Key Concepts

Relevant source files

This page defines the core terminology and technical abstractions of the oh-my-openagent (formerly oh-my-opencode) system. It explains how high-level user intents are translated into coordinated actions across agents, tools, and background processes.

1. Ultrawork (`/ultrawork` or `ulw`)

Ultrawork is the primary high-level orchestration mode that engages the full multi-agent "Discipline" system. It is designed to run autonomously until a task is completed, utilizing parallel execution and self-correcting loops README.md109 It is the flagship feature of the project, designed to "just work" until the task is done README.md81-82

Implementation and Data Flow

When a user triggers ultrawork, the system follows a specific activation pipeline:

Keyword Detection: The keyword-detector hook AGENTS.md58 identifies the /ultrawork or ulw string in the message.
Mode Activation: It injects a mode-specific system prompt that enforces the ultrawork protocol on the agent.
The Ralph Loop: If the task is complex, the ralph-loop (or ulw-loop) AGENTS.md63 ensures continuation. If an agent becomes idle but tasks remain, the todoContinuationEnforcer (also known as the Boulder mechanism) AGENTS.md60 forces a session resume to maintain momentum.

Sources: README.md103-109 AGENTS.md43-63 docs/guide/overview.md134-141

2. Discipline Agents

The "Discipline Agents" are a specialized set of AI personas, each with distinct models, prompts, and tool permissions. They work in a hierarchy to prevent "cognitive drift" where a single agent loses track of a complex task.

Agent Inventory and Model Mapping

The system includes 11 built-in agents docs/reference/features.md5

Agent	Role	Default Model	Purpose
Sisyphus	Main Orchestrator	`claude-opus-4-7`	Main planner and delegator; follows ~1,100 line prompt docs/reference/features.md13
Hephaestus	Deep Worker	`gpt-5.5`	Deep technical specialist for autonomous multi-file reasoning docs/reference/features.md14
Oracle	Consultant	`gpt-5.5`	Architecture consultation and read-only expert docs/reference/features.md15
Librarian	Researcher	`gpt-5.4-mini-fast`	External documentation and multi-repo analysis docs/reference/features.md16
Explore	Searcher	`gpt-5.4-mini-fast`	Fast, contextual grep and codebase exploration docs/reference/features.md17
Prometheus	Strategic Planner	`claude-opus-4-7`	High-level planning via interview mode docs/reference/features.md23

Natural Language to Code Entity Space: Agent Resolution

The mapping from a user's natural language request to a specific code entity is handled by the agent registry and builder pattern during the initialization flow AGENTS.md45-57

Title: Agent Resolution Flow

Sources: AGENTS.md45-57 docs/reference/features.md5-32 docs/guide/agent-model-matching.md43-70

3. IntentGate

IntentGate (implemented via the keyword detector system) is a critical layer that analyzes user input before any tools are executed. It ensures that complex requests are routed to the correct reasoning level AGENTS.md43

Functionality

Classification: It identifies if the request requires a specific specialist or mode (e.g., triggering ultrawork or team modes) AGENTS.md58
Validation: It ensures models follow structured reasoning protocols, such as thinking blocks assets/oh-my-opencode.schema.json168-186
Context Management: The system manages token limits during high-intensity tasks through aggressive truncation flags docs/reference/configuration.md140

Sources: AGENTS.md43-58 assets/oh-my-opencode.schema.json168-186 docs/reference/configuration.md140

4. Hash-Anchored Edit (`hashline_edit`)

To solve the "hallucination" problem in code editing, OmO uses a Hash-Anchored Edit system AGENTS.md43

Implementation

Hash Injection: When an agent reads a file, the system appends a unique LINE#ID content hash to every line AGENTS.md43
Verification: When the agent submits an edit, the system verifies that the target line's current content matches the hash provided by the agent assets/oh-my-opencode.schema.json106-108
Configuration: This feature is enabled by default but can be toggled via the hashline_edit configuration key assets/oh-my-opencode.schema.json106

Sources: AGENTS.md43 assets/oh-my-opencode.schema.json106-108 docs/reference/configuration.md34

5. Categories and Skills

Skills

Skills are modular extensions that allow agents to discover and use new tools dynamically assets/oh-my-opencode.schema.json45

Discovery: Skills are loaded from specialized directories like .agents/skills/ AGENTS.md17
Embedded MCPs: Skills can include their own Model Context Protocol servers AGENTS.md60
Lifecycle: The system includes 53-60 lifecycle hooks that skills can tap into AGENTS.md43

Natural Language to Code Entity Space: Delegation Flow

The transition from a natural language intent to a technical delegation call is mediated by the task tool and the skill-mcp-manager.

Title: Delegation and Skill Flow

Sources: AGENTS.md17-60 docs/reference/configuration.md99-124 docs/guide/overview.md69-71 assets/oh-my-opencode.schema.json45-65