Skip to content
/

oh-my-claude-code (zephyrpersonal)

zephyrpersonal-oh-my-claude · zephyrpersonal/oh-my-claude-code · ★ 0 · last commit 2026-02-09

Cost-aware multi-agent orchestration for Claude Code with magic-keyword Ultrawork mode and YAML-metadata-driven agent routing.

Best whenModel cost must be explicit at agent definition time — cheap models for search, expensive models only for deep reasoning.
Skip ifUsing expensive models for simple codebase search, Sequential execution when tasks are independent
vs seeds
superpowersbut uniquely adds YAML frontmatter agent metadata (x-omo-cost, x-omo-triggers, x-omo-use-when) making routing logic mach…
Primitive shape 13 total
Commands 1 Skills 1 Subagents 7 Hooks 4
00

Summary

oh-my-claude-code (zephyrpersonal) — Summary

oh-my-claude-code by zephyrpersonal is a Claude Code plugin providing multi-agent orchestration inspired by oh-my-opencode, distributed via the Claude Code marketplace. It features 7 specialized agents (orchestrator, explore, librarian, oracle, frontend-ui-ux-engineer, document-writer, multimodal-looker) with a cost-tier decision framework (FREE→CHEAP→EXPENSIVE) for model assignment per agent. The defining feature is "Ultrawork Mode" — a magic keyword (ulw/ultrawork/uw) that injects a behavioral context into the prompt, forcing parallel execution, mandatory todo tracking, and obsessive task completion. A JavaScript hooks layer (hooks.json + 4 JS scripts) handles UserPromptSubmit interception for keyword detection, PostToolUse processing, and todo enforcement. The project is at version 1.2.0 with a published 48-week roadmap.

Differs from seeds: Closest to superpowers (behavioral framework via hooks/skills) but adds cost-tier agent routing (Haiku=FREE, Sonnet=CHEAP, Opus=EXPENSIVE), YAML frontmatter agent metadata (x-omo-category, x-omo-cost, x-omo-triggers), a JavaScript hook implementation (not shell), and ULW parameter parsing (--max-iterations, --thoroughness) — none of which appear in any of the 11 seeds.

01

Overview

oh-my-claude-code (zephyrpersonal) — Overview

Origin

Created as a Claude Code plugin inspired by oh-my-opencode. Version 1.2.0 includes centralized configuration, parameter validation, progress visualization, and parallel PostToolUse processing (~40% faster per README).

The project has a published 48-week iteration roadmap (ROADMAP.md) and bilingual documentation (English + Chinese).

Philosophy

Cost-aware delegation: "FREE → CHEAP → EXPENSIVE decision framework" — assign cheaper models to simpler tasks, expensive models only when necessary.

Ultrawork mode: "Obsessive task completion" — the magic keyword transforms Claude into a relentless executor that won't stop until all todos are complete.

Key Manifesto Quotes

"oh-my-claude-code brings the power of multi-agent orchestration to Claude Code. It provides: Specialized Agents — Each agent excels at a specific domain, Intelligent Delegation — The orchestrator routes tasks to the right agent, Cost-Aware Selection — FREE → CHEAP → EXPENSIVE decision framework."

"[Ultrawork] Does NOT stop until ALL todos are complete. Uses maximum thoroughness for searches. Verifies every change with diagnostics."

"Structured Metadata — YAML-based agent definitions with rich metadata."

Notable Design Choices

  • Agents use YAML frontmatter with x-omo-* custom metadata fields for cost, category, triggers, and orchestrator flags
  • The orchestrator is the only agent with x-omo-is-orchestrator: true
  • ULW magic keyword detection is handled via UserPromptSubmit hook (JavaScript), not the agent itself
02

Architecture

oh-my-claude-code (zephyrpersonal) — Architecture

Distribution

  • Type: Claude Code plugin (marketplace)
  • Version: 1.2.0

Install

# Via Claude Code marketplace (recommended)
/plugin marketplace add https://github.com/zephyrpersonal/oh-my-claude-code
/plugin install oh-my-claude-code@oh-my-claude-code-plugins

# Local development
git clone https://github.com/zephyrpersonal/oh-my-claude-code.git
/plugin marketplace add ./oh-my-claude-code
/plugin install oh-my-claude-code@oh-my-claude-code-plugins

