Run your coding agents in tmux and let a local model tell you which replies to read A developer built TLDRack, an open-source (MIT) tool that runs coding agents such as Claude Code and Codex side by side in tmux and displays them on a single local page. A small local model reads each agent's closing message and grades whether work is still open, cutting the 1,523 prompts the developer sent over two weeks down to 696 messages that needed reading. The developer cautions that the model judges the message rather than the work, so a turn that falsely claims completion still reads as done. In two weeks I sent my coding agents 1,523 prompts and read every closing message. I was fed up, so I put a small local model in front of them. It left me 696 to read. The tool is TLDRack. It is open source MIT and runs on macOS with Claude Code or Codex. Here is how to get it running and what you get. TLDRack runs your coding agents side by side in tmux and shows every one of them on one local page. Sessions name themselves from your prompt, so you can find a task at a glance. When a turn finishes, a local model reads the closing message and marks it done or needs you. You need git, tmux, ttyd, Node 22.18 or newer, uv, and one agent CLI you have logged in to once. brew install tmux ttyd node uv git clone https://github.com/voidcraft-world/tldrack cd tldrack uv tool install --editable ./bridge puts tldrack on your PATH tldrack doctor every missing part prints its fix tldrack hooks --install your agent CLIs report their state to the rack cd /path/to/your/project new agents start where you run tldrack up tldrack up builds the page, then opens http://127.0.0.1:7682/ tldrack laya install the model's runtime and base weights, about 800 MB It takes about ten minutes, and most of that is downloads. One row per agent session, each with a name it gave itself and a grade once its turn ends. Next to that: the selected agent's terminal, its worktree, a live view of what every agent is doing, and any files or drawings it made for you to review. The grade is the part that saves reading. The model answers one question about the closing message: does this say work is still open? It works with no training at all, and that is your first run. After that you fix its misses with one tap per turn, and those grades train a version fitted to your agents. That ranking is my judgment from running all three. I have not measured it. The model judges the message, not the work. A turn that claims it is done and is not still reads as done. Treat the grade as where to look first. Video one minute : https://youtu.be/dnTCRYkfKiY https://youtu.be/dnTCRYkfKiY