oh-my-claudecode

oh-my-claudecode turns Claude Code from a chat into a pipeline: plan, PRD, execute, verify, fix, with tmux workers for Codex, Gemini, Antigravity, Grok or Cursor when you want a second model to check the first. The trait that decides it is the verify stage, which keeps looping until the evidence says done rather than until the model says so.

You need Claude Code as your daily driver and tasks that take hours, not minutes. Pick it if that is you. Pick Ruflo if you want a hundred-agent swarm and a memory, and plain Claude Code if the pipeline sounds like ceremony.

The risk is a dependency on a flag. The canonical team mode requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, an experimental switch in Claude Code that Anthropic can rename or remove in a release note, and the harness has no fallback that keeps the pipeline intact. The package ships on npm as oh-my-claude-sisyphus while the repository is oh-my-claudecode, so the install line and the docs disagree on the name.

Pin Claude Code's version alongside it. What it does right: the docs insist on one primary loop authority per session and warn against stacking modes, which is a rare admission that loops fight.

The design is a routed pipeline. 1. Nineteen specialised agents with tier variants cover architecture, research, design, testing and analysis. 2. Routing sends simple steps to Haiku and complex reasoning to Opus, and the README claims a 30 to 50 percent token saving from this; no task set, baseline or run count is given, so the figure is asserted, not demonstrated. 3. Six hook events, from session start to post tool use, are where the harness intercepts the host.

No benchmark is published. The observation: a saving measured against an unstated baseline is a saving of unstated size.

Yeachan Heo is an individual and the asset is 38,984 stars attached to somebody else's product. Distribution is borrowed twice, once through the Claude Code plugin marketplace and once through the OpenClaw gateway integration that relays stop notifications, and there is no pricing, so the project's value is the author's reputation. Pricing power: none, and a plugin cannot raise a price its host could undercut with a feature.

Moat: a workflow habit. Likely outcome: the author is hired by Anthropic or by a harness company, and the pipeline ideas ship first-party. Position: use it, expect it to be absorbed.

The demo is a pipeline finishing overnight. Procurement: team mode requires tmux, named autopilot profiles require the Linux flock utility, and stop callbacks post to Telegram, Discord or Slack, each engineer choosing their own. Sixty engineers means sixty personal configurations on personal Claude Code subscriptions, optional Codex or Gemini plans on top, no SSO, no central log and no vendor.

It can run headless, so a CI experiment is possible. Onboarding is /omc-setup plus a docs afternoon. Not yet; a shared config committed to the repo and a single notification channel first.

MIT and configured entirely from files I can commit: .claude/omc.jsonc for the project, ~/.config/claude-omc/config.jsonc for me, JSONC so I can leave comments, and a .mcp.json for servers. State lives in .omc/, which is deleted with a linked worktree unless I set OMC_STATE_DIR or drop a .omc-workspace marker for a multi-repo layout; that is documented, which is more than most.

No local-model route; the routing table is Claude, Codex, Gemini, Antigravity, Grok and Cursor, all hosted. Forkable, readable, and the config story is the best of the Claude Code plugins.