Core Concepts
How OmO connects intent, planning, models, agents, categories, skills, tools, and continuation.
System Shape
Intent Before Action
OmO does not treat every prompt as literal text. It first asks what the user really needs: explanation, investigation, implementation, planning, review, or delegation. That matters because the same sentence can imply different work depending on context.
| Surface request | Harness interpretation |
|---|---|
| “Why is this broken?” | Reproduce, diagnose root cause, fix if feasible. |
| “Can you look into this?” | Investigate and resolve, not just summarize. |
| “What's the best way?” | Choose the route and implement when the codebase implies action. |
| “Review this” | Find bugs, risks, missing tests, and verification gaps first. |
Planning vs Execution
Prometheus, Metis, Momus, and Atlas exist to separate decisions from implementation.
| Agent | Boundary |
|---|---|
| Prometheus | Interviews, researches, and writes plans. It should not edit production code. |
| Metis | Finds hidden intent, ambiguity, missing constraints, and AI failure points before a plan hardens. |
| Momus | Rejects vague plans until file references, acceptance criteria, and verification are concrete. |
| Atlas | Executes an approved plan through focused tasks, accumulates learnings, and verifies each step. |
| Sisyphus-Junior | Category-spawned executor. It focuses on one assigned unit and cannot re-delegate. |
The important rule is “Decision Complete”: a plan should leave no hidden product or architecture decisions to the implementer.
Model Families
OmO routes by agent and category because models behave differently.
| Family | Typical use |
|---|---|
| Claude-like models | Structured orchestration and disciplined instruction following. Claude Opus, Sonnet, Kimi K2.5, GLM 5 fit many Sisyphus-style flows. |
| GPT models | Explicit reasoning and architecture. GPT-5.4 is used for Hephaestus, Oracle, Momus, and hard logic categories. |
| Gemini models | Visual/frontend and creative work where a different reasoning style helps. |
| Fast utility models | Explore, Librarian, quick fixes, grep-style work, and cheap background research. |
Do not “upgrade” every task to the most expensive model. Explore and Librarian are intentionally fast and cheap because retrieval should not burn orchestration tokens.
Category + Skill
Categories describe the work; skills inject domain expertise and sometimes MCP servers.
| Category | Use |
|---|---|
visual-engineering | Frontend, UI/UX, design, styling, animation. |
ultrabrain | Hard logic, architecture, deep reasoning. |
deep | Autonomous research and implementation. |
artistry | Creative or unconventional problem solving. |
quick | Trivial single-file changes, typo fixes, simple modifications. |
unspecified-low | Low-effort work that does not fit another category. |
unspecified-high | High-effort work that does not fit another category. |
writing | Documentation, prose, technical writing. |
A good delegation names both:
task({
category: "visual-engineering",
load_skills: ["frontend-ui-ux", "playwright"],
run_in_background: false,
prompt: "Implement and browser-verify the responsive sidebar.",
});Continuation as a Harness Feature
The Sisyphus metaphor is literal: the system pushes unfinished work back into motion.
| Mechanism | What it protects against |
|---|---|
| Todo continuation | Agent stops while todos remain. |
| Ralph loop / ulw loop | Long tasks that need repeated self-continuation. |
| Boulder state | /start-work can resume a Prometheus plan across sessions. |
| Session recovery | Missing tool results, unavailable tools, empty content, context-limit failures. |
| Runtime fallback | Model/provider failures that should move to a configured fallback. |