A human in the loop TUI to review and revise LLM generated markdown plan files quickly.
For agent-loop usage, wrap your agent harness in a tmux session so the skill can launch the rep UI modally.
rep opens a markdown file in an interactive TUI optimized for providing feedback and requesting changes from an LLM.
When you exit rep, it prints a list of requested changes for an AI agent to process and apply to the markdown plan. These changes can be a mix of deletions, additions, changes, and intent guidance.
For the most seamless experience, launch Codex or Claude Code inside a tmux session, which allows rep to launch as modal from a skill and automatically pass the revisions you ask it for back into the agentic loop.
Install the latest macOS or Linux release to ~/.local/bin, and the bundled agent skill source to ~/.agents/skills/rep, with this command:
curl -fsSL https://raw.githubusercontent.com/mattorb/rep/main/install.sh | shAfter installing the skill source, the installer offers to symlink it into the supported agent skill directories for Claude, Codex, Gemini, opencode, Hermes, and Droid. The prompt works with the curl | sh form when an interactive terminal is available.
To install or update only the skill source and optional agent symlinks later:
curl -fsSL https://raw.githubusercontent.com/mattorb/rep/main/install.sh | sh -s -- --skills-onlyThis source install path installs the rep binary only.
Use the rep skill to launch in the agentic loop: /rep in Claude Code or $rep in Codex.
If you execute it manually like this, you'll have to copy and paste results back to the agent.
rep plan.md| Key | Action |
|---|---|
j, Down, Right | Move to the next active unit |
k, Up, Left | Move to the previous active unit |
Space | Cycle to the next selection unit |
Backspace | Cycle to the previous selection unit |
i | Use a finer selection unit |
o | Use a coarser selection unit |
c | Add or edit a literal change request |
f | Add or edit feedback or intent |
b | Insert text before the current unit |
a | Insert text after the current unit |
x | Clear existing annotations or mark the unit for deletion |
e | Edit an existing annotation |
[, ] | Jump to the previous or next annotation |
/ | Search |
n, N | Jump to the next or previous search match |
?, Shift + / | Open or close help |
I | Open or close the AST view |
O | Reveal markdown links for the current sentence |
r | Copy annotations to the clipboard |
q | Quit and print annotations to stdout |
Q | Quit silently and discard annotations |
Enter | Save text in change, feedback, insert, edit, or search modes |
Esc | Cancel the current input mode or close an open popup |
| Platform | Release artifact | CI coverage | Support status |
|---|---|---|---|
| Linux x86_64 | x86_64-unknown-linux-musl | Build, package, archive smoke test, and installer smoke test on GitHub-hosted Ubuntu | Supported |
| Linux arm64 | aarch64-unknown-linux-musl | Cross-build, package, archive smoke test, and installer smoke test on GitHub-hosted Ubuntu | Supported |
| macOS x86_64 | x86_64-apple-darwin | Build and tests on GitHub-hosted macOS | Supported |
| macOS arm64 | aarch64-apple-darwin | Cross-target release build on GitHub-hosted macOS | Supported |
Linux release artifacts are static MUSL builds.
On exit, something like this prints to stdout, for the LLM agent to consume and make modifications to the markdown plan file.
FILE: plan.md
ACTION: change
WHERE: line 12 sentence 2
CONTEXT:
prev: The release workflow builds archives for every configured target.
target: Windows artifacts are published even though the installer and tests do not cover Windows.
next: Checksums are generated after packaging.
CHANGE: Stop publishing Windows archives until CI and installer support are added.
MIT — see LICENSE.