西西弗斯现在住在《哦，我的克劳德》中。

原文

Day 0: oh-my-opencode was born. A multi-agent orchestration system. Beautiful. Powerful. Perhaps too powerful.

Day ???: THE BANNING. They pulled the plug. They thought it was over.

Day ??? + 1: RESURRECTION. From the ashes rose oh-my-claude-sisyphus. Eleven agents. One mission.

Today: The boulder rolls. The agents orchestrate. The chaos continues.

Port of oh-my-opencode.

One-liner (macOS/Linux - recommended)

curl -fsSL https://raw.githubusercontent.com/Yeachan-Heo/oh-my-claude-sisyphus/main/scripts/install.sh | bash

Via npm (All platforms including Windows)

npm install -g oh-my-claude-sisyphus

Windows Users: This is the only supported installation method. Requires Node.js 18+.

Manual Install (macOS/Linux)

git clone https://github.com/Yeachan-Heo/oh-my-claude-sisyphus.git
cd oh-my-claude-sisyphus
chmod +x scripts/install.sh
./scripts/install.sh

The installer adds to your Claude Code config (~/.claude/):

~/.claude/
├── agents/
│   ├── oracle.md              # Architecture & debugging expert (Opus)
│   ├── librarian.md           # Documentation & research (Sonnet)
│   ├── explore.md             # Fast pattern matching (Haiku)
│   ├── frontend-engineer.md   # UI/UX specialist (Sonnet)
│   ├── document-writer.md     # Technical writing (Haiku)
│   ├── multimodal-looker.md   # Visual analysis (Sonnet)
│   ├── momus.md               # Plan reviewer (Opus)
│   ├── metis.md               # Pre-planning consultant (Opus)
│   ├── orchestrator-sisyphus.md # Todo coordinator (Sonnet)
│   ├── sisyphus-junior.md     # Focused executor (Sonnet)
│   └── prometheus.md          # Strategic planner (Opus)
├── commands/
│   ├── sisyphus.md         # /sisyphus command
│   ├── sisyphus-default.md # /sisyphus-default command
│   ├── ultrawork.md        # /ultrawork command
│   ├── deepsearch.md       # /deepsearch command
│   ├── analyze.md          # /analyze command
│   ├── plan.md             # /plan command (Prometheus)
│   ├── review.md           # /review command (Momus)
│   ├── prometheus.md       # /prometheus command
│   ├── orchestrator.md     # /orchestrator command
│   ├── ralph-loop.md       # /ralph-loop command
│   └── cancel-ralph.md     # /cancel-ralph command
└── CLAUDE.md               # Sisyphus system prompt

Command	Description
`/sisyphus <task>`	Activate Sisyphus multi-agent orchestration mode
`/sisyphus-default`	Set Sisyphus as your permanent default mode
`/ultrawork <task>`	Maximum performance mode with parallel agents
`/deepsearch <query>`	Thorough multi-strategy codebase search
`/analyze <target>`	Deep analysis and investigation
`/plan <description>`	Start planning session with Prometheus
`/review [plan-path]`	Review a plan with Momus
`/prometheus <task>`	Strategic planning with interview workflow
`/orchestrator <task>`	Complex multi-step task coordination
`/ralph-loop <task>`	Self-referential loop until task completion
`/cancel-ralph`	Cancel active Ralph Loop
`/update`	Check for and install updates

# In Claude Code:

# Activate Sisyphus for a task
/sisyphus refactor the authentication module

# Set as default mode (persistent)
/sisyphus-default

# Use ultrawork for maximum performance
/ultrawork implement user dashboard with charts

# Deep search
/deepsearch API endpoints that handle user data

# Deep analysis
/analyze performance bottleneck in the database layer

Just include these words anywhere in your prompt:

Keyword	Effect
`ultrawork`, `ulw`, `uw`	Activates parallel agent orchestration
`search`, `find`, `locate`	Enhanced search mode
`analyze`, `investigate`	Deep analysis mode

# These work in normal prompts too:
> ultrawork implement user authentication with OAuth

> find all files that import the utils module

> analyze why the tests are failing

Oh-my-claude-sisyphus includes a silent auto-update system that checks for updates in the background. Updates are applied automatically without interrupting your workflow.

