This page introduces oh-my-openagent (formerly oh-my-opencode) as an OpenCode plugin and explains its architecture, initialization flow, and core components at a high level. Oh My OpenAgent extends OpenCode with a multi-model AI agent orchestration system, providing 11 specialized agents, 54–61 lifecycle hooks, 20–39 integrated tools, and batteries-included tooling for code editing, task delegation, and background execution.
For practical installation and setup, see Quick Start. For detailed explanation of core terminology and agent behaviors, see Key Concepts and Agents. For how to use tools and workflow guidance, see Tools and Features.
Sources: README.md103-110
Oh My OpenAgent is a powerful plugin for the OpenCode ecosystem that transforms it from a single agent platform into a coordinated multi-agent orchestration system. It is designed to run on heterogeneous AI models from multiple providers, orchestrating them intelligently to maximize performance and reduce costs.
Key characteristics include:
Multi-Agent System: 11 specialized agents perform distinct roles, such as main orchestration (Sisyphus), deep autonomous work (Hephaestus), planning (Prometheus), execution management (Atlas), research (Librarian), exploration (Explore), consultation (Oracle), and more docs/reference/features.md4-18
Lifecycle Hooks: 54–61 hooks spanning Session management, Tool Guards, Message Transforms, Continuation workflows, and Skills integration provide fine-grained extension and control AGENTS.md48
Rich Tooling: A comprehensive suite of 20–39 tools supports code manipulation (LSP, AST grep, hashline edits), task delegation, background processing, session management, interactive shells, and consensus decision making AGENTS.md48
Hash-Anchored Editing: The unique hashline_edit system ensures content integrity by using line hashes to validate edits and prevent mismatches AGENTS.md48
Multi-Layer MCP System: Incorporates built-in MCPs (LSP, AST-grep), Claude Code compatibility MCPs, and skill-embedded MCPs to extend capabilities AGENTS.md48
Team Mode: Supports parallel multi-agent coordination with member sessions, shared worktrees, and synchronized communication AGENTS.md48
Open Provider Support: Not locked to Claude or OpenAI; designed to orchestrate models from Anthropic, OpenAI, Google (Gemini), GPT, Kimi, GLM, and open-source providers.
Oh My OpenAgent ships in two editions:
Ultimate Edition: Full-featured OpenCode plugin experience with 11 agents, extended hooks, Team Mode, and complete tools docs/guide/installation.md3-8
Light Edition: Portable subset designed to run inside the OpenAI Codex CLI’s plugin system with 8 components and compatible MCPs docs/guide/installation.md3-8
Sources: README.md103-110 AGENTS.md48 docs/guide/installation.md3-8 docs/reference/features.md4-18
This diagram illustrates the plugin's modular architecture. The entry point pluginModule in packages/omo-opencode/src/index.ts delegates creation to createPluginModule which loads configurations, creates managers (tmux, background tasks, MCPs), registers tools, composes hooks in tiers, and finally assembles the plugin interface integrating with OpenCode’s runtime hooks.
The core agent system consists of highly specialized agents designed for efficient collaboration, while tools provide rich APIs for code manipulation and workflow automation.
Sources: packages/omo-opencode/src/index.ts1-19 AGENTS.md48
The plugin follows a strict initialization flow triggered by OpenCode loading pluginModule.server. It loads and merges configuration from both user and project scopes, validates with Zod, and then proceeds to create core managers responsible for concurrency, background tasks, and terminal multiplexing. Tools are registered next, followed by composition of a five-tiered hook system that intercepts OpenCode events. Finally, the plugin interface is created to bind the components into OpenCode's runtime.
Sources: packages/omo-opencode/src/index.ts1-19 AGENTS.md58-80
Oh My OpenAgent provides 11 specialized agents, each tailored with dedicated prompt engineering, model preferences, and tool access.
| Agent | Purpose | Agent Mode | Recommended Model(s) |
|---|---|---|---|
| Sisyphus | Main Orchestrator: overall planning, delegation | primary | claude-opus-4-7, kimi-for-coding/k2p5, GLM 5+ |
| Hephaestus | Deep Autonomous Worker: end-to-end coding | primary | gpt-5.5 |
| Atlas | Todo List Orchestrator: executes plans | primary | claude-sonnet-4-6 |
| Prometheus | Strategic Planner: interview/interrogation mode | primary | claude-opus-4-7 |
| Oracle | Architecture Consultant: read-only expert | subagent | gpt-5.5 |
| Librarian | Documentation and OSS Search | subagent | google/gemini-3-flash |
| Explore | Fast Codebase Grep and Contextual Exploration | subagent | github-copilot/grok-code-fast-1 |
| Metis | Plan Consultant: pre-planning analysis | subagent | claude-sonnet-4-6 |
| Momus | Plan Reviewer: validates plans | subagent | gpt-5.5 |
| Multimodal-Looker | Visual Content Specialist: PDFs, images, diagrams | subagent | gpt-5.5 |
| Sisyphus-Junior | Category-based executor for specialized tasks | subagent | Dynamic category-based model matching |
Agents coordinate work through a delegation hierarchy, exchange context, and enforce task completion with typical fallback models configured per agent.
