Specialized Assistants
Relevant source files
The following files were used as context for generating this wiki page:
Specialized Assistants are a modular collection of AI agents within the OhMyOpenCode ecosystem, each designed with a specific operational focus, reasoning capacity, and toolset. These agents function as subagents or advisors to the primary orchestrator, Sisyphus, enabling the system to handle complex tasks by delegating specialized work to the most appropriate model. The architecture relies on the "Conductor-Specialist" model, where a central agent manages parallel execution and verification of outputs from these assistants.
Sources: sisyphus-prompt.md:12-20, src/agents/sisyphus.ts:162-175, src/agents/atlas/gpt.ts:12-16
Architecture and Core Assistants
The system utilizes specialized subagents to separate concerns such as planning, research, architecture consultation, and plan verification. This specialization ensures that high-cost reasoning models are only invoked for complex decision-making, while cheaper, faster models handle data retrieval or contextual grep operations.
Assistant Taxonomy
| Agent | Category | Role | Cost |
|---|---|---|---|
| Prometheus | Planning | Strategic planning and interview conductor | N/A |
| Sisyphus | Utility | Primary orchestrator and Deep Worker | EXPENSIVE |
| Oracle | Advisor | Strategic technical advisor and architecture consultant | EXPENSIVE |
| Metis | Advisor | Pre-planning consultant and failure point analyst | EXPENSIVE |
| Momus | Advisor | Practical work plan reviewer | EXPENSIVE |
| Librarian | Exploration | External reference and documentation researcher | CHEAP |
| Explore | Exploration | Internal codebase contextual grep | FREE |
Sources: src/agents/oracle.ts:13-17, src/agents/librarian.ts:9-13, src/agents/metis.ts:250-264, src/agents/momus.ts:160-174, sisyphus-prompt.md:160-165
Orchestration Flow
The following diagram illustrates the interaction between the user, the primary orchestrator (Sisyphus), and the specialized advisory agents during a standard work cycle.
The flow highlights the mandatory review gates (Metis and Momus) for non-trivial tasks. Sources: src/agents/prometheus/plan-generation.ts:20-40, src/agents/momus.ts:114-125, src/agents/metis.ts:220-240
Research and Exploration Agents
Librarian
The Librarian is specialized in multi-repository analysis and official documentation retrieval. It operates as a "Reference Grep" for external resources, using tools like GitHub CLI and websearch_exa to find evidence-backed implementation examples.
- Logic: Classifies requests into conceptual (Type A), implementation (Type B), context (Type C), or comprehensive (Type D).
- Documentation Discovery: Performs a sequential discovery process:
websearch-> version check -> sitemap discovery -> targeted investigation. - Citations: Mandates permalink construction (e.g.,
github.com/<owner>/<repo>/blob/<sha>/<path>#L<start>-L<end>) for every evidence-based claim.
Sources: src/agents/librarian.ts:40-100, src/agents/librarian.ts:182-200
Explore
The Explore agent serves as a "Contextual Grep" for internal codebases. It is used liberally as a peer tool to direct search functions to find patterns, structure, and hidden conventions within the current project.
Sources: sisyphus-prompt.md:200-215, src/agents/dynamic-agent-prompt-builder.ts:89-100
Planning and Review Agents
Prometheus: The Strategic Planner
Prometheus is strictly a planner and is forbidden from writing code or executing implementation tasks. Its primary output consists of clarified requirements and markdown-based work plans saved to .sisyphus/plans/*.md.
- Interview Mode: Default behavior involves interviewing the user to clear a "Clearance Checklist" (Core objective defined, Scope established, Technical approach decided).
- Draft Protocol: Must continuously record decisions to a draft file in
.sisyphus/drafts/to prevent context loss. - Plan Template: Generates plans containing a TL;DR, Context, Work Objectives, and a mandatory Verification Strategy (QA scenarios).
Sources: src/agents/prometheus/identity-constraints.ts:12-60, src/agents/prometheus/plan-template.ts:10-50
Metis and Momus: The Quality Gates
Metis and Momus act as defensive agents to prevent AI failures and scope creep.
- Metis: Analyzes requests before Prometheus starts planning. It identifies hidden intentions, flags over-engineering patterns (AI-slop), and prepares directives for the planner.
- Momus: Acts as a ruthless reviewer for the final plan. It focuses exclusively on executability and reference validity (e.g., "Do the referenced files actually exist?"). It uses an "Approval Bias," rejecting plans only for true blockers like missing files or internal contradictions.
Sources: src/agents/metis.ts:10-30, src/agents/momus.ts:15-50, src/agents/momus.ts:80-95
Advisor Agent: Oracle
Oracle is a high-IQ, read-only reasoning model invoked for debugging hard problems and high-difficulty architecture design.
Oracle Consultation Triggers
| Domain | Trigger Condition |
|---|---|
| Architecture | Multi-system tradeoffs, unfamiliar patterns |
| Self-Review | Completion of significant implementation |
| Hard Debugging | 2+ failed fix attempts |
| Risk Assessment | Security or performance concerns |
Sources: src/agents/oracle.ts:13-28, sisyphus-prompt.md:465-475
Response Structure
Oracle responses are structured in three tiers to ensure actionable insights:
- Essential: Bottom line (2-3 sentences), Action plan (numbered steps), and Effort estimate.
- Expanded: Rationale, trade-offs, and "watch out for" risks.
- Edge Cases: Escalation triggers and alternative high-level outlines.
Sources: src/agents/oracle.ts:88-105
Implementation Constraints and Safety
All specialized assistants operate under a set of "Hard Blocks" to maintain codebase integrity.
- No Type Safety Suppression: Use of
as any,@ts-ignore, or@ts-expect-erroris strictly forbidden. - Verification Guarantee: No task is marked "done" without evidence (clean
lsp_diagnostics, passing builds, and verified test runs). - Atomic Operations: Implementation agents must follow a "Red-Green-Refactor" cycle when TDD is enabled.
Sources: src/agents/sisyphus.ts:390-410, src/agents/hephaestus.ts:400-420, sisyphus-prompt.md:520-530
Summary
Specialized Assistants provide a structured, multi-agent approach to software engineering. By isolating planning (Prometheus), research (Librarian), analysis (Metis), review (Momus), and architectural reasoning (Oracle), the system ensures high-quality code generation and project-scale consistency while minimizing reasoning costs and implementation errors.