Built my own tui tool. It groups repos into projects and manages a task list for each repo. Creating a new task creates a new worktree, the task view contains two tmux sessions, one of which is a claude or codex session, the other a terminal. I have status dots for the current state of claude/codex and the current state of the pr if it exists. And i have a flow to review and annotate pr feedback and auto paste it back into my llm session with the instructions to address it and push. I have a task switcher which incorporates the status colors so i can tend to the red tasks, wait on the yellow ones, merge the green ones and close out the purple ones. Each task has a status that includes the last thing i typed to the llm. I also have hotkeys to launch cursor, embedded lazygjt, tower, or jump to the pr in the browser. I integrated linear and can create a task from the linear ticket and pull the ticket and all the comments into the context. People keep telling me i should add audio alerts, but i never have to step away. If everything is yellow i just start a new task.

Currently I mostly use codex app (desktop one) and ghostty terminal.

For the code I usually started from codex app to implement new feature or fixing bugs. And after that I review the code changes on ghostty using `lazygit`. Somehow it easier for me to check the code changes on lazygit compared to check it on Codex app.

After that I'll usually try to run test and run the app from terminal (`npm` or `mvn` command) after that try to test locally on browser for FE apps, and using Bruno for the APIs.

vscode with claude and codex extensions. two running chats are fine for me, allows me to keep my eyes on the code rather than running 75 agents.

[dead]

[dead]