omc（奥麦克）

omc — Claude-first topology-aware router for oh-my-claudecode

When to use this skill

• The user wants Claude Code–native orchestration through OMC / oh-my-claudecode
• The user needs to choose between the marketplace plugin, the omc shell CLI, or a local checkout / --plugin-dir workflow
• The user asks about /team, /autopilot, /ralph, /ultrawork, omc team, omc ask, omc setup, omc update, hooks, HUD, or OMC state behavior
• The user is hitting duplicate installs, plugin-dir confusion, worktree/state collisions, setup drift, or HUD / rate-limit / resume trouble
• The user really wants a Claude-first runtime owner, not jeo, ralphmode, omx, ohmg, or browser-review skills

Instructions

Step 1: Identify the install topology before the packet

Before suggesting commands, identify which OMC topology the operator is actually using:

1. marketplace-plugin — Claude Code plugin / slash-skill surface installed from the marketplace
2. shell-cli — npm-installed omc command for shell-side setup, tmux workers, or provider asks
3. local-plugin-dir — local checkout / --plugin-dir / OMC_PLUGIN_ROOT workflow
4. mixed-or-unknown — overlapping plugin + CLI + local checkout, duplicate commands, or unclear setup state

If the topology is mixed-or-unknown, prefer recovery and topology clarification before recommending more commands.

Step 2: Pick one request packet

After topology is clear enough, classify the actual job into one packet:

1. install-topology — install, first-run setup, or local-checkout / plugin-dir decisions
2. in-session-runtime — choose the right slash skill inside Claude Code
3. terminal-runtime — choose the right omc... shell command
4. recovery-and-update — fix duplicate installs, setup drift, worktree/state trouble, HUD/hooks, or resume issues
5. boundary-and-route-out — the request really belongs to jeo, ralphmode, plannotator, browser-review skills, or another runtime

Lead with the topology and packet mentally, then give only the commands that fit that combination.

Step 3: Tell the operator which surface they are on

OMC has two real runtime surfaces. Do not flatten them.

A. Claude Code plugin / in-session surface

Use this when the user is already inside Claude Code and wants slash skills, hooks, HUD, or native team behavior.

Examples:

/team 3:executor "fix all TypeScript errors"
/autopilot "build a REST API for tasks"
/ralph "keep going until verified done"
/ultrawork "parallelize this cleanup"
/deep-interview "help me clarify the feature"

This surface owns:

• slash skills
• hooks and HUD behavior
• native team orchestration inside Claude Code
• in-session keywords like autopilot: or ralph:

B. Terminal CLI surface

Use this when the user wants shell-side setup, updates, provider consultation, tmux workers, or runtime management.

Examples:

omc setup
omc update
omc ask codex "review this patch"
omc team 2:codex "review auth flow"
omc team status review-auth-flow

This surface owns:

• omc setup / omc update
• omc ask
• omc team tmux workers
• shell-side runtime management

Important truth:

• /team and omc team are different runtimes
• /autopilot, /ralph, and /ultrawork are in-session skills, not normal omc CLI subcommands
• the npm package name is oh-my-claude-sisyphus, not oh-my-claudecode

Step 4: Use the right topology + packet pair

Packet: install-topology

Use when the user needs installation, setup, or topology cleanup.

marketplace-plugin

Inside Claude Code:

/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode

Then run one setup entrypoint inside Claude Code:

setup omc

or:

/oh-my-claudecode:omc-setup

shell-cli

If they also want omc shell commands:

npm i -g oh-my-claude-sisyphus@latest

local-plugin-dir

Do not pretend the marketplace flow is enough. Point them to the plugin-dir / local-checkout path and the dedicated reference before improvising cleanup:

• omc --plugin-dir <path>...
• OMC_PLUGIN_ROOT
• omc setup --plugin-dir-mode
• references/install-topology-and-recovery.md

mixed-or-unknown

Treat duplicate commands, duplicate slash skills, or unclear state as a recovery problem first. Do not stack more installs on top of an ambiguous topology.

Enable native teams in ~/.claude/settings.json when Claude-native teams are expected:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Packet: in-session-runtime

Use when the user wants the best Claude Code slash skill for the job.

Pick one truthful recommendation plus at most 1 nearby alternative.

Packet: terminal-runtime

Use when the user wants the shell CLI rather than the in-session plugin surface.

If the user says “team mode” but is already inside Claude Code, prefer /team. If the user says “team mode” and clearly wants shell/tmux workers, prefer omc team.

Packet: recovery-and-update

Use when the real job is repairing OMC.

