a, C-e, C-f, C-d, C-b when editing text.