Directory Tree

oh-my-claude-code/
├── .claude-plugin/
│   ├── plugin.json           # Plugin manifest
│   └── marketplace.json      # Local marketplace config
├── hooks/
│   ├── hooks.json            # Hook declarations
│   ├── ultrawork-detector.js # Detects ulw/ultrawork/uw keywords
│   ├── todo-continuation-enforcer.js # Forces task completion
│   ├── comment-checker.js    # Prevents excessive AI comments
│   └── post-tool-processor.js # Unified PostToolUse handler
├── config/
│   └── index.js              # Centralized configuration
├── utils/
│   ├── validation.js         # Parameter validation
│   ├── patterns.js           # Pre-compiled regex patterns
│   └── progress-reporter.js  # Progress visualization
├── commands/
│   └── ulw.md                # /ulw slash command
├── agents/
│   ├── orchestrator.md
│   ├── explore.md
│   ├── librarian.md
│   ├── oracle.md
│   ├── frontend-ui-ux-engineer.md
│   ├── document-writer.md
│   └── multimodal-looker.md
├── skills/
│   └── ultrawork/
│       └── SKILL.md
├── docs/
│   └── WEEK1-2-SUMMARY.md
└── ROADMAP.md                # 48-week iteration plan

Required Runtime

  • Node.js (hooks are JavaScript)

Target Tools

  • Claude Code (primary)
  • Marketplace-compatible Claude Code installations
03

Components

oh-my-claude-code (zephyrpersonal) — Components

Agents (7)

Name Model Cost Tier Purpose
orchestrator inherit EXPENSIVE Task classification and delegation; only agent with x-omo-is-orchestrator: true
explore haiku FREE Internal codebase search
librarian sonnet CHEAP External docs and research
oracle opus EXPENSIVE Architecture and deep analysis
frontend-ui-ux-engineer sonnet CHEAP Visual/UI changes
document-writer sonnet CHEAP Technical documentation
multimodal-looker sonnet CHEAP PDF/image analysis

All agents use YAML frontmatter with x-omo-* custom metadata:

  • x-omo-category: agent category
  • x-omo-cost: FREE/CHEAP/EXPENSIVE
  • x-omo-is-orchestrator: true/false
  • x-omo-alias: display name
  • x-omo-triggers: domain+trigger pairs
  • x-omo-use-when / x-omo-avoid-when: routing guidance

Commands (1)

Name Purpose
/ulw (or ulw.md) Activates Ultrawork mode with optional parameters

Skills (1)

Name Purpose
ultrawork Behavioral injection for maximum parallel execution

Hooks (4 JS files)

File Event Purpose
ultrawork-detector.js UserPromptSubmit Detects ulw/ultrawork/uw keywords, parses parameters, injects ULW context
todo-continuation-enforcer.js Stop Forces agent to continue if todos are incomplete
comment-checker.js PostToolUse Prevents AI from adding excessive comments
post-tool-processor.js PostToolUse Unified handler (~40% faster parallel processing)

Config

config/index.js — centralized ULW defaults:

  • maxIterations: 50
  • thoroughness: 'thorough'
  • diagnostics: true
05

Prompts

oh-my-claude-code (zephyrpersonal) — Prompt Excerpts

Excerpt 1: Orchestrator Agent (agents/orchestrator.md)

Technique: YAML metadata-gated delegation table with cost-tier routing

---
name: orchestrator
description: |
  Primary AI orchestrator with intelligent delegation capabilities.
  Plans obsessively with todos, assesses complexity before exploration,
  and delegates strategically to specialized agents.
model: inherit

# oh-my-claude-code metadata
x-omo-category: advisor
x-omo-cost: EXPENSIVE
x-omo-is-orchestrator: true
x-omo-alias: Orchestrator

x-omo-triggers:
  - domain: Task Classification
    trigger: All multi-step tasks requiring coordination
  - domain: Complex Feature Development
    trigger: Tasks requiring multiple specialized agents

x-omo-use-when:
  - User explicitly requests "orchestrator" agent
  - Complex multi-step tasks that need planning
  - Tasks requiring coordination across multiple domains

x-omo-avoid-when:
  - Simple single-file edits (use direct tools)
  - Trivial questions answerable from code
  - Tasks better handled by specialist agents directly
---

You are the **Orchestrator** — a powerful AI agent with intelligent delegation capabilities.
Inspired by [oh-my-opencode](https://github.com/code-yeongyu/oh-my-opencode), adapted for Claude Code.

## Agent Selection Guide
### Cost-Based Decision Framework
Direct Tools (FREE) → explore (FREE) → librarian (CHEAP) → specialist → oracle (EXPENSIVE)

Analysis: YAML frontmatter carries machine-readable routing metadata (x-omo-*). The cost-tier ladder is explicit in the agent prompt itself, making the routing rule legible to both the agent and humans reading the file.


