Category System

Categories are agent profiles that map task types to optimized models. They're not skills — they're the execution layer that determines how a task runs. Skills determine what the agent knows.

How Categories Work

When Sisyphus delegates a task, it doesn't pick a model. It picks a category. The category maps to a specific model and fallback chain:

User: "Debug the auth timeout issue"
  → Sisyphus picks category: deep
    → deep maps to: deepseek-v4-pro (primary)
    → Fallback: glm-5.1 → minimax-m2.7 → minimax-m2.5-free

The agent says what kind of work it is. The harness picks the right model. You touch nothing.

Available Categories

Category	Domain	Typical Model Tier	Cost
visual-engineering	Frontend, UI/UX, design, animation	Premium	$$$
ultrabrain	Hard logic, architecture decisions, complex reasoning	Premium	$$$
deep	Thorough autonomous research + execution	Premium	$$$
artistry	Creative problem-solving, exploration	Medium	$$
quick	Single-file changes, typos, small fixes	Free	$
unspecified-low	Simple misc tasks that don't need a premium model	Free	$
unspecified-high	Complex misc tasks that need a premium model	Premium	$$$
writing	Documentation, prose, copy	Free	$

When Each Category Fires

• visual-engineering: Any task involving UI, styling, layout, animation, or visual design. The model is tuned for design taste and CSS expertise.
• ultrabrain: Architecture decisions, complex algorithmic problems, multi-step reasoning. Routes to the strongest reasoning model available.
• deep: Autonomous work that requires thorough exploration: debugging, refactoring, research-heavy tasks. The model has room to think deeply.
• artistry: Creative exploration, brainstorming, alternative approaches. Medium-tier model with creative tendencies.
• quick: Trivial changes: typos, single-line fixes, small config edits. Uses the fastest available model.
• unspecified-low: Tasks that don't fit other categories but don't need premium compute.
• unspecified-high: Tasks that don't fit other categories but need premium compute.
• writing: Documentation, README files, comments, prose. Uses a model good at natural language.

Category + Skill Composition

Categories and skills are independent axes. The category provides the execution profile; the skill provides domain instructions:

// Visual work with design skill
task(category="visual-engineering", load_skills=["frontend-ui-ux"], prompt="...")

// Deep debugging with diagnosis skill
task(category="deep", load_skills=["diagnose"], prompt="...")

// Quick fix, no skill needed
task(category="quick", load_skills=[], prompt="...")

// Architecture review with deep thinking
task(category="ultrabrain", load_skills=["improve-codebase-architecture"], prompt="...")

Composition Matrix

Task	Category	Skill	Why
Build a React component	visual-engineering	frontend-ui-ux	Design-first UI on a visual model
Debug a production issue	deep	diagnose	Thorough exploration with structured diagnosis
Write tests for a feature	deep	tdd	Methodical test-first on a strong model
Plan a new project	ultrabrain	grill-with-docs	Hard reasoning with domain-aware planning
Quick typo fix	quick	—	Fast model, no skill overhead
Review your work	unspecified-high	review-work	Premium model for thorough review
Write documentation	writing	docs-writer	Prose-optimized model with docs skill
Refactor architecture	deep	improve-codebase-architecture	Deep model with architecture analysis

Configuration

Categories are configured in oh-my-openagent.json:

{
  "categories": {
    "visual-engineering": {
      "model": "opencode-go/deepseek-v4-pro",
      "fallback_models": [
        {"model": "opencode-go/glm-5.1"},
        {"model": "opencode-go/minimax-m2.7"},
        {"model": "opencode-go/minimax-m2.5-free"}
      ]
    },
    "deep": {
      "model": "opencode-go/deepseek-v4-pro",
      "fallback_models": [
        {"model": "opencode-go/glm-5.1"},
        {"model": "opencode-go/minimax-m2.7"}
      ]
    },
    "quick": {
      "model": "opencode-go/minimax-m2.5-free",
      "fallback_models": [
        {"model": "opencode-go/big-pickle"}
      ]
    }
  }
}

Each category has independent fallback chains. If the primary model is unavailable or rate-limited, the next model in the chain takes over automatically.

Custom Categories

You can define custom categories beyond the defaults. A custom category is just a named profile with a model and fallback chain:

{
  "categories": {
    "security-audit": {
      "model": "opencode-go/deepseek-v4-pro",
      "fallback_models": [
        {"model": "opencode-go/glm-5.1"}
      ]
    }
  }
}

Then use it in task delegation:

task(category="security-audit", load_skills=[], prompt="Audit the auth module for vulnerabilities")

Custom categories let you tune model selection for your specific workflow without modifying the defaults.