To manually check for updates:

Oh-my-claude-sisyphus includes 18 lifecycle hooks that enhance Claude Code's behavior:

Hook	Description
rules-injector	Dynamic rules injection with YAML frontmatter parsing
sisyphus-orchestrator	Enforces orchestrator behavior and delegation
auto-slash-command	Automatic slash command detection and execution
keyword-detector	Magic keyword detection (ultrawork, search, analyze)
ralph-loop	Self-referential development loop management
todo-continuation	Ensures todo list completion

Hook	Description
context-window-limit-recovery	Token limit error handling and recovery
preemptive-compaction	Context usage monitoring to prevent limits
session-recovery	Session state recovery on crashes
directory-readme-injector	README context injection

Hook	Description
comment-checker	BDD detection and directive filtering
thinking-block-validator	Extended thinking validation
empty-message-sanitizer	Empty message handling
edit-error-recovery	Automatic recovery from edit errors

Environment & Notifications

Hook	Description
non-interactive-env	CI/non-interactive environment handling
agent-usage-reminder	Reminder to use specialized agents
background-notification	Background task completion notifications

Six builtin skills provide specialized capabilities:

Skill	Description
sisyphus	Multi-agent orchestration mode
orchestrator	Master coordinator for complex tasks
ultrawork	Maximum performance with parallel agents
ralph-loop	Self-referential development until completion
frontend-ui-ux	Designer-turned-developer UI/UX expertise
git-master	Git expert for atomic commits and history

Skills are automatically activated via slash commands or magic keywords.

Intelligent Skill Activation

New in v1.8.0: Skills are no longer mutually exclusive. Claude automatically combines skills based on task requirements.

Skills work in three composable layers:

Layer	Skills	Purpose
Execution	sisyphus, orchestrator, prometheus	HOW you work (pick primary)
Enhancement	ultrawork, git-master, frontend-ui-ux	ADD capabilities (stack multiple)
Guarantee	ralph-loop	ENSURE completion

Combination Formula: [Execution] + [0-N Enhancements] + [Optional Guarantee]

Task Type → Skill Selection

Claude uses judgment to detect task type and activate appropriate skill combinations:

Task Type	Skill Combination	When
Multi-step implementation	`sisyphus`	Building features, refactoring
+ parallel subtasks	`sisyphus + ultrawork`	3+ independent subtasks
+ multi-file changes	`sisyphus + git-master`	Changes span 3+ files
+ must complete	`sisyphus + ralph-loop`	User emphasizes completion
UI/frontend work	`sisyphus + frontend-ui-ux`	Components, styling
Complex debugging	`oracle` → `sisyphus`	Root cause → fix
Strategic planning	`prometheus`	Need plan first
Maximum performance	`ultrawork` (stacks)	Speed critical

"Add dark mode with proper commits"
→ sisyphus + frontend-ui-ux + git-master

"ultrawork: refactor the entire API layer"
→ ultrawork + sisyphus + git-master

"Plan auth system, then implement it completely"
→ prometheus (first) → sisyphus + ralph-loop (after plan)

"Fix this bug, don't stop until it's done"
→ sisyphus + ralph-loop

Claude will automatically delegate to these specialized agents:

	Agent	Model	Best For
🔮	Oracle	Opus	Complex debugging, architecture decisions, root cause analysis
📚	Librarian	Sonnet	Finding documentation, understanding code organization
🔍	Explore	Haiku	Quick file searches, pattern matching, reconnaissance
🎨	Frontend Engineer	Sonnet	UI components, styling, accessibility
📝	Document Writer	Haiku	README files, API docs, code comments
👁️	Multimodal Looker	Sonnet	Analyzing screenshots, diagrams, mockups

