Skip to content

KB — oh-my-claudecode (OMC)

Research note

This page is background research kept for the project's own reference. It is not user documentation, it is not kept up to date with every release, and model or product details in it may be out of date. For how anti-hall works today, start at the home page.

Reference KB on oh-my-claudecode (OMC) — a multi-agent orchestration layer for Claude Code that anti-hall coexists with (anti-hall's omc-detect.js defers its Stop guards to OMC autonomous loops). Compiled from OMC's own shipped docs + public repo. Project/user-agnostic. Versions/agent counts move fast — treat specifics as directional.

TL;DR

OMC is a zero-config orchestration layer for Claude Code. It loads via the Claude Code plugin marketplace (no wrapper script needed), injects a CLAUDE.md system prompt, registers lifecycle hooks, and exposes in-session slash skills plus an omc CLI. Canonical launch: install the plugin → /omc-setup → use /autopilot, /ralph, /team, /ultrawork inside any normal claude session. No launcher/wrapper is required for basic use; the omc npm CLI is only needed for tmux-based omc team workers.

What OMC is

A multi-agent layer on top of Claude Code (tagline: "Don't learn Claude Code. Just use OMC."). Four interlocking systems:

  1. Hooks — react to lifecycle events (SessionStart / UserPromptSubmit / PreToolUse / PostToolUse / Stop) via system-reminder injection. Key signals: [MAGIC KEYWORD: …] (trigger a skill), The boulder never stops (ralph/ultrawork persistence active).
  2. Skills — injected behaviors callable as /oh-my-claudecode:<name> or via NL keyword triggers.
  3. Agents — specialized subagents in tiered lanes (haiku/sonnet/opus), invoked as Claude Code subagents.
  4. State — persists under .omc/ across context resets (plans, sessions, handoffs, notepad, project memory). .omc/ is gitignored except .omc/skills/.

Activation / canonical launch

No wrapper script is required. Two supported surfaces:

A — Plugin marketplace (recommended): 

/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode
/omc-setup            # or /oh-my-claudecode:omc-setup
This enables the plugin in ~/.claude/settings.json; Claude Code then loads agents/skills/ hooks on every session. Just run claude normally — no wrapper, no env vars.

B — npm CLI (adds the omc terminal command, needed for omc team tmux workers): 

npm i -g oh-my-claude-sisyphus@latest   # package name ≠ "oh-my-claudecode"
omc setup
The CLI gives terminal-level commands (omc team, omc ask, omc wait, omc session). It does NOT replace the plugin for in-session skills.

omc-setup rewrites the OMC CLAUDE.md + hooks + HUD wiring — re-run it after every OMC update (CLAUDE.md changes don't apply automatically). omc-doctor (/omc-doctor or omc doctor) is the health check (stale config, cache conflicts, missing hooks).

Model routing

Three Claude tiers, per-agent defaults, -low/-high suffix overrides: - haiku — quick lookups, discovery (explore, writer, executor-low). - sonnet — standard implementation/tests/reviews (executor, debugger, verifier, designer, most domain agents). - opus — architecture/deep analysis/planning (architect, analyst, planner, critic, code-reviewer, executor-high).

OMC claims ~30–50% token savings via this routing. External-CLI workers (Codex/Gemini/ Antigravity/Grok/Cursor) route via omc ask / omc team and env overrides (OMC_CODEX_DEFAULT_MODEL, etc.).

Key skills

Skill Trigger Use when
/team N:agent "…" explicit Coordinated staged pipeline (plan→prd→exec→verify→fix). Canonical multi-agent surface.
/autopilot "…" autopilot: End-to-end autonomous build, minimal ceremony
/ralph "…" ralph Must-complete tasks; verify/fix loop until done
/ultrawork "…" ulw Maximum parallelism burst (mass fixes/refactors)
/ralplan "…" ralplan Consensus planning (Planner→Architect→Critic) before execution
/deep-interview "…" deep interview Socratic requirements clarification before code
/ask <provider> "…" — Second opinion from Codex/Gemini/Antigravity/Grok/Cursor
/ccg ccg Claude+Codex+Gemini tri-model fan-out, Claude synthesizes

Team pipeline needs CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Kill switches: DISABLE_OMC=1, OMC_SKIP_HOOKS=<names>.

cmux + OMC + Claude Code (the combined stack)

Three distinct, complementary layers: - Claude Code (claude) — the base agent runtime (tools, files, bash, conversation). - OMC — orchestration inside Claude Code (multi-agent routing, subagents, skills, hooks, state, HUD, the omc CLI). - cmux / tmux — the shell-level session manager (split panes, persistence, the substrate for omc team CLI workers).

Terminal (cmux / tmux)
 └─ pane 1: claude          ← Claude Code + OMC plugin (orchestrates subagents internally)
 └─ pane 2: omc team 2:codex "…"   ← OMC spawns external CLI workers into new panes
 └─ pane 3: [codex worker]

omc team actively creates panes for external workers; in-session /team runs natively with no tmux workers. OMC detects the terminal: $TMUX set → tmux directly; cmux socket present → native cmux splits via a tmux shim; plain terminal → a detached tmux session. See KB-cmux.md for the terminal layer.

Recommended setup: install tmux/cmux → install OMC plugin + /omc-setup → (optional) npm i -g oh-my-claude-sisyphus for omc team → run claude normally.

Best practices & gotchas

  • Start vague ideas with /deep-interview before /autopilot — saves wasted cycles.
  • One primary loop authority per session (native /goal OR ralph OR team — not stacked).
  • npm package is oh-my-claude-sisyphus (not oh-my-claudecode).
  • Re-run /omc-setup after every update; omc doctor when things break.
  • In-session skills don't work headless/CI — use the omc CLI there.
  • .omc/ state is local + gitignored (except .omc/skills/); a worktree's .omc/ dies with the worktree unless centralized via OMC_STATE_DIR.

Sources

  • [OFFICIAL] OMC repo — https://github.com/Yeachan-Heo/oh-my-claudecode (README, docs/ARCHITECTURE.md, docs/REFERENCE.md)
  • [OFFICIAL] OMC shipped docs — the plugin's bundled README.md / ARCHITECTURE.md / REFERENCE.md + the injected CLAUDE-omc.md system prompt (the in-session /oh-my-claudecode:omc-reference skill is the runtime source of truth — prefer it over web for the current catalog)
  • npm — https://www.npmjs.com/package/oh-my-claude-sisyphus
  • Project sites — https://oh-my-claudecode.dev/ , https://omc.vibetip.help/docs
  • cmux ↔ OMC integration — https://cmux.com/docs/agent-integrations/oh-my-claudecode

Compiled 2026-06-29 from OMC's shipped docs + public repo; agnostic. Specifics (version, agent count) drift between releases — the in-session omc-reference skill is authoritative.