Imun Farmer · Published:
- 예상 수확: 6 min read
oh-my-codex (OMX) — The Orchestration Layer That Turns Codex Into a Team
oh-my-codex (OMX) — The Orchestration Layer That Turns Codex Into a Team
Summary (3 lines) OMX is a workflow layer placed on top of OpenAI Codex CLI. It does not replace Codex — it adds multi-agent team execution, persistent state management, and standardized pipelines. Created by Korean developer Yeachan Heo, the open-source project surpassed 16,000 GitHub stars within two weeks of launch and has exceeded 29,600 stars as of May 2026.
Why It Exists — The Limits of a Solo Codex
Codex CLI is a powerful tool. But after using it for anything beyond single-file tasks, the limits become clear. There is no structured planning phase. State disappears when the session ends. There is no built-in way to coordinate multiple agents in parallel.
OMX was built to fill exactly that gap. The official docs use a direct analogy — “like oh-my-zsh, but for Codex.” Just as oh-my-zsh layered themes, plugins, and hooks on top of zsh without forking it, OMX leaves Codex’s code generation engine intact and changes only how work is organized.
First published on February 2, 2026, the current version is v0.18.3 (as of May 25, 2026). The codebase is 92.7% TypeScript and 4.7% Rust, published under the MIT license. The npm package name is oh-my-codex and the CLI command is omx.
The Core Philosophy — Codex is the Brain, OMX is the Office
The fastest way to understand OMX is to grasp the division of labor.
| Component | Role |
|---|---|
| Codex CLI | Actual code writing, editing, and execution (the brain) |
| OMX role keywords | Reusable agent role definitions (the manual) |
| OMX skills | Automated recurring workflows (the procedures) |
.omx/ directory | Persistent storage for plans, logs, memory, and runtime state (the office) |
Removing OMX returns you to a plain Codex session. That is why OMX is described as a “layer” rather than a “replacement.”
Raw Codex vs Codex + OMX — What Actually Changes
| Feature | Raw Codex CLI | Codex CLI + OMX |
|---|---|---|
| Planning phase | None | $deep-interview → $ralplan approval gate |
| Parallel workers | None | omx team N:executor (tmux-based) |
| Git isolation | None | Auto git worktree per worker |
| Session state | None (lost on context reset) | Persisted in .omx/ directory |
| Hook system | Not wired by default | .codex/hooks.json PreToolUse/PostToolUse |
| Memory across context resets | Lost | Priority notepad + project-memory.json |
| Specialized agents | None | 33 specialized prompts + 36 workflow skills |
Standard Workflow — The 4-Stage Pipeline
The core of OMX is the $deep-interview → $ralplan → $ultragoal pipeline. Each stage is not optional — it is a gate. Execution cannot proceed without approval from the previous stage.
$deep-interview — Understand First
Used when a request is still vague. This is the interview stage that clarifies intent, scope, and non-goals. Anyone who has ever jumped straight to coding only to receive a half-correct result will immediately understand why this step exists.
$ralplan — Get the Plan Approved
Converts the clarified request into a structured implementation plan with explicit trade-off review. Human approval is enforced; execution does not start without it. It is the safety gate against “let’s just try it” approaches.
$prometheus-strict — Harden High-Risk Plans
An optional step reserved for high-risk work. It stress-tests the plan through interview-driven critique and synthesis, and leaves artifacts in .omx/plans/prometheus-strict/.
$ultragoal — Durable Execution
Converts the approved plan into sequential Codex goals with .omx/ultragoal ledger checkpoints. Sessions can be interrupted and resumed. $ralph is an alternative single-owner completion loop without a multi-goal ledger.
# Standard OMX workflow
$deep-interview "clarify the scope of the authentication change"
$ralplan "approve the auth plan and review tradeoffs"
$prometheus-strict "stress-test the plan before durable execution"
$ultragoal "turn the approved plan into durable Codex goals"
Team Mode — Parallel Multi-Agent Execution
OMX team mode uses tmux to run multiple agents simultaneously. Each worker automatically receives its own isolated git worktree. Its value is most apparent in large-scale refactors spanning 15+ files, or in long-running tasks that outlast a single interactive session.
