Turn Superlogical's Rex into an AI Harness

demo.mp4 #

_{One minute, sped up while the agent works. No player? Download the video.}

Rex already runs your agents. rexmod builds the harness around them: it records every turn, so you see exactly what the agent changed; it opens a real editor beside the agent to read and review that code; and it hands your review back to the agent as its next input. Prompt, change, review, feedback: the whole loop stays in one Rex window, for Claude Code, Codex, OpenCode or Pi.

Note

rexmod is an unofficial mod, not made by or affiliated with Superlogical. It patches your local copy of Rex so Rex loads the panel; the uninstaller puts the official build back.

1. Prompt. Talk to your agent as usual. rexmod snapshots the working tree the moment you send.
2. Change. TheChanges panel on the right, styled like Rex's own sidebar, fills in while the agent works:Last Turn is exactly what this reply changed, next toUncommitted ,Branch and aFiles tree.
3. Review. Click a file and it opens in Neovim beside the agent, at the changed line, with LSP, completion and live diagnostics.Space r on a line or a selection leaves a review comment.
4. Feedback. Your comments ride along with your next message to the agent, or go right away from the panel. The agent answers them, and the next turn starts at step 2.

Around the loop:

• Any language, on demand. Open a Go, Rust, Python, Elixir… file and its tree-sitter parser and language server are installed (Mason) and started on the spot.
• ⌘P file picker over the agent's repository, with the editor's syntax highlighting in the preview.
• Several agents per tab. The harness follows the agent you focused last; its name at the bottom of the panel is a menu to switch. A note always goes to the agent it was written for.
• Media previews. Images and PDFs open as pictures in the editor, videos show their first frame (Space u o plays them in the macOS app), and markdown renders in place, images included (Space c p opens a live browser preview).
• The agent knows the harness. Every agent gets therexmod skill, so it can open a file for you in the editor pane (rexmod open file:line ) when it points you at something.
• Out of the way. Hidden, the panel takes no space; the toggle sits with the pane actions,⌘⇧E from the keyboard. When Rex updates itself, rexmod re-applies the patch and asks you to reopen it.

You need a Mac with Apple silicon on macOS 15 or later, Rex installed and opened once, and Homebrew (for Neovim, fzf and Node, if you don't have them).

With Homebrew:

brew trust --tap alpcanaydin/rexmod   # Homebrew asks you to trust third-party taps
brew install alpcanaydin/rexmod/rexmod
rexmod-setup

Or in one line:

curl -fsSL https://raw.githubusercontent.com/alpcanaydin/rexmod/main/install.sh | sh

Then quit Rex (⌘Q) and open it again. Nothing is compiled on your machine. After brew upgrade rexmod, run rexmod-setup again.

What the installer does #

• Copies rexmod into ~/Library/Application Support/rexmod and links~/.local/bin/rexmod .
• Downloads the free Flow Icons pack for the file tree.
• Editor: with no ~/.config/nvim , installs the bundled LazyVim config there; with a lazy.nvim config, adds one plugin spec (lua/plugins/rexmod.lua ); with any other config, installs the bundled one as~/.config/rexmod-nvim and leaves yours alone.
• Adds ⌘P to~/.config/rex/init.lua (in a managed block).
• Hooks into each agent it finds and installs the rexmod skill. JSON files are backed up first.
• Backs up the official Rex, patches it, and adds a LaunchAgent that re-patches after Rex updates.

Options: --update-rex installs the latest official Rex first, --no-patch skips the panel, --no-hooks skips the agents. From a checkout, ./install.sh builds what it needs if Go and Xcode are there.

Important

Codex runs new hooks only after you trust them: start Codex and run /hooks once. OpenCode loads plugins when its background server starts, so quit OpenCode once after installing.

The comment badge in the panel header sends pending comments right away. Without it they ride along with the next message you type to the agent, as a short list of file:line, the code, and your note.

Several agents in one tab. The panel, Last Turn and your comments follow the agent pane you focused last; working in the editor or a shell doesn't change that. The agent name at the bottom of the panel becomes a menu to switch.

flowchart LR
  subgraph Rex.app
    panel[Changes panel<br/>RexModPanel.dylib]
  end
  agent[Agent pane<br/>Claude Code · Codex · OpenCode · Pi]
  editor[Editor pane<br/>Neovim + rexmod.nvim]
  cli((rexmod))
  git[(.git/rexmod<br/>turns · comments)]

  panel -- "state, open, send" --> cli
  agent -- "hook prompt / stop" --> cli
  editor -- "comments" --> cli
  cli -- "panes, splits, input" --> rex[Rex server API]
  cli --> git
  cli -- "--listen socket" --> editor

• The panel is a Swift dylib. The installer adds a weak load command for it to Rex's executable and re-signs the app ad hoc. It insets Rex's own chrome and draws on Rex's backdrop, so it looks native.
• rexmod , a Go CLI, does the work: it finds the tab's agent pane through Rex's API, diffs git trees, opens files in the editor pane, and keeps comments.
• Last Turn comes from two snapshots of the working tree (written with a throwaway git index, so your index is untouched): one when you send a message, one when the agent finishes.
• Comments live in.git/rexmod/ of the worktree, each tied to the agent pane it was written for.

Updates: brew upgrade rexmod && rexmod-setup, or run the one-line install again.

Rex itself: Sparkle can't update a patched Rex (its installer only accepts an app signed by Superlogical), so rexmod turns Sparkle's checks off and its LaunchAgent updates Rex instead, every hour: it downloads the official build, checks its signature, Team ID and notarization, patches it and asks you to reopen Rex. To do it now: ~/Library/Application\ Support/rexmod/scripts/update.sh.

Uninstall, which restores the official Rex and removes the hooks, the ⌘P block and the files:

~/Library/Application\ Support/rexmod/scripts/uninstall.sh

The panel doesn't show up #

Check that Rex is patched: ~/Library/Application\ Support/rexmod/scripts/patch.sh --check. If Rex just updated, reopen it once the re-patch notification arrives. The panel logs to ~/Library/Logs/rexmod-panel.log; start Rex with REXMOD_PANEL_DISABLE=1 to run it without the panel.

Last Turn stays empty #

The agent's hooks aren't running. Run ~/Library/Application\ Support/rexmod/scripts/install-hooks.sh again, then for Codex trust the hooks with /hooks, and for OpenCode quit it once.

Do my comments go to an agent outside Rex? #

No. A note is tied to the agent pane of the tab you wrote it in. An agent running elsewhere (another terminal or app) in the same repository only gets notes written outside Rex.

Two Codex panes in the same repository #

Codex runs its hooks in one background server for every pane, so rexmod finds the pane by looking for the Codex that works in that folder. With two of them in the same repository it can't tell which one sent the message, and rather than guess it leaves the notes pending and records the turn without a pane. Send those notes with the comment badge in the panel: it pastes them straight into the pane the panel follows. Claude Code, OpenCode and Pi don't have this limit.

Debugging the panel #

Darwin notifications drive it from a shell: notifyutil -p dev.rexmod.toggle (also .refresh, .snapshot, .dumpviews, .dumpchrome, .dumpheader, .dumpsidebar). Output lands in ~/Library/Caches/rexmod/.

make build, make panel, make test, make lint. See CONTRIBUTING.md for the setup and how the pieces fit together.