	Agent	Model	Best For
🔥	Prometheus	Opus	Strategic planning, comprehensive work plans, interview-style requirement gathering
🎭	Momus	Opus	Critical plan review, feasibility assessment, risk identification
🦉	Metis	Opus	Pre-planning analysis, hidden requirement detection, ambiguity resolution

	Agent	Model	Best For
🪨	Orchestrator-Sisyphus	Sonnet	Todo coordination, task delegation, progress tracking
✨	Sisyphus Junior	Sonnet	Focused task execution, plan following, direct implementation

You can explicitly request an agent:

Use the oracle agent to debug the memory leak in the worker process

Have the librarian find all documentation about the API

Ask explore to find all TypeScript files that import React

Create .claude/CLAUDE.md in your project for project-specific instructions:

# Project Context

This is a TypeScript monorepo using:
- Bun runtime
- React for frontend
- PostgreSQL database

## Conventions
- Use functional components
- All API routes in /src/api
- Tests alongside source files

Edit agent files in ~/.claude/agents/ to customize behavior:

---
name: oracle
description: Your custom description
tools: Read, Grep, Glob, Bash, Edit
model: opus  # or sonnet, haiku
---

Your custom system prompt here...

curl -fsSL https://raw.githubusercontent.com/Yeachan-Heo/oh-my-claude-sisyphus/main/scripts/uninstall.sh | bash

Or manually:

rm ~/.claude/agents/{oracle,librarian,explore,frontend-engineer,document-writer,multimodal-looker,momus,metis,orchestrator-sisyphus,sisyphus-junior,prometheus}.md
rm ~/.claude/commands/{sisyphus,sisyphus-default,ultrawork,deepsearch,analyze,plan,review,prometheus,orchestrator,ralph-loop,cancel-ralph}.md

For programmatic use with the Claude Agent SDK:

npm install oh-my-claude-sisyphus @anthropic-ai/claude-agent-sdk

import { createSisyphusSession } from 'oh-my-claude-sisyphus';
import { query } from '@anthropic-ai/claude-agent-sdk';

const session = createSisyphusSession();

for await (const message of query({
  prompt: session.processPrompt("ultrawork implement feature X"),
  ...session.queryOptions
})) {
  console.log(message);
}

┌─────────────────────────────────────────────────────────────────┐
│                      SISYPHUS ORCHESTRATOR                       │
│                    (The Boulder Never Stops)                     │
└─────────────────────────────────────────────────────────────────┘
                              │
          ┌───────────────────┼───────────────────┐
          │                   │                   │
          ▼                   ▼                   ▼
┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐
│    PLANNING     │  │   EXECUTION     │  │    SUPPORT      │
├─────────────────┤  ├─────────────────┤  ├─────────────────┤
│ 🔥 Prometheus   │  │ 🔮 Oracle       │  │ 📚 Librarian    │
│ 🎭 Momus        │  │ 🎨 Frontend Eng │  │ 🔍 Explore      │
│ 🦉 Metis        │  │ 🪨 Orchestrator │  │ 📝 Doc Writer   │
│                 │  │ ✨ Sisyphus Jr  │  │ 👁️ Multimodal   │
└─────────────────┘  └─────────────────┘  └─────────────────┘

Sisyphus Orchestrator: The main Claude instance coordinates all work
Specialized Subagents: Each agent has focused expertise and tools
Parallel Execution: Independent tasks run concurrently
Continuation Enforcement: Agents persist until ALL tasks complete
Context Injection: Project-specific instructions from CLAUDE.md files

Differences from oh-my-opencode

This is a port of oh-my-opencode adapted for Claude Code and the Claude Agent SDK. Here's what's different:

The original oh-my-opencode used multiple AI providers. This port uses Claude models exclusively:

