oh-my-codex (OMX)
Start Codex stronger, then let OMX add better prompts, workflows, and runtime help when the work grows.
Liked OmX but find it a bit overkill? Try gajae-code.
Keep Codex OAuth with a faster, cheaper, simpler, and more powerful SDK-based path for OpenClaw, Hermes, Grokbot, and other integrations.
Website: https://yeachan-heo.github.io/oh-my-codex-website/
Docs: Getting Started · Agents · Skills · Integrations · Demo · OpenClaw guide
Community: Discord — shared Gajae community server for oh-my-codex, gajae-code, and related tooling.
Official project and package
The official/original OMX project is this repository, Yeachan-Heo/oh-my-codex, and the official npm package for this project is oh-my-codex. Install this project with npm install -g oh-my-codex (or alongside Codex CLI as shown below).
Third-party projects or forks that use names such as “OMX v2” are not official continuations, replacements, or release lines for this repository unless this README or the docs explicitly say so. When in doubt, trust this repository and the oh-my-codex package as the official install target.
OMX is a workflow layer for OpenAI Codex CLI.
🚨 CAUTION — RECOMMENDED DEFAULT ONLY: macOS or Linux with Codex CLI.
OMX is primarily designed and actively tuned for that path.
Native Windows and Codex App are not the default experience, may break or behave inconsistently, and currently receive less support. |
It keeps Codex as the execution engine and makes it easier to:
- start a stronger Codex session by default
- run one consistent workflow from clarification to completion
- invoke the canonical workflow with
$plan, $ultragoal, $team, $code-review, and $ultraqa — each independently, no fixed chain
- keep project guidance, plans, logs, and state in
.omx/
Core Maintainers
Ambassadors
Top Collaborators
Recommended default flow
If you want the default OMX experience, start here:
Choose one install path. If Codex CLI is already installed (Homebrew, npm, or another supported method):
codex --version
npm install -g oh-my-codex
# from the git project you want Codex to edit; choose a task-specific name
omx --worktree=feat/task --madmax --xhigh
If you do not have Codex CLI yet and want npm to manage it:
npm install -g @openai/codex
npm install -g oh-my-codex
Do not run a combined npm install -g @openai/codex oh-my-codex over an existing Homebrew-owned codex binary such as /opt/homebrew/bin/codex; npm may fail with EEXIST when @openai/codex tries to create the same binary. OMX only needs a working, authenticated codex command on PATH; it does not require Codex to be installed through npm.
On a real oh-my-codex version bump, the global npm install now prints an explicit reminder instead of launching setup automatically. When you're ready, run the scoped setup command below or use omx update to check npm and then run the same setup refresh path.
OMX also checks for npm updates at launch on a throttled cadence and prompts before scheduling the update after the current session exits. Set OMX_AUTO_UPDATE=0 to disable the launch-time check, or set OMX_AUTO_UPDATE=defer to schedule the same deferred update without prompting.
Choose the setup scope deliberately:
- Use
omx setup --scope project --merge-agents from the git project you want OMX to operate on when that repository should own the durable AGENTS.md guidance.
- Use
omx setup --scope user for user-level Codex setup when you are not preparing the current directory as an OMX project.
- Avoid running project-scoped setup from a broad home directory or operating hub unless that directory is intentionally the project under review. A home-level
AGENTS.md often contains global safety and routing rules; keep project-specific OMX runtime guidance in the real repository instead.
Persisting an explicit AGENTS merge policy
omx setup --merge-agents, omx setup --no-merge-agents, and omx setup --clear-merge-agents-policy are the only policy selectors; use their bare forms (not =value spellings). Repeating an identical selector is harmless, but mixing set and clear choices fails before setup changes anything. An explicit set overrides a saved policy. A successful explicit set records mergeAgents: true or false in the current working root's ./.omx/setup-scope.json, even when setup scope is user; it never becomes a global user preference or leaks to another root. Later omx update replays a valid matching policy for both immediate and deferred refreshes.
true takes the existing merge branch. false only suppresses that branch: it does not promise preservation, replacement, or any new safety mode, so the existing prompt, skip, managed-refresh, plugin-default, and force behavior still applies. A matching-scope review retains the policy while unrelated settings change. Reset or a scope change removes the inherited policy unless the same setup run explicitly sets true or false; clear always removes the policy and cannot be combined with a set selector. Malformed, unknown, nonboolean, or wrong-scope saved data is ignored safely.