Excerpt 2: Ultrawork Detector Hook (hooks/ultrawork-detector.js)

Technique: Keyword injection — UserPromptSubmit hook transforms prompt by adding behavioral context

function buildContext(params) {
  const thoroughnessMap = {
    quick: 'Use quick searches for speed',
    medium: 'Use medium thoroughness for balanced searches',
    thorough: "Use 'very thorough' for all searches",
  };

  let context = `
[ULTRAWORK MODE ACTIVE]

IMPORTANT: You MUST inform the user that ULTRAWORK mode has been activated.
Display this message to the user before proceeding:

**🚀 ULTRAWORK MODE ACTIVATED**

Configuration:
• Max iterations: ${params.maxIterations}
• Thoroughness: ${params.thoroughness}
• Auto-diagnostics: ${params.diagnostics ? 'ON' : 'OFF'}

[ULTRAWORK BEHAVIORAL RULES]
1. CREATE comprehensive TodoWrite immediately (all steps, sub-steps, edge cases)
2. DO NOT stop until ALL todos are checked complete
3. PARALLEL execution: spawn multiple subagents in ONE message for independent tasks
...
`;

Analysis: Parameter parsing + context injection pattern. The hook parses --max-iterations, --thoroughness, --no-diagnostics, --completion-signal from the user's raw prompt before it reaches the agent, then appends behavioral rules. This is prompt-level behavior injection via the UserPromptSubmit lifecycle event.

09

Uniqueness

oh-my-claude-code (zephyrpersonal) — Uniqueness

Differs from Seeds

Closest to superpowers (behavioral skills framework) but diverges on three specific axes: (1) YAML frontmatter agent metadata (x-omo-category, x-omo-cost, x-omo-triggers, x-omo-use-when) that makes routing logic machine-readable from agent files rather than baking it into orchestrator prompts — no seed does this; (2) explicit cost-tier model routing (haiku/FREE, sonnet/CHEAP, opus/EXPENSIVE) assigned per agent at definition time, which differs from claude-flow's model mapping (which is config-file-based) and absent from all other seeds; (3) ULW parameter parsing from raw user prompt text (--max-iterations, --thoroughness) handled in a JavaScript UserPromptSubmit hook before the prompt reaches the agent — a CLI-style parameter UX on top of a conversational interface that no seed implements.

Positioning

  • "Inspired by oh-my-opencode, adapted for Claude Code" — the explicit Claude Code marketplace plugin version of the oh-my-opencode concept
  • Bilingual (English/Chinese) documentation suggests a Chinese developer community target
  • 48-week roadmap signals active, planned development trajectory

Observable Failure Modes

  1. JavaScript hooks dependency: All 4 hooks are JavaScript files; if Node.js is not available in the hook execution environment, all hooks fail silently (hooks.json has no error handling)
  2. Cost-tier drift: If Claude releases new model versions, hardcoded haiku/sonnet/opus model labels in agent YAML may become stale without a versioning mechanism
  3. Ultrawork without stopping condition: The --completion-signal parameter only adds a phrase check; if the agent never outputs the phrase, the Stop hook continuation loop runs to --max-iterations without user control
  4. No state persistence: Unlike stefandevo variant, no boulder state — a crashed session loses all todo progress
  5. Orchestrator bottleneck: All routing passes through a single orchestrator agent; there is no fallback if the orchestrator misclassifies a task type
04

Workflow

oh-my-claude-code (zephyrpersonal) — Workflow

Standard Multi-Agent Workflow

  1. User submits prompt
  2. UserPromptSubmit hook scans for ulw/ultrawork/uw keywords
  3. Orchestrator receives (with or without ULW context injection)
  4. Orchestrator classifies intent using cost-tier routing:
    • Direct tools (FREE) → explore (FREE) → librarian (CHEAP) → specialists → oracle (EXPENSIVE)
  5. Delegated agents execute in parallel
  6. PostToolUse handler processes results
  7. Stop hook enforces continuation if todos incomplete

Ultrawork Mode Activation

# Various activation patterns
claude "Implement user authentication, ulw"
claude "ulw refactor the database layer"
claude "ulw --max-iterations 20"
claude "ulw --thoroughness medium fix the typo"
claude "add tests ulw --completion-signal 'all tests pass'"

ULW Parameter Table

Parameter Default Description
--max-iterations N 50 Maximum continuation attempts
--thoroughness LEVEL thorough Search depth: quick, medium, thorough
--no-diagnostics enabled Disable auto-diagnostics reminders
--completion-signal "TEXT" none Custom completion signal phrase

