Sisyphus (Main Orchestrator)

Relevant source files

Purpose and Scope

This page documents Sisyphus, the primary orchestrator agent in Oh My OpenCode. Sisyphus coordinates work across specialized sub-agents, tools, and skills to complete user requests from start to finish. It serves as the default agent and handles most user interactions through intelligent delegation and parallel execution.

For information about:

Planning workflows with Prometheus and Atlas, see 3.2 Planning Agents
Specialized worker agents that Sisyphus delegates to, see 3.3 Worker Agents
How Sisyphus selects models for different work types, see 3.4 Agent-Model Matching
Dynamic prompt construction mechanics, see 3.5 Dynamic Prompt Building
Task delegation system, see 4.1 Task Delegation
Background execution, see 4.4 Background Task Tools

Overview

Sisyphus is the main orchestrator agent that analyzes user requests, determines optimal approaches, and coordinates specialized agents and tools for execution. Rather than attempting direct solutions, Sisyphus operates on a delegation-first principle—routing work to the most appropriate specialized agents while managing parallel execution and verification packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts101-103

The name reflects the agent's philosophy: just as Sisyphus pushes his boulder uphill daily, developers push code daily. Sisyphus aims to produce code indistinguishable from a senior engineer's work packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts88-90

Key Characteristics

Characteristic	Value	Notes
Mode	`primary`	Main entry point for user interaction packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts21-22
Default Model Class	`GPT-5.4`, `Gemini`, `Claude`, `Kimi K2.7`	Optimized prompts for each provider packages/omo-opencode/src/agents/sisyphus-junior/agent.ts7-10
Core Strategy	Delegation-first	Prefers specialist agents over direct work packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts102
Execution Style	Parallel	Multiple background tasks and tool calls simultaneously packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts103

Sources: packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts1-112 packages/omo-opencode/src/agents/sisyphus-junior/agent.ts31-40 packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts12-31

Architecture

Configuration and Creation Pipeline

Sisyphus is initialized via a factory pattern that resolves the best available model based on user overrides or system defaults packages/omo-opencode/src/agents/sisyphus-junior/agent.ts95-106

Sisyphus creation logic:

Model Resolution: Resolves the best available model (e.g., GPT-5.4, Kimi K2.7, Claude Opus 4.7) based on provider connectivity and requirements packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts33-65
Prompt Building: Triggers the dynamic prompt builder to inject available agents, tools, and skills into the system instruction packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts58-76
Permission Guarding: Applies model-specific tool permissions, such as blocking task for subagents while allowing call_omo_agent for research packages/omo-opencode/src/agents/sisyphus-junior/index.test.ts196-224

System Integration Diagram

Sources: packages/omo-opencode/src/agents/sisyphus-junior/agent.ts95-157 packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts11-31

Dynamic Prompt Construction

The buildKimiK27SisyphusPrompt and similar functions construct a massive system instruction by aggregating various sections based on the current environment packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts58-81

Section Builders (packages/omo-opencode/src/agents/dynamic-agent-prompt-builder.ts)

Key Triggers: Identifies phrases (e.g., "External library mentioned → fire librarian") that should bypass normal classification packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts66
Tool Selection: A table mapping task complexity to agents or direct tools (LSP, AST, Search) packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts67
Category/Skills Guide: Maps user-defined categories and skills to the delegation system packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts70
Task Management: Instructions for TaskCreate or todowrite based on the useTaskSystem flag packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts38-56

Sources: packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts58-81 packages/omo-opencode/src/agents/sisyphus/claude-opus-4-7.ts44-81

Core Workflows

Intent Gate: Request Classification

Sisyphus uses an intent gate to prevent "AI slop" and literal misinterpretations packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts114-116 It is mandated to classify intent before classification packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts119-130

Intent → Routing Map Sisyphus is instructed to verbalize its intent before acting packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts132:

Surface Form	True Intent	Routing Decision
"explain X", "how does Y work"	understanding	explore, then answer in prose packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts123
"implement X", "add Y"	code changes	plan, then delegate or execute packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts124
"look into X", "check Y"	investigation	explore, report, wait packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts125
"what do you think?"	judgment	evaluate, propose, wait packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts126