--force is independent and transient: it is neither recorded nor replayed, and does not override an explicit merge policy. Setup atomically commits explicit set or clear intent only after all setup work succeeds, including when active-session or plugin-symlink safeguards skip the current AGENTS.md write, so the next refresh can honor the requested policy. This does not make merge the default or revive the rejected #2892 merge-by-default approach. Older OMX versions safely ignore the field, but may erase it when rewriting their known setup preferences.
Codex plugin install note: this repo also ships an official Codex plugin layout at plugins/oh-my-codex with marketplace metadata in .agents/plugins/marketplace.json. That plugin bundles the mirrored skill surface plus plugin-scoped companion metadata for official Codex lifecycle hooks, optional MCP compatibility servers, and apps. It is still not a replacement for the global oh-my-codex CLI plus scoped setup: plugin-scoped hooks launch the installed omx CLI, legacy setup mode installs native agents and prompts, and plugin setup mode relies on plugin discovery for bundled skills while archiving/removing legacy OMX-managed prompts/native-agent TOMLs so stale role files cannot shadow plugin behavior. Plugin mode still needs a persistent scope AGENTS.md (~/.codex/AGENTS.md for user setup or ./AGENTS.md for project setup) as the durable orchestration guidance layer; session-scoped AGENTS files only compose that durable guidance with runtime overlays and are not a replacement.
Then work normally inside Codex:
# Durable objective/checkpoints for a long task:
/goal Create a safe authentication refactor plan, implement it, and verify login, logout, and refresh-token behavior.
$deep-interview "clarify the authentication change"
$ralplan "approve the auth plan and review tradeoffs"
$ultragoal "turn the approved plan into durable Codex goals"
That is the main path.
Before you treat the runtime as ready, run the quick-start smoke test below: omx doctor verifies the install shape, while omx exec proves the active Codex runtime can actually authenticate and complete a model call from the current environment.
Start OMX strongly, clarify first when needed, approve the plan, then use $ultragoal as the default durable completion wrapper. Use $team inside that execution path only when a specific Ultragoal story needs coordinated parallel work; use $ralph when you intentionally want a single-owner completion loop instead of a durable multi-goal run.
What OMX is for
Use OMX if you already like Codex and want a better day-to-day runtime around it:
- a standard workflow built around
$deep-interview -> $ralplan -> $ultragoal
- research boundaries: use
$best-practice-research for ordinary pre-planning official/upstream evidence, $autoresearch for bounded validator-gated research artifacts, $autoresearch-goal for goal-mode research missions, and feed any research findings into $ralplan for architecture synthesis
- durable multi-goal handoffs with
$ultragoal and .omx/ultragoal artifacts as the default completion path after planning
- specialist roles and supporting skills when the task needs them
- project guidance through scoped
AGENTS.md
- durable state under
.omx/ for plans, logs, memory, and mode tracking
If you want plain Codex with no extra workflow layer, you probably do not need OMX.
Quick start
Requirements
- Node.js 20+
- Codex CLI installed, verified with
codex --version, and authenticated (Homebrew or npm are both fine; do not reinstall @openai/codex with npm if Homebrew already owns codex)
- Codex auth configured and visible in the same shell/profile that will run OMX
tmux on macOS/Linux if you want the recommended durable team runtimepsmux on native Windows only if you intentionally want the less-supported Windows team path
A good first session
After install, check both boundaries:
omx doctor
codex login status
omx exec --skip-git-repo-check -C . "Reply with exactly OMX-EXEC-OK"
omx doctor catches missing OMX files, hooks, and runtime prerequisites. The real smoke test catches auth, profile, and provider/base-URL problems that only appear when Codex performs an actual request.