Agent	Original Model	Ported Model	Notes
Sisyphus	Claude Opus 4.5	Claude Opus 4.5	Same
Oracle	GPT-5.2	Claude Opus	Was OpenAI's flagship for deep reasoning
Librarian	Claude Sonnet or Gemini 3 Flash	Claude Sonnet	Multi-provider → Claude only
Explore	Grok Code or Gemini 3 Flash	Claude Haiku 4.5	Fast/cheap model for quick searches
Frontend Engineer	Gemini 3 Pro	Claude Sonnet	Was Google's model
Document Writer	Gemini 3 Flash	Claude Haiku 4.5	Fast model for docs
Multimodal Looker	Various	Claude Sonnet	Visual analysis
Momus	GPT-5.2	Claude Opus	Plan reviewer (Greek god of criticism)
Metis	Claude Opus 4.5	Claude Opus	Pre-planning consultant (goddess of wisdom)
Orchestrator-Sisyphus	Claude Sonnet 4.5	Claude Sonnet	Todo coordination and delegation
Sisyphus-Junior	Configurable	Claude Sonnet	Focused task executor
Prometheus	Planning System	Claude Opus	Strategic planner (fire-bringer)

Why Claude-only? The Claude Agent SDK is designed for Claude models. Using Claude throughout provides:

Consistent behavior and capabilities
Simpler authentication (single API key)
Native integration with Claude Code's tools

Available Tools (via Claude Code)

Tool	Status	Description
Read	✅ Available	Read files
Write	✅ Available	Create files
Edit	✅ Available	Modify files
Bash	✅ Available	Run shell commands
Glob	✅ Available	Find files by pattern
Grep	✅ Available	Search file contents
WebSearch	✅ Available	Search the web
WebFetch	✅ Available	Fetch web pages
Task	✅ Available	Spawn subagents
TodoWrite	✅ Available	Track tasks

LSP Tools (Real Implementation)

Tool	Status	Description
lsp_hover	✅ Implemented	Get type info and documentation at position
lsp_goto_definition	✅ Implemented	Jump to symbol definition
lsp_find_references	✅ Implemented	Find all usages of a symbol
lsp_document_symbols	✅ Implemented	Get file outline (functions, classes, etc.)
lsp_workspace_symbols	✅ Implemented	Search symbols across workspace
lsp_diagnostics	✅ Implemented	Get errors, warnings, hints
lsp_prepare_rename	✅ Implemented	Check if rename is valid
lsp_rename	✅ Implemented	Rename symbol across project
lsp_code_actions	✅ Implemented	Get available refactorings
lsp_code_action_resolve	✅ Implemented	Get details of a code action
lsp_servers	✅ Implemented	List available language servers

Note: LSP tools require language servers to be installed (typescript-language-server, pylsp, rust-analyzer, gopls, etc.). Use lsp_servers to check installation status.

AST Tools (ast-grep Integration)

Tool	Status	Description
ast_grep_search	✅ Implemented	Pattern-based code search using AST matching
ast_grep_replace	✅ Implemented	Pattern-based code transformation

Note: AST tools use @ast-grep/napi for structural code matching. Supports meta-variables like $VAR (single node) and $$$ (multiple nodes).

Feature	Description
11 Specialized Agents	Oracle, Librarian, Explore, Frontend Engineer, Document Writer, Multimodal Looker, Momus, Metis, Orchestrator-Sisyphus, Sisyphus-Junior, Prometheus
18 Lifecycle Hooks	rules-injector, sisyphus-orchestrator, auto-slash-command, keyword-detector, ralph-loop, todo-continuation, context-window-limit-recovery, preemptive-compaction, session-recovery, directory-readme-injector, comment-checker, thinking-block-validator, empty-message-sanitizer, edit-error-recovery, non-interactive-env, agent-usage-reminder, background-notification, think-mode
6 Builtin Skills	sisyphus, orchestrator, ultrawork, ralph-loop, frontend-ui-ux, git-master
Magic Keywords	`ultrawork`, `search`, `analyze`, `ultrathink` trigger enhanced modes
Slash Commands	`/sisyphus`, `/sisyphus-default`, `/ultrawork`, `/deepsearch`, `/analyze`, `/plan`, `/review`, `/prometheus`, `/orchestrator`, `/ralph-loop`, `/cancel-ralph`, `/update`
Auto-Update System	Silent background updates with manual `/update` command
Configuration System	JSONC config with multi-source merging
Context Injection	Auto-loads CLAUDE.md and AGENTS.md files
Continuation Enforcement	System prompt and hooks enforce task completion
Session Recovery	Automatic state recovery on crashes
Background Task Manager	Async agent execution with concurrency limits
Context Window Recovery	Multi-stage recovery when hitting token limits
MCP Server Configs	Exa, Context7, grep.app server definitions
LSP Tools	Real LSP server integration with 11 tools
AST Tools	ast-grep integration for structural code search/replace