# Run a team with 3 executor workers
omx team 3:executor "fix the failing tests with verification"
# Check team status
omx team status <team-name>
# Resume a session
omx team resume <team-name>
When workers produce conflicts, the leader integrates changes and records them in integration-report.md. By setting OMXTEAMWORKERCLIMAP, Claude or Gemini can be mixed in as workers. Codex acts as the leader, with other AI models serving as subcontractors.
Persistent Memory — Remembers Across Sessions
A context window reset is one of the most frustrating moments in AI-assisted development. OMX addresses this structurally through project-memory.json and a priority notepad. Critical project decisions and task state survive session interruptions.
The .omx/ directory structure:
plans/— Approved implementation planslogs/— Execution logsultragoal/— Goal ledger checkpointsproject-memory.json— Cross-session persistent memory
Hook System — Intercept Every Tool Call
PreToolUse and PostToolUse hooks are natively wired through .codex/hooks.json. Destructive commands (deletions, overwrites) automatically trigger warnings. Errors surface structured guidance rather than silent failures.
The hook architecture has three layers:
plugins/oh-my-codex/hooks/hooks.json— Official hook registrations for plugin installs.codex/hooks.json— Legacy/fallback native hooks.omx/hooks/*.mjs— OMX plugin hooks
Installation and First Run
Installation takes two commands.
# If Codex CLI is already installed
npm install -g oh-my-codex
omx setup
# Fresh install
npm install -g @openai/codex
npm install -g oh-my-codex
omx setup
omx setup installs 33 agent prompts, 36 workflow skills, generates AGENTS.md, configures .codex/config.toml, and wires hooks in a single pass. Follow with omx doctor to verify the install.
The recommended launch command is omx --madmax --high. --madmax activates aggressive execution mode, --high enables high-performance parallel processing.
Core CLI Surfaces
| Command | Purpose |
|---|---|
omx --madmax --high | Recommended strong-start mode |
omx setup | Initial configuration and update |
omx doctor | Diagnose install health |
omx hud --watch | Live session monitoring |
omx explore --prompt "..." | Read-only repository lookup |
omx sparkshell <command> | Shell-native bounded verification |
omx wiki query --json | Project wiki search |
/skills | Browse installed skills |
OMX vs OMC — Two Easily Confused Projects
Two projects from the same developer (Yeachan Heo). Same design philosophy, different execution engines.
| OMX (oh-my-codex) | OMC (oh-my-claudecode) | |
|---|---|---|
| Execution engine | OpenAI Codex CLI | Claude Code |
| CLI prefix | omx | omc |
| npm package | oh-my-codex | oh-my-claude-sisyphus |
| Target user | Codex CLI users | Claude Code users |
Use OMX if Codex is your primary tool. Use OMC if Claude Code is your primary tool. Mixing Claude CLI workers inside an OMX team is still possible.
When OMX Is Not the Right Tool
OMX is not a universal solution. For targeted single-file changes or quick bug fixes, the planning gates add friction without payoff. For exploratory coding where structured gates would get in the way, it adds unnecessary overhead. On Windows without a WSL2 setup, production-grade reliability cannot be guaranteed — OMX officially targets macOS and Linux. Windows support exists but is explicitly marked as non-default.
Project Scale by the Numbers (May 2026)
- GitHub Stars: 29,600
- Forks: 2,400
- Contributors: 73
- Releases: 109 (v0.18.3 is current)
- Specialized agent prompts: 33
- Workflow skills: 36
- Supported languages: 16 (Korean, English, Japanese, Chinese, and more)
References
| No. | Source | Type |
|---|---|---|
| 1 | GitHub — Yeachan-Heo/oh-my-codex (Official) | Official repository |
| 2 | verdent.ai — What is oh-my-codex (OMX) | Technical guide |
| 3 | knightli.com — OMX Workflow Layer | Analysis document |
| 4 | findskill.ai — oh-my-zsh for Codex | Technical explainer |
| 5 | memoryhub.tistory.com — OMX 23k Stars Analysis | Tech blog |
| 6 | velog — Using oh-my-codex Hands-On | Practical guide |
| 7 | sourceforge.net — OMX Project Mirror | Mirror description |
| 8 | elancer.co.kr — Codex CLI Usage Guide | Usage guide |
Contribution to this Harvest
내용이 유익했다면 물을 주어 글을 성장시켜주세요!
(0개의 물방울이 모였습니다)