Launch OMX the recommended way from a git project:
omx --worktree=feat/task --madmax --xhigh
On macOS/Linux interactive terminals with tmux available, this starts the
leader in OMX-managed detached tmux by default so the HUD/runtime panes can be
created and recovered. --worktree also moves the launch into a separate git
checkout, which is the safer default when using --madmax. Replace
feat/task with a branch-like name for the task.
Concurrent standard conversations
A standard launch owns one writable session pointer under its selected OMX_ROOT. A second ordinary omx launch from the same checkout therefore fails closed instead of sharing or silently changing that root. Give each additional conversation an explicit, distinct root:
omx
OMX_ROOT="$HOME/.omx/instances/second-conversation" omx
OMX_ROOT="$HOME/.omx/instances/third-conversation" omx
PowerShell:
$env:OMX_ROOT = "$HOME/.omx/instances/second-conversation"
omx
Command Prompt:
set "OMX_ROOT=%USERPROFILE%\.omx\instances\second-conversation"
omx
User-specified roots are literal: launching twice with the same explicit OMX_ROOT remains a fatal owner conflict. Separate checkouts have separate default roots, while --worktree and --madmax keep their existing isolation behavior.
Madmax and worktree launch safety
--madmax is OMX shorthand for Codex
--dangerously-bypass-approvals-and-sandbox. It removes the normal approval and
sandbox guardrails, so only use it in trusted repositories and environments.
--high and --xhigh are shorthand for -c model_reasoning_effort="high|xhigh"; a normal strong session is omx --madmax --xhigh (or omx --worktree=feat/task --madmax --xhigh from a git project).
When you use --madmax from a git repository, prefer a worktree launch instead
of running directly in the current checkout. For repeatable or concurrent work,
use a named worktree:
omx --worktree=feature/auth --madmax --xhigh
If you are outside a git repository, omit --worktree; worktree launches
require Git.
For concurrent --madmax sessions, do not run them all in the same
directory. Give each session its own named worktree:
omx --worktree=feature/auth --madmax --xhigh
omx --worktree=fix/flaky-tests --madmax --xhigh
--worktree / -w with no name creates or reuses a detached launch worktree at
../<repo>.omx-worktrees/launch-detached. --worktree=<name>,
--worktree <name>, or -w <name> creates or reuses a named launch worktree
under ../<repo>.omx-worktrees/ and checks out that branch name. OMX consumes
the worktree flag before starting Codex; it is not forwarded to Codex itself.
Treat the unnamed detached form as a one-off convenience: if the source checkout
advances after that worktree is created, a later unnamed launch can fail with
worktree_target_mismatch because launch-detached still points at the old
HEAD. Use a named worktree for repeated work, or remove the old detached
worktree before retrying.
If the target launch worktree is already dirty, OMX warns and launches as-is, so
clean, commit, or stash that worktree before relying on it for isolation.
For omx team, workers already use dedicated worktrees automatically by
default; --worktree on omx team is only a legacy-compatible override.
Repo-aware tools receive the same canonical context in launch, team-worker, and autoresearch runtimes: OMX_REPO_ROOT, OMX_WORKTREE_ROOT, OMX_GIT_COMMON_DIR, OMX_WORKTREE_SCOPE, OMX_CODEGRAPH_MODE, and OMX_CODEGRAPH_PROJECT_PATH. OMX_CODEGRAPH_MODE=auto prefers a worktree-local .codegraph/codegraph.db, then a leader/repo .codegraph/codegraph.db, otherwise resolves to off. Explicit shared, local, and off are honored. OMX does not install CodeGraph, auto-index worktrees, or copy/symlink .codegraph; shared leader indexes are useful for baseline navigation but are not branch-accurate for worktree-only changes.
