lazycodex: an agent harness for Codex that provides project memory, strategic planning, and verified task completion for complex codebases

What it solves

LazyCodex provides a structured agent harness for managing complex codebases within Codex. It solves the problem of AI agents losing context in large repositories, lacking a strategic plan for implementation, or providing hopeful status updates instead of verified completion of tasks.

How it works

It acts as a distribution layer for the OmO (oh-my-openagent) engine. It integrates into Codex via plugins and hooks, providing a set of specialized commands and agent roles. It uses a multi-model routing system to assign tasks to the most appropriate model (e.g., high-reasoning models for logic, faster models for small edits) to optimize quota usage.

Who it’s for

Developers using Codex who want to automate complex software engineering tasks with a disciplined, multi-agent team (including roles like explorer, librarian, and reviewer) without extensive manual setup.

Highlights

• Strategic Planning: Uses $ulw-plan to create decision-complete plans in Markdown before any product code is written.
• Verified Completion: The $ulw-loop command runs tasks until they are verified by evidence, capping at 500 iterations in ultrawork mode.
• Project Memory: The $init-deep skill generates hierarchical AGENTS.md files to provide landmarks and context for future agents.
• Multi-Agent Orchestration: Supports parallel execution and specialized sub-agent roles (explorer, librarian, etc.) via Codex's native tools.
• Specialized Skill Library: Includes tools for AST-grep structural search, LSP diagnostics, and AI-slop removal.