Revdiff is how I solved it. From a Claude Code session I type `/revdiff main` or just say "review diff for last 3 commits" - it opens as a terminal overlay on top of the agent session. I read the diff, drop annotations on any line, hunk, or file, then quit. The annotations go straight back to the agent, which picks them up and starts working on them. When the agent is done, revdiff reopens on the new changes so I can annotate again - and this loop continues until I quit without leaving any annotations.
The same flow turned out to be useful for plans too. There's a separate `revdiff-planning` plugin that hooks into Claude Code's plan mode - the moment the agent finishes a plan and calls ExitPlanMode, revdiff opens automatically on the plan text. I annotate the parts I disagree with or want expanded, quit, and the agent revises the plan before writing any code.
It's a two-pane TUI: file tree on the left, syntax-highlighted diff on the right. Vim-style navigation, /search, hunk jumping, word-level change highlighting, blame gutters, collapsed diff mode. Fully customizable - 7 bundled color themes, remappable keybindings, and per-color overrides via CLI flags, env vars, or config file. Works with git and Mercurial.
It's a single binary - just `revdiff HEAD~3` to review your last 3 commits. Ready-to-use plugins are available for Claude Code (terminal overlay via tmux, Zellij, kitty, wezterm, Kaku, cmux, ghostty, iTerm2, or Emacs vterm - whichever you're in), Codex, and pi.
Written in Go, open source (MIT). Install via brew: `brew install umputun/apps/revdiff`
GitHub: https://github.com/umputun/revdiff Site: https://revdiff.com
Would love to hear how others approach reviewing AI-generated diffs and PRs, any unusual use cases, and ideas for what to improve or add.