If you want a one-off launch with no OMX tmux/HUD management, use --direct:
omx --direct --yolo
For a persistent shell/profile preference, set an environment policy:
OMX_LAUNCH_POLICY=direct omx --yolo
Return to the auto/default behavior with:
unset OMX_LAUNCH_POLICY
CLI policy flags win over the environment, and the last CLI policy flag before
-- wins:
OMX_LAUNCH_POLICY=direct omx --tmux --yolo
Use OMX_LAUNCH_POLICY=direct|tmux|detached-tmux|auto. This iteration only
adds CLI and environment controls; it intentionally does not add a config-file
setting. If you run --direct from inside an existing tmux pane, OMX will not
create HUD splits, enable mouse mode, or wrap extended-key handling, but the
process still runs inside that already-open terminal pane.
Then try the canonical workflow:
# Copy/pasteable durable-goal example:
/goal Ship the checkout bug fix with a durable objective, checkpoints for reproduction, implementation, regression tests, and final verification.
$plan "approve the checkout bug-fix plan and review tradeoffs"
$ultragoal "execute the approved checkout fix with checkpoint evidence"
Use $team when an active Ultragoal story needs coordinated parallel work.
/goal and skill selection
Start a normal strong session with omx --madmax --xhigh (or add --worktree=<task> in a git repo). Inside that session, execute directly or use $ultragoal for durable multi-goal runs, $team for coordinated parallel work, or $plan when the task needs explicit planning first. Use /goal when the task itself needs a durable objective/checkpoint structure that Codex should keep reconciling across turns.
Add only 2-5 relevant skills by default. More skills are allowed when the task scope justifies them, but loading a large catalog is usually a context-budget and attention-quality problem, not a hard parser/runtime blocker. Treat it as a concrete runtime blocker only when a command actually errors.
Anti-pattern:
omx --madmax --xhigh
# Then immediately load 20 skills "just in case" before stating the task.
# This bloats session context and makes the model spend attention on irrelevant workflows.
A simple mental model
OMX does not replace Codex.
It adds a better working layer around it:
- Codex does the actual agent work
- OMX role keywords make useful roles reusable
- OMX skills make common workflows reusable
.omx/ stores plans, logs, memory, and runtime state
Most users should think of OMX as better task routing + better workflow + better runtime, not as a command surface to operate manually all day.
Start here if you are new
- If Codex CLI already exists, verify it with
codex --version and install or update OMX with npm install -g oh-my-codex; otherwise install @openai/codex separately first if you want npm to manage Codex
- After install or real OMX version bumps, run
omx setup --scope project --merge-agents from the target git project or omx setup --scope user for user-level Codex setup, or use omx update when you also want npm to check for and install the latest build before refreshing setup
- Run
omx doctor
- Run a real execution smoke test:
codex login status and omx exec --skip-git-repo-check -C . "Reply with exactly OMX-EXEC-OK"
- Launch with a named worktree from a git repo, for example
omx --worktree=feat/task --madmax --xhigh; if you run concurrent --madmax sessions, use distinct named worktrees such as --worktree=feature/auth
- Use
$deep-interview "..." when the request or boundaries are still unclear
- Use
$ralplan "..." to approve the plan and review tradeoffs
- Use
$ultragoal or $team when the task needs durable or parallel execution; add /goal when durable objective/checkpoint structure should be explicit
Recommended workflow
$autopilot is the first-class canonical orchestrator for the staged workflow $deep-interview -> $ralplan -> $ultragoal. The chain is its defining default, while each stage remains independently invocable when earlier input contracts are already satisfied.
$deep-interview — iterative Socratic ambiguity clearance, resumable state, and execution-ready requirements artifacts.$ralplan — architecture, feasibility, and consensus planning over the deep-interview artifact.$ultragoal — durable multi-goal execution with .omx/ultragoal ledger checkpoints.$team — coordinated parallel execution when a story benefits from multiple lanes.
$deep-interview is an independent requiremen
OMX（oh-my-codex）是叠加在 OpenAI 的 Codex CLI 之上的工作流层，增加了钩子、智能体团队、HUD、提示词，以及标准工作流命令，如 $plan、$ultragoal、$team、$code-review 和 $ultraqa，同时保留 Codex 本身作为执行引擎。