What it does

OMX is a TypeScript workflow layer that sits on top of the OpenAI Codex CLI. It keeps Codex as the execution engine but adds a structured path from clarification to completion: interview the problem, plan the approach, then execute through durable goals. It also manages git worktrees, tmux sessions, and project state under .omx/ so your AI-assisted work stays isolated and recoverable.

The interesting bit

The project treats prompt engineering as infrastructure. Instead of ad-hoc chatting, you invoke named workflow stages like $deep-interview, $ralplan, and $ultragoal — essentially turning a coding session into a state machine with predefined transitions. The --madmax flag (shorthand for bypassing Codex’s safety approvals) paired with automatic git worktree creation shows the authors are optimizing for speed in trusted environments, not demo safety.

Key highlights

• Canonical workflow: $deep-interview → $ralplan → optional $prometheus-strict → $ultragoal for multi-goal execution
• Git worktree isolation by default when using --madmax; concurrent tasks get separate checkouts automatically
• Tmux-based HUD/runtime panes on macOS/Linux for durable, recoverable sessions
• omx doctor and omx exec smoke tests to catch install, auth, and runtime problems before real work
• Ships both as global npm CLI (npm install -g oh-my-codex) and as an official Codex plugin, though the plugin is not a full replacement
• Durable state tracking under .omx/ for plans, logs, memory, and mode history

Caveats

• macOS or Linux with Codex CLI is the actively tuned default; native Windows and Codex App “may break or behave inconsistently” with less support
• --madmax removes approval and sandbox guardrails — the README repeatedly warns to only use it in trusted repositories
• Unnamed detached worktrees can hit worktree_target_mismatch if the source checkout advances; the README explicitly calls this a “one-off convenience”

Verdict

Worth a look if you’re already using Codex CLI daily and want more structure than raw prompting. Skip it if you want point-and-click AI coding or if you’re on Windows expecting first-class support.