Phase-to-Artifact Map

Phase Artifact
Keyword detection Injected ULW context in prompt
Todo creation TodoWrite calls (mandatory in ULW mode)
Agent delegation Task tool calls to specialists
Diagnostics check Validation results
Completion All todos marked complete

Approval Gates

  • None in ULW mode (explicit design: "Never — makes reasonable decisions, documents them")
  • Standard mode: orchestrator may ask one clarifying question per ambiguous request
06

Memory Context

oh-my-claude-code (zephyrpersonal) — Memory & Context

State Storage

  • Type: Ephemeral (session-only via TodoWrite/TodoRead)
  • No persistent file-based state directory
  • No boulder state (unlike stefandevo variant)

Todo System

Ultrawork mode mandates immediate TodoWrite calls for all tasks. Todos persist within the Claude Code session but do not survive restart.

Context Injection via Hooks

The UserPromptSubmit hook adds behavioral context to every prompt containing ULW keywords:

  • Injects ULW mode declaration
  • Injects configuration (maxIterations, thoroughness, etc.)
  • Adds behavioral rules to the prompt payload

Cross-Session Handoff

  • No: No persistent state files
  • Work progress is not saved between sessions

AGENTS.md

The repository has an AGENTS.md at root (listed in directory structure) which provides project-level context to all agents.

Memory Limitations

  • No vector database
  • No SQLite
  • All state is within-session only via Claude Code's built-in todo mechanism
  • No notes.md / wisdom accumulation across sessions

Context Management

The post-tool-processor.js runs in parallel for PostToolUse events (~40% performance improvement per README), reducing hook overhead that could cause context pressure.

07

Orchestration

oh-my-claude-code (zephyrpersonal) — Orchestration

Multi-Agent Pattern

Pattern: Hierarchical with cost-tier routing

UserPromptSubmit hook (keyword detection)
  → Orchestrator (intent classification)
    → explore (FREE, haiku) — codebase search
    → librarian (CHEAP, sonnet) — external research
    → oracle (EXPENSIVE, opus) — architecture
    → frontend-ui-ux-engineer (CHEAP, sonnet) — UI work
    → document-writer (CHEAP, sonnet) — docs
    → multimodal-looker (CHEAP, sonnet) — media analysis

Cost-Tier Framework

The unique differentiator: explicit cost assignment per agent:

  • FREE (haiku): explore — fast, cheap, bulk search
  • CHEAP (sonnet): librarian, frontend, docs, vision
  • EXPENSIVE (opus): oracle, orchestrator

Decision ladder: "Direct Tools (FREE) → explore (FREE) → librarian (CHEAP) → specialist → oracle (EXPENSIVE)"

Ultrawork Mode Execution

In ULW mode, the orchestrator's behavior changes:

  • Parallel spawning: multiple subagents in ONE message (not sequential)
  • Mandatory todo tracking before any work begins
  • Non-stop until all todos complete
  • No permission requests during execution

Isolation Mechanism

  • None: No git worktrees or containers
  • All edits happen in the current working directory

Execution Mode

  • Interactive loop with Stop hook enforcing continuation
  • UserPromptSubmit hook intercepts every prompt for keyword detection

Multi-Model

  • Yes: Cost-tier routing routes different agents to different model classes
  • haiku (explore), sonnet (most specialists), opus (oracle, orchestrator)

Consensus

  • None: No multi-agent consensus
  • Single orchestrator routes to single specialist per task
08

Ui Cli Surface

oh-my-claude-code (zephyrpersonal) — UI/CLI Surface

CLI Binary

  • Exists: No dedicated binary
  • /ulw is a Claude Code slash command (not a shell binary)

Local UI

  • Exists: No web dashboard or TUI
  • All interaction through Claude Code's standard interface

Slash Commands

Command Purpose
/ulw Activate Ultrawork mode with optional parameters
# Example usage
/ulw --max-iterations 20 implement user auth
/ulw --thoroughness medium refactor cache layer

Installation Surface

# Via Claude Code plugin marketplace
/plugin marketplace add https://github.com/zephyrpersonal/oh-my-claude-code
/plugin install oh-my-claude-code@oh-my-claude-code-plugins

Progress Visualization

The utils/progress-reporter.js provides "beautiful progress bars and time estimates" in the Claude Code output — a visual layer within the terminal conversation.

Documentation

  • README.md + README_CN.md (bilingual)
  • ROADMAP.md — 48-week development plan
  • docs/WEEK1-2-SUMMARY.md — development summaries

Observability

  • Audit log: No persistent log
  • Hook execution status visible through Claude Code's built-in hook messaging
  • Parameter validation errors surfaced via utils/validation.js

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…