Specialized Agents
Relevant source files
The following files were used as context for generating this wiki page:
Specialized Agents are domain-specific AI entities within the Oh-My-OpenCode ecosystem designed to handle specific phases of the software development lifecycle, from strategic planning and documentation research to deep implementation and rigorous review. These agents operate under a primary orchestrator, such as Sisyphus, which delegates tasks based on the specialized capabilities and "skill" requirements of each agent.
The architecture emphasizes a "multi-model agent orchestration" approach, utilizing various frontier models (Claude 4.5, GPT-5.2, Gemini 3) to optimize for task complexity, cost, and reasoning requirements.
Sources: AGENTS.md:162-167, src/agents/sisyphus.ts:108-118
Agent Ecosystem Overview
The project categorizes agents into several functional roles. Some serve as primary executors, while others act as advisors or explorers.
| Agent | Category | Primary Role | Preferred Model |
|---|---|---|---|
| Sisyphus | Primary | Orchestrator & Multi-task Worker | Claude Opus 4.5 |
| Prometheus | Strategic | Strategic Planning Consultant | Claude Opus 4.5 |
| Hephaestus | Primary | Autonomous Deep Worker | GPT-5.2 Codex |
| Oracle | Advisor | Architecture & Debugging Consultant | GPT-5.2 |
| Metis | Advisor | Pre-Planning & Intent Analysis | (Subagent) |
| Librarian | Exploration | Docs & Open Source Code Search | GLM-4.7 |
| Momus | Advisor | Plan Review & Quality Assurance | (Subagent) |
| Explore | Exploration | Fast Contextual Codebase Grep | Grok-Code-Fast |
Sources: AGENTS.md:195-207, src/agents/oracle.ts:11-30, src/agents/momus.ts:21-36
Strategic Planning: Prometheus & Metis
The planning phase is strictly separated from implementation. This is handled by Prometheus, a dedicated planner that is forbidden from writing implementation code.
Metis: Pre-Planning Consultant
Metis is invoked before a plan is generated to identify hidden intentions, unstated requirements, and "AI-slop" patterns like over-engineering or scope creep.
Metis provides "directives" that Prometheus must follow, such as mandatory verification tools or specific codebase patterns to emulate.
Sources: src/agents/metis.ts:14-35, src/agents/metis.ts:167-185
Prometheus: The Strategic Planner
Prometheus acts as a consultant that interprets implementation requests as requests for a Work Plan. It generates a single atomic plan file in .sisyphus/plans/*.md.
Key Constraints for Prometheus:
- Planning Only: Forbidden from editing source code; may only modify
.sisyphus/*.mdfiles. - Single Plan Mandate: Everything must go into one plan regardless of size to maintain architectural consistency.
- Verification Strategy: Every task must be verifiable without human intervention, using tools like Playwright or curl.
Sources: src/agents/prometheus/identity-constraints.ts:14-45, src/agents/prometheus/identity-constraints.ts:100-117, src/agents/prometheus/plan-template.ts:71-80
Implementation: Sisyphus & Hephaestus
Once a plan is established, implementation is handled by primary workers who follow the plan's TODO list.
Hephaestus: Autonomous Deep Worker
Hephaestus is designed for autonomous problem-solving. It uses an "Explore-First" protocol, where it searches for information using tools before asking the user for clarification.
Sisyphus: The Orchestrator
Sisyphus manages the dev team. It uses a Category + Skills system to delegate tasks to specialized subagents. For example, UI tasks are sent to visual-engineering with the frontend-ui-ux skill.
Sources: src/agents/sisyphus.ts:114-128, src/agents/sisyphus.ts:203-216
Advisory & Quality: Oracle & Momus
Specialized agents also provide safety nets and high-level consultation.
Oracle: The Advisor
Oracle is a high-IQ, read-only reasoning model. It is consulted for multi-system tradeoffs, security concerns, or when implementers fail to fix a bug after two attempts.
Decision Framework:
- Bias toward simplicity: Recommendations must favor the least complex solution.
- Pragmatic Minimalism: Favor existing dependencies over new ones.
- Effort Tagging: Estimates work as Quick (<1h), Short (1-4h), Medium (1-2d), or Large (3d+).
Sources: src/agents/oracle.ts:40-54, src/agents/oracle.ts:60-65
Momus: The Plan Reviewer
Momus reviews work plans generated by Prometheus. Unlike a perfectionist, it has an Approval Bias, looking only for "Blockers" that would prevent a developer from starting work.
| Check Category | Failure Condition (REJECT) | Pass Condition (OKAY) |
|---|---|---|
| References | Referenced file does not exist. | Reference exists but is imperfect. |
| Executability | Task is too vague to begin. | Developer can figure out minor gaps. |
| Blockers | Internal contradictions in plan. | "Could be clearer" suggestions. |
Sources: src/agents/momus.ts:43-65, src/agents/momus.ts:88-112
Knowledge Retrieval: Librarian
The Librarian is a specialized agent for searching remote codebases and official documentation. It follows a specific "Date Awareness" protocol to ensure it does not use outdated search results.
Librarian Investigation Types:
- Type A (Conceptual): Uses
context7andwebsearchfor documentation discovery. - Type B (Implementation): Clones remote repos using
gh repo cloneto find source code permalinks. - Type C (History): Searches GitHub issues and PRs to find context on "Why" something changed.
Sources: src/agents/librarian.ts:50-70, src/agents/librarian.ts:108-125, src/agents/librarian.ts:160-170
Conclusion
The specialized agent system focuses on the separation of concerns between planning, implementation, and verification. By utilizing specialized entities like Oracle for reasoning and Momus for review, the system ensures high-quality output while leveraging autonomous workers like Hephaestus to reduce manual developer overhead.
Sources: src/agents/prometheus/identity-constraints.ts:38-42, src/agents/hephaestus.ts:110-120