Top recovery patterns:

• Duplicate install / duplicate commands → inspect whether marketplace plugin, npm CLI, and local-checkout/plugin-dir installs are overlapping
• Setup drift after upgrade → rerun the truthful setup/update path for the current topology (setup omc, /oh-my-claudecode:omc-setup, omc update, omc setup)
• Local checkout not behaving like marketplace install → verify --plugin-dir, OMC_PLUGIN_ROOT, and plugin-dir mode instead of reinstalling blindly
• Team mode not working → verify CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
• Worktree/state weirdness → inspect .omc/state/, worktree namespace behavior, and team/session names
• HUD / rate-limit / ephemeral-environment trouble → treat it as environment/runtime recovery, not a mode-selection question

When the fix gets detailed, route to:

• references/install-topology-and-recovery.md
• references/intake-packets-and-route-outs.md
• references/cli-reference.md
• references/hooks-reference.md

Packet: boundary-and-route-out

Do not force OMC when another skill owns the job better.

Route out when:

• long-lived plan / execution ledger / resumable multi-skill loop → jeo
• spec-first persistence method beyond OMC runtime detail → ralph
• approval posture, trusted-folder / bypass policy, permission surface → ralphmode
• plan review / approval gate → plannotator
• fresh-session browser verification → agent-browser
• running authenticated browser reuse → playwriter
• exact rendered-UI critique / annotation handoff → agentation
• Codex-first runtime orchestration → omx
• Gemini / Antigravity portable harness adoption → ohmg

Step 5: Keep the answer truthful about volatility

OMC is a fast-moving upstream project. Prefer:

• topology truth first
• current surface distinctions
• current install/setup commands
• current route-outs
• reference docs for deeper operator detail

Avoid:

• pretending all commands live on one surface
• treating /team and omc team as interchangeable
• claiming there is a general omc autopilot CLI
• stacking installs when duplicate-path symptoms suggest recovery first
• absorbing browser review, approvals, or non-Claude runtime ownership into OMC

Examples

Example 1: Marketplace plugin install

User: "OMC 설치하고 Claude Code 안에서 바로 팀 모드까지 쓰고 싶어"

Response: Use the install-topology packet with marketplace-plugin topology. Give the plugin install path, then setup omc or /oh-my-claudecode:omc-setup, and mention CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 when native teams are expected.

Example 2: Local checkout / plugin-dir

User: "로컬 checkout으로 OMC 수정 중인데 Claude plugin cache 때문에 계속 예전 동작이 남아. plugin-dir 쪽으로 어떻게 잡아야 해?"

Response: Use the install-topology packet with local-plugin-dir topology. Point to --plugin-dir, OMC_PLUGIN_ROOT, and omc setup --plugin-dir-mode, then send the operator to the install-topology reference instead of treating this like a normal marketplace install.

Example 3: Shell-side worker team

User: "shell에서 codex 두 명 붙여서 auth 플로우 리뷰 돌리고 싶어"

Response: Use the terminal-runtime packet. Recommend omc team 2:codex "review auth flow", keep the answer on the CLI runtime, and mention status/shutdown commands only if helpful.

Example 4: Long-loop orchestration boundary

User: "기획부터 실행, 검증, 재개까지 하나의 장기 루프로 굴리고 싶어"

Response: Use the boundary-and-route-out packet. Route to jeo as the long-loop coordination layer, and mention OMC only as the Claude-first runtime that jeo may compose with.

Example 5: Mixed duplicate install symptoms

User: "업데이트 후 slash command가 중복으로 보이고 worktree마다 상태도 이상해"

Response: Use the recovery-and-update packet with mixed-or-unknown topology. Tell them to inspect overlapping plugin/CLI/plugin-dir installs, then rerun the truthful setup/update path and inspect worktree/state behavior before choosing another mode.

Best practices

1. Name the topology before the packet when install or recovery is involved.
2. Prefer one recommended mode plus one nearby alternative, not a giant command dump.
3. Treat /team and omc team as related but different runtimes.
4. Keep the npm/package-name mismatch explicit: oh-my-claude-sisyphus provides the omc CLI.
5. Do not improvise over --plugin-dir or duplicate-install symptoms; route to the topology reference.
6. Route long-loop orchestration, approval posture, browser review, and other runtime ownership outward.
7. Use references for volatile operator detail; keep the front door focused on truthful routing.

References

• references/install-topology-and-recovery.md
• references/intake-packets-and-route-outs.md
• references/modes-reference.md
• references/hooks-reference.md
• references/cli-reference.md
• references/agents-catalog.md