Sources: packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts114-132 packages/prompts-core/prompts/ultrawork/gemini.md7-35

Delegation Protocol

Sisyphus follows an "Orchestrate by default" rule packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts102 It uses the delegation system with a structured prompt format often involving specialized sub-agents. The orchestrator determines if a task is "Trivial", "Moderate", "Complex", or "Research" to decide whether to execute directly or delegate packages/prompts-core/prompts/ultrawork/gpt.md41-49

Sisyphus-Junior (Subagent Variant)

For focused execution tasks that do not require full orchestration, Sisyphus can spawn Sisyphus-Junior. This variant acts as a focused executor that takes one delegated task and carries it to completion itself packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts32-36 It executes but does not orchestrate, though it may fire explore or librarian via call_omo_agent for research packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts36

Sources: packages/omo-opencode/src/agents/sisyphus-junior/agent.ts1-11 packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts1-88 packages/prompts-core/prompts/ultrawork/gpt.md39-55

Model-Specific Variants

Sisyphus adapts its core instructions based on the underlying model's architecture to maximize performance.

Kimi K2.7 Variant

The Kimi K2.7 variant is authored from the ground up to be restrained and outcome-first packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts1-13

Decision Rules: Replaces rituals with judgment-based operating rules packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts98-106
Task System: Uses TaskCreate or todowrite based on configuration to track multi-step work packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts38-56

Gemini Variant

Gemini models receive specific corrective overlays to ensure tool usage compliance.

Intent Gate: Mandates a Step 0 classification before any tool call packages/prompts-core/prompts/ultrawork/gemini.md7-35
Tool Mandate: Explicitly warns that "Thinking is not doing" and mandates tool use packages/prompts-core/prompts/ultrawork/gemini.md113-126
Thinking Config: Injects thinking with a 32,000 token budget packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts145-148

Claude Variant (Opus 4.7)

Parallel Tool Calls: Explicitly enabled via <use_parallel_tool_calls> packages/omo-opencode/src/agents/sisyphus/claude-opus-4-7.ts107-109
Literal Following: Instructed to counter 4.7's tendency to generalize by following instructions literally packages/omo-opencode/src/agents/sisyphus/claude-opus-4-7.ts99-102

Sources: packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts1-132 packages/prompts-core/prompts/ultrawork/gemini.md1-143 packages/omo-opencode/src/agents/sisyphus/claude-opus-4-7.ts1-129

Task Creation and Verification Loops

Sisyphus uses tasks to track progress and ensures every step is verified packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts40-46

Natural Language to Code Entity Space Mapping

Verification Strategy

Trivial Changes: Requires lsp_diagnostics on the file packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts70
Behavioral Changes: Requires running tests that import the changed module and watching them pass packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts71
Cross-cutting Changes: Requires full build exit 0 and real-surface verification via Bash packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts72

Sources: packages/omo-opencode/src/agents/sisyphus/kimi-k2-7.ts38-56 packages/omo-opencode/src/agents/sisyphus-junior/kimi-k2-7.ts66-74 packages/prompts-core/prompts/ultrawork/gpt.md108-113

Testing and Verification

Sisyphus's prompt logic and model-specific configurations are verified via unit tests.

Key tests cover:

Model Overrides: Ensuring createSisyphusJuniorAgentWithOverrides correctly applies model, temperature, and description overrides packages/omo-opencode/src/agents/sisyphus-junior/index.test.ts11-53
Reasoning Configuration: Verifying that GPT models receive reasoningEffort: "medium" while Claude models get injected thinking configs packages/omo-opencode/src/agents/sisyphus-junior/index.test.ts146-182
Tool Safety: Ensuring that subagents like Sisyphus-Junior have the task tool blocked but call_omo_agent allowed packages/omo-opencode/src/agents/sisyphus-junior/index.test.ts196-224

Sources: packages/omo-opencode/src/agents/sisyphus-junior/index.test.ts1-224 packages/omo-opencode/src/agents/sisyphus-agent-factory.test.ts1-151