Features

Features you’ll think should’ve always existed. Once you use them, you can’t go back.

Ralph Loop / /ulw-loop

Self-referential loop. Doesn’t stop until 100% done.

The Ralph Loop is the engine behind ultrawork. When activated:

Agent executes a task cycle
Checks if everything is truly complete
If not, loops back and continues
Repeats until every subtask is verified done

This is why ultrawork doesn’t stop halfway. The loop enforces completion.

Activate via:

ultrawork or ulw — automatic Ralph Loop
/ulw-loop — explicit loop activation

Todo Enforcer

Agent goes idle? System yanks it back.

Todo Enforcer monitors task progress. When an agent finishes a subtask and has nothing queued, it doesn’t stop — the system pulls the next todo item and forces the agent to continue.

Your task gets done, period.

Comment Checker

No AI slop in comments. Code reads like a senior wrote it.

Comment Checker rejects low-quality AI-generated comments:

Obvious restatements of code (// increment i by 1 for i++)
Boilerplate JSDoc that adds no information
Generic doc strings that could apply to any function

/init-deep

Auto-generates hierarchical AGENTS.md files throughout your project:

project/
├── AGENTS.md              ← project-wide context
├── src/
│   ├── AGENTS.md          ← src-specific context
│   └── components/
│       └── AGENTS.md      ← component-specific context

Agents auto-read relevant context. Zero manual management. Great for both token efficiency and agent performance.

Run:

/init-deep

Session Tools

List, read, search, and analyze session history:

Action	What it does
List	Show all sessions with metadata
Read	Load a previous session’s full context
Search	Search across session transcripts
Analyze	Extract patterns and insights from past work

Session Recovery

Automatic recovery from:

Session errors and crashes
Context window limits (auto-truncation and resume)
API failures (automatic retry with backoff)

Your work survives interruptions. The system picks up where it left off.

Think Mode

Enhanced reasoning capability. When enabled, the agent spends more compute on planning and analysis before responding.

Useful for:

Complex architectural decisions
Multi-step debugging
Tasks requiring deep analysis

File Prompts

Load prompts from files using file:// syntax in agent configurations:

{
  "agents": {
    "oracle": {
      "prompt_append": "file://.opencode/prompts/oracle-custom.md"
    }
  }
}

This lets you manage agent instructions as version-controlled files instead of inline JSON strings.

Doctor Command

Built-in diagnostics:

bunx oh-my-opencode doctor

Checks:

Plugin registration in opencode.json
Configuration file validity
Model availability and API key access
Environment and dependency status

Run this first when something doesn’t work.

Model Fallbacks

Configure fallback models so tasks don’t fail when a provider is down:

{
  "fallback_models": {
    "default": ["claude-opus-4-6", "kimi-k2.5"],
    "hephaestus": [
      "gpt-5.4",
      { "model": "claude-opus-4-6", "reasoning_effort": "high" }
    ]
  }
}

Supports both plain model strings and per-fallback configuration objects in the same array.

Model Setup

Agent-model matching is built into the installation process. Each agent defaults to the model best suited for its role:

Agent	Default Model	Why
Sisyphus	claude-opus-4-6 / kimi-k2.5 / glm-5	Strong orchestration and planning
Hephaestus	gpt-5.4	Deep autonomous reasoning
Prometheus	claude-opus-4-6 / kimi-k2.5 / glm-5	Interview and planning
Experts	Varies by category	Category-based routing

See the Installation Guide for model setup details.

Next Steps

Agent System — How agents use these features
Tools — The developer tools powering agents
Configuration — Configure fallbacks, hooks, and more