Sources: docs/reference/features.md4-33 docs/guide/agent-model-matching.md13-39
Tools facilitate agent interaction with code, sessions, background tasks, and shell environments:
delegate_task, call_omo_agent for passing work among agents.hashline_edit (hash-anchored editing), ast_grep_search, lsp_rename, diagnostics.interactive_bash (tmux integration), background_task, session_read, skill tools.All tools are gated by agent permissions and category configurations.
Sources: AGENTS.md48
Hooks allow injection of custom behavior and system-wide policies:
todoContinuationEnforcer.The system has a complete lifecycle, enabling flexible extension and fine-grained control over the AI workflow.
Sources: AGENTS.md48 packages/omo-opencode/src/hooks/AGENTS.md1
Configuration is loaded and merged from multiple locations in order:
$HOME in .opencode/oh-my-openagent.json[c] (or legacy oh-my-opencode.json[c]).~/.config/opencode/oh-my-openagent.json[c].User configs override project configs when conflicts occur. The system supports JSONC format for comments and trailing commas.
Sources: docs/reference/configuration.md44-58
| Field Category | Description |
|---|---|
| Agent Overrides | Customize agent model, fallback chains, temperature, tools. |
| Categories | Define categories that map to specific models and prompts. |
| Disabled Features | Disable agents, tools, hooks, skills, MCPs, commands, providers. |
| Behavior Flags | Enable or disable features like hashline_edit, model_fallback. |
| Background Task Limits | Concurrency limits per provider and model. |
These allow fine-tuning the AI agent orchestration and resource usage.
Sources: assets/oh-my-opencode.schema.json1-101 docs/reference/configuration.md44-58
The ultrawork (alias ulw) command triggers the main orchestration process. It engages the Sisyphus agent to:
This mode embodies "just do it" autonomy.
Sources: README.md109
By pressing Tab or invoking Planning Agents, you enter an interview-style planning mode that:
Running /start-work launches Atlas to orchestrate and execute the plan.
Sources: docs/reference/features.md22
The hashline_edit tool ensures the integrity of code edits by embedding line content hashes (LINE#ID) within files. This prevents unintentional overwrite and enforces that the content matches expected hashes before applying changes.
Sources: AGENTS.md48
The recommended way is to delegate installation to an LLM agent (such as Claude Code or Cursor) to avoid configuration errors:
Paste into your agent session:
Install and configure oh-my-openagent by following the instructions here:
https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/refs/heads/dev/docs/guide/installation.md
This installs the Ultimate Edition by default, configuring providers, models, and agent authorization.
For manual setup and troubleshooting, see the full Installation Guide
Sources: docs/guide/installation.md20-29
ultrawork CommandOnce installed, simply type:
ultrawork
The system autonomously engages the full multi-agent architecture to complete your requested task end-to-end, including research, coding, and verification.
For more control, press Tab to enter the planning mode, then use /start-work to run a detailed work plan.
Sources: README.md109
oh-my-openagent's multi-agent orchestration centers on the following foundational concepts:
ultraworkThe main command to start an autonomous agent orchestration session managed by the Sisyphus agent, enabling parallel, iterative task execution until done.
Sources: README.md109
Specialized AI agents with distinct roles, prompts, and permissions, collaborating via delegation. They include orchestrators, planners, deep workers, consultants, explorers, and specialists.
Sources: docs/reference/features.md4-33
A classifier that analyzes user input to extract intent and classify the request, enabling Sisyphus to delegate to the most suitable agent or category.
Sources: AGENTS.md48
A robust code editing mechanism that leverages line content hashes (LINE#ID) to validate file modifications and ensure edit integrity.
Sources: AGENTS.md48
A dynamic routing system where tasks are assigned to category-specific agents instead of fixed model names. Categories like visual-engineering, ultrabrain, or quick automatically map to optimized models.
Sources: docs/reference/configuration.md99-123
Reusable, discoverable capabilities embedded in the system or MCPs, which agents can invoke to perform complex operations or evaluations.
Sources: AGENTS.md48
For further information on installation, detailed concepts, and workflows, please see the child pages:
Sources: README.md103-110 docs/guide/overview.md1-146 AGENTS.md48 docs/reference/configuration.md docs/reference/features.md3-63
Refresh this wiki