FAQ

General

What is the difference between OpenCode and oMO?

OpenCode is the AI coding CLI — sessions, context, language models. oMO (oh-my-openagent) is an orchestration layer on top that adds a multi-agent pipeline (Prometheus → Atlas → Sisyphus), category-based task delegation, and a skills ecosystem. OpenCode runs the agents. oMO tells them how to coordinate.

Do I need oMO?

No. OpenCode works great on its own. oMO adds planning, delegation, and verification for larger projects. Try OpenCode first, add oMO when you want the pipeline.

Does oMO work on Linux and Windows?

Yes. oMO is a configuration layer, not a binary. As long as OpenCode runs on your machine, oMO works. The install steps use standard paths (~/.opencode/ or %APPDATA%\opencode\ on Windows). On Linux, replace any brew commands in guides with your package manager or the npm/source install method.

What if my project doesn't have a CLAUDE.md or AGENTS.md?

oMO still works — it just won't have project-specific context. Sisyphus explores the codebase to build its own understanding. Add a CLAUDE.md (for Claude Code sessions) or AGENTS.md (for OpenCode/Codex sessions) when you want consistent conventions across sessions: stack, test commands, naming rules, anything an agent should always know.

---

Pipeline

When does the pipeline fire?

Only when you type /start-work. Normal conversation doesn't trigger it — Sisyphus handles those directly. If you just say "add a login endpoint," Sisyphus does it without involving Prometheus or Atlas.

Can I skip Prometheus and go straight to Atlas?

Yes. If you already have a plan at .sisyphus/plans/your-plan.md, Atlas can read it directly. Useful when you've written the plan yourself or want to reuse one from a previous session.

Can I stop the pipeline mid-execution?

Yes. Tell Sisyphus to stop at any point:

"Stop — I want to change the approach before you continue"
"Pause after Task 3 and wait for me to review"

Atlas will halt. In-progress tasks finish first; queued tasks don't start.

What if build verification fails during the pipeline?

Atlas catches the failure and tries to fix it. If it fails twice, it stops and asks you. You can:

• Give Atlas more context about why it failed
• Skip verification for that task and continue
• Stop the pipeline and debug manually

What happens if I run /start-work on an existing codebase?

Prometheus reads the existing code before planning. It discovers your stack, conventions, and structure, then plans around them — it won't scaffold from scratch if there's already a project. This makes /start-work safe to use for large new features on existing projects too.

---

Skills

What's the difference between a skill and a category?

A skill injects domain instructions into an agent — tdd adds red-green-refactor rules, diagnose adds structured debugging steps. A category is an execution profile with a specific model — visual-engineering uses a premium UI-tuned model, quick uses a fast free model. They're independent axes that compose: task(category="deep", load_skills=["tdd"]) gives you a strong model that follows TDD.

How do I disable a skill or MCP?

Add it to the disabled_skills or disabled_mcps list in oh-my-openagent.json:

{
  "disabled_skills": ["ai-slop-remover"],
  "disabled_mcps": ["playwright"]
}

Disabled skills won't load even if an agent tries to use them. Disabled MCPs won't start even if a skill references them.

Do I need Matt Pocock's skills?

They're optional. They add structured engineering processes — grilling sessions, TDD loops, backlog triage, architecture reviews. Worth trying /grill-me and /tdd at minimum. See Matt Pocock Skills for the full list.

How do I create my own skill?

Use /write-a-skill and follow the SKILL.md format. The most important part is the description field — it's how agents decide whether to load your skill. See Create Your Own for the full guide.

Which skill should I use for a code review?

Use /review-work for a general review of what was just implemented. Use /grill-me before implementation to align on intent. Use /improve-codebase-architecture when you want a broader architectural assessment. These are sequential, not interchangeable: grill → implement → review.

---

Models

Which model should I use?

Start with the defaults. oMO ships with sensible category-to-model assignments. Tune only if you notice quality issues with a specific task type, or if you want to reduce costs by routing simple work to free models. See Model Selection.

How do I know which model is being used for a task?

Sisyphus reports the category and skill for each delegated task. The category tells you the model tier. To see the exact model string, check the categories section of your oh-my-openagent.json for that category.

What if a model is down?

The fallback chain kicks in automatically. Each category tries its primary model first, then falls through the list in order. You don't need to do anything. If all models in the chain fail, the agent stops and tells you.

Can I use a different AI provider?

oMO uses whatever models OpenCode supports. Model strings in the config follow the opencode-go/<model-name> format for models available through OpenCode's model routing. Check OpenCode's documentation for the current list of supported providers and model strings.

---

Troubleshooting

Sisyphus picked the wrong category for my task — how do I fix it?

You can override category selection explicitly:

"Handle this with the visual-engineering category"
"Use ultrabrain for this — the default picked quick and it's not enough"

If it keeps picking the wrong category for a specific task type, add explicit routing notes to your CLAUDE.md or AGENTS.md: "All auth-related tasks use the deep category." Sisyphus reads these files at session start.

An agent keeps making the same mistake — what do I do?

Add a rule to CLAUDE.md / AGENTS.md so every session starts with the constraint:

## Rules
- Never use any ORM — raw SQL only via database/sql
- Always check errors explicitly, no _ assignment for errors

For one-off corrections, just tell Sisyphus mid-session. For recurring mistakes, put it in the project file.

A skill isn't loading — how do I debug it?

1. Check the skill is in the right directory (~/.opencode/skills/ or the project's skills/ folder)
2. Check the SKILL.md frontmatter is valid — name, description, and type are required
3. Check the skill isn't in disabled_skills in your config
4. Try loading it explicitly: "Load the [skill-name] skill and use it for this task"