The oh-my-opencode project, also known as Sisyphus, presents itself as a batteries-included agent harness designed to significantly enhance the capabilities and productivity of Large Language Models (LLMs) in coding and development tasks within the OpenCode environment. The central objective is to transform LLM agents into highly efficient, disciplined, and collaborative "teammates" that emulate human development teams, overcoming common LLM limitations such as quitting halfway, inefficient context management, and limited tool interaction.

The core methodology revolves around a sophisticated agent orchestration framework centered on a main agent, "Sisyphus" (typically leveraging anthropic/claude-opus-4-5), which acts as a project manager. Sisyphus is designed for strategic planning, intelligent delegation, and relentless execution, emphasizing a "todo-driven workflow" and aggressive parallel task processing. Instead of a single monolithic agent, oh-my-opencode implements a multi-agent system where Sisyphus delegates specialized tasks to a curated team of subagents, each optimized for specific domains and often utilizing different, task-appropriate LLM models:

• oracle (openai/gpt-5.2): For architectural design, code review, and strategic problem-solving due to its strong logical reasoning.
• frontend-ui-ux-engineer (google/gemini-3-pro-preview): Specializes in UI development, leveraging Gemini's creative capabilities.
• librarian (opencode/glm-4.7-free): Conducts multi-repository analysis, documentation lookup, and provides implementation examples, focusing on evidence-based answers.
• explore (dynamic, e.g., opencode/grok-code, google/gemini-3-flash, anthropic/claude-haiku-4-5): Facilitates blazing-fast codebase exploration and pattern matching.
• document-writer (google/gemini-3-flash): Handles technical writing.
• multimodal-looker (google/gemini-3-flash): Analyzes visual content like PDFs and diagrams.

A fundamental technical detail of this orchestration is the heavy reliance on background agents and parallel execution. Sisyphus doesn't waste its primary context on auxiliary tasks like file hunting or exhaustive searches. Instead, it fires off these tasks to faster, cheaper subagents in parallel using the call_omo_agent tool with a run_in_background parameter. The main agent is notified upon completion and can retrieve results, allowing for complex, concurrent workflows (e.g., debugging with GPT while Claude explores alternative solutions, or frontend work by Gemini concurrently with backend work by Claude). This approach significantly reduces the main agent's context load, making it more efficient and focused on high-level planning.

Beyond agent orchestration, oh-my-opencode provides agents with a rich suite of developer tools that are typically exclusive to human IDEs, allowing them to perform more deterministic and surgical operations:

• Full LSP/AST-Grep Support: Agents can leverage Language Server Protocol (LSP) functionalities (lsp_hover, lsp_goto_definition, lsp_find_references, lsp_rename, lsp_code_actions, etc.) for precise refactoring, navigation, and analysis. Additionally, ast_grep_search and ast_grep_replace enable Abstract Syntax Tree (AST)-aware code pattern matching and replacement across 25 languages, ensuring contextually aware code modifications.
• Curated MCPs (Multi-Capability Providers): Integrated external capabilities via tools like websearch (Exa AI), context7 (official documentation lookup), and grep_app (GitHub code search) for real-time, external information gathering. Skills can also embed their own MCP servers for specialized operations (e.g., playwright for browser automation).
• Session Management Tools: Agents have tools (session_list, session_read, session_search, session_info) to navigate, search, and maintain continuity across their own historical sessions, preventing redundant effort and enabling long-running projects.
• Contextual Awareness Mechanisms:
  • AGENTS.md/README.md Injector: Automatically injects relevant AGENTS.md and README.md files from the file's directory path up to the project root into the agent's context when a file is read, providing hierarchical, context-specific instructions.
  • Conditional Rules Injector: Injects rules from .claude/rules/ based on matching conditions (via globs in frontmatter), ensuring agents follow dynamic project-specific or user-defined guidelines.
  • look_at tool: Leverages subagents to intelligently extract only necessary information from large files or documents, preventing context window bloat.
• Compatibility Layer: Features a Claude Code compatibility layer, allowing existing configurations and plugins from that ecosystem (settings.json for hooks, custom commands, skills, agents, and MCPs) to function directly within oh-my-opencode.

To ensure task completion and address common LLM behavioral issues, the project incorporates several behavioral enforcement and enhancement mechanisms:

• ultrawork (or ulw) Keyword: A "magic word" that, when included in a prompt, activates the maximum performance mode, triggering parallel agent orchestration, deep exploration, and relentless execution until task completion.
• Ralph Loop: A self-referential development loop (/ralph-loop) that continuously runs the agent until a task is explicitly completed (indicated by $<promise>DONE</promise>$) or maximum iterations are reached, preventing agents from stopping prematurely.
• Todo Continuation Enforcer: Forces agents to complete all identified TODOs before concluding a task, combating the LLM tendency to quit halfway.
• Comment Checker: Reminds agents to justify or remove excessive comments, promoting clean, human-like code output.
• Think Mode: Automatically detects phrases like "ultrathink" and adjusts model settings for extended reasoning.
• Context Window Monitor: Reminds agents about available context window headroom to encourage thoroughness and prevent hurried outputs.
• Anthropic Auto Compact: Automatically summarizes and compacts session history for Claude models when token limits are approached, maintaining continuity without manual intervention.

Installation for humans involves a simple bunx oh-my-opencode install command, with prompts for configuring Claude, ChatGPT, and Gemini subscriptions. For LLM agents, a detailed step-by-step guide is provided, including how to query user subscription types and run the CLI installer with appropriate flags (--claude=<yes|no|max20>, --chatgpt=<yes|no>, --gemini=<yes|no>) and configure authentication for various providers (Anthropic, Google Gemini with Antigravity OAuth, OpenAI with Codex Auth). It explicitly outlines how to override default agent model settings to utilize specific models provided by authentication plugins (e.g., google/antigravity-gemini-3-pro-high).

In summary, oh-my-opencode is a comprehensive framework that transforms LLM interaction from simple prompt-response into a sophisticated, multi-agent collaborative environment. By providing LLMs with advanced tooling (LSP, AST-Grep, MCPs), robust orchestration capabilities (parallel background tasks, intelligent delegation), and clever behavioral enforcers, it aims to unlock unprecedented levels of autonomy and effectiveness for AI-driven development.