opencode-codex-orch

Multi-agent orchestration plugin for OpenCode. Multiple models, multiple agents, one coordinated team.

Origin

This project is derived from oh-my-openagent (omo) and its slim variant oh-my-opencode-slim (omo-slim). It has been purpose-built for OpenAI Codex: all Anthropic/Claude-specific logic, prompts, and model routing have been stripped out, leaving a lean, GPT/Codex-focused orchestration core.

On top of the slim foundation, this fork adds several practical skills and fork-specific workflows (see Built-in skills and Fork-specific skills below).

What it does

opencode-codex-orch turns a single AI agent session into a coordinated development team. Sisyphus orchestrates, specialized agents handle research, planning, code search, and execution in parallel across multiple model providers.

Key capabilities:

• Multi-agent orchestration with 10 specialized agents (Sisyphus, Atlas, Oracle, Prometheus, Librarian, Explore, Metis, Momus, Multimodal-Looker, Sisyphus-Junior)
• Multi-model routing across GPT, Kimi, Gemini, GLM and more — no Claude/Anthropic dependencies
• ultrawork / ulw prefix for autonomous, execution-biased operation
• Hash-anchored edits (LINE#ID content hashing) for reliable file modifications
• Background agents running 5+ tasks in parallel
• Built-in MCPs (web search, Context7, Grep.app)
• LSP + AST-Grep for IDE-level precision
• Skill system with embedded per-skill MCP servers
• 13 lifecycle hooks for error recovery, model fallback, context injection, etc.

Built-in skills

The plugin ships with these built-in skills:

• playwright / agent-browser — browser automation, page inspection, screenshots, and web testing (provider-dependent)
• dev-browser — persistent browser workflows for navigation, extraction, and app testing
• frontend-ui-ux — broad UI/UX design skill for visually strong frontend work, including restrained, brand-forward landing pages and consumer UI (GPT-optimized prompt)
• git-commit / git-rebase / git-search — git workflow skills covering atomic commits, history rewriting, and code archaeology (GPT-optimized; split from the old monolithic git-master to save ~85% tokens per invocation)
• skill-creator — create and refine reusable SKILL.md-based skills for this plugin

Fork-specific skills

These skills were added specifically for this fork:

• github-triage — unified GitHub issue & PR triage with background-task parallelism, fork-aware review lens
• merge-upstream — guided upstream merge workflow that preserves fork identity, slim architecture, and Codex-first design decisions

Install

Add to your OpenCode config (~/.config/opencode/opencode.json):

{
  "plugin": ["opencode-codex-orch@latest"]
}

For local development:

{
  "plugin": ["file:///absolute/path/to/opencode-codex-orch/dist/index.js"]
}

Interactive setup:

bunx opencode-codex-orch install

Config files

• Project: .opencode/opencode-codex-orch.json
• User: ~/.config/opencode/opencode-codex-orch.json

Usage

opencode
# then type:
ulw fix the failing tests

ulw triggers autonomous execution mode. The agent explores, plans, implements, and verifies on its own.

Documentation

• Overview
• Installation Guide
• Orchestration Guide
• Agent-Model Matching
• Configuration Reference
• Features Reference

Credits

Derived from oh-my-openagent (omo) and oh-my-opencode-slim (omo-slim). Prompt design referenced OpenAI Codex CLI prompt.md. Hash-anchored editing inspired by oh-my-pi.

License

SUL-1.0