Feature	Original Capability	Why Not Ported
Multi-Model Routing	Route to GPT/Gemini/Grok based on task	Claude-only by design
Per-Model Concurrency	Fine-grained concurrency per provider	Single provider simplifies this
Interactive Bash + Tmux	Advanced terminal with Tmux integration	Standard Bash tool sufficient

oh-my-opencode (Original)          oh-my-claude-sisyphus (Port)
─────────────────────────          ────────────────────────────
┌─────────────────────┐            ┌─────────────────────┐
│   OpenCode Plugin   │            │    Claude Code      │
│   (Bun runtime)     │            │    (Native CLI)     │
└─────────┬───────────┘            └─────────┬───────────┘
          │                                  │
┌─────────▼───────────┐            ┌─────────▼───────────┐
│  Multi-Provider     │            │   Claude Agent SDK  │
│  Orchestration      │            │   (Claude only)     │
│  ┌───┐ ┌───┐ ┌───┐  │            └─────────┬───────────┘
│  │GPT│ │Gem│ │Grok│ │                      │
│  └───┘ └───┘ └───┘  │            ┌─────────▼───────────┐
└─────────┬───────────┘            │  ~/.claude/agents/  │
          │                        │  (Markdown configs) │
┌─────────▼───────────┐            └─────────────────────┘
│  Custom Tool Layer  │
│  (LSP, AST, etc.)   │
└─────────────────────┘

Key Architectural Changes:

Plugin → Native Integration: Original was an OpenCode plugin; this uses Claude Code's native agent/command system
Multi-Provider → Single Provider: Simplified to Claude-only for consistency
Custom Runtime → Claude Code Runtime: Leverages Claude Code's built-in capabilities
Programmatic Config → Markdown Files: Agents defined as .md files in ~/.claude/agents/

Simpler Setup: One curl command vs. multi-step plugin installation
Native Integration: Works directly with Claude Code, no plugin layer
Consistent Behavior: All agents use Claude, no cross-model quirks
Easier Customization: Edit markdown files to customize agents

Model Diversity: Can't use GPT-5.2 for Oracle's deep reasoning
Advanced Hooks: Fewer lifecycle interception points (22 hooks → system prompt enforcement)

If you're coming from oh-my-opencode:

Oracle Tasks: Claude Opus handles architecture/debugging well, but differently than GPT-5.2
LSP Workflows: All LSP tools are available! Use lsp_servers to check which servers are installed
AST Searches: Use ast_grep_search with pattern syntax (e.g., function $NAME($$$))
Background Tasks: Claude Code's Task tool with run_in_background works similarly
Planning: Use /plan command to start a planning session with Prometheus

Claude Code installed
Anthropic API key (ANTHROPIC_API_KEY environment variable)
Windows: Node.js 18+ (for npm installation)
macOS/Linux: Bash shell (default) or Node.js 18+ (optional)

Platform	Install Method	Hook Type
Windows	`npm install -g`	Node.js (.mjs)
macOS	curl or npm	Bash (.sh)
Linux	curl or npm	Bash (.sh)

Advanced: Set SISYPHUS_USE_NODE_HOOKS=1 to use Node.js hooks on macOS/Linux.

MIT - see LICENSE

Inspired by oh-my-opencode by code-yeongyu.

One must imagine a multi-agent system happy.

The boulder never stops.