LazyCodex is a distribution layer that installs the OmO (oh-my-openagent) agent harness into Codex, so a Codex session gains project memory, planning, execution, skills, hooks, model routing, and verification defaults in one pass. The harness itself lives in the oh-my-openagent project, included here as a submodule; LazyCodex packages it for Codex the way LazyVim packages lazy.nvim for Neovim.
Work is driven by commands typed with a $ prefix in the Codex composer. $init-deep scores complex directories and writes hierarchical AGENTS.md guidance next to the code that needs it. $ulw-plan runs a Socratic interview and parallel codebase exploration, then writes a decision-complete plan to plans/ .md without touching product code. $start-work executes that plan until every checkbox is done, tracking durable progress across turns. $ulw-loop repeats a task until an Oracle role verifies completion, capped at 500 iterations in ultrawork mode and 100 otherwise.
Sub-agent roles — explorer, librarian, plan, momus, metis, and codex-ultrawork-reviewer — are installed into ~/.codex/agents/ and selected by passing agenttype to Codex's spawnagent tool; when a Codex build lacks that parameter, the skills fall back to describing the role inside the message. Model selection is routed per task category rather than fixed: small edits go to a smaller model, hard logic to a high-reasoning one, and agentic coding paths to Codex-tuned models.
Features
- Project memory: $init-deep generates hierarchical AGENTS.md context so agents start from local guidance in a large repository
- Planning: $ulw-plan turns ambiguous work into a decision-complete plan file and never writes product code
- Plan execution: $start-work runs a plan to completion with durable progress state and parallel subagents
- Verified completion: $ulw-loop keeps iterating until a completion claim is backed by verification, with iteration caps per mode
- Sub-agent roles: explorer, librarian, plan, momus, metis, and a review role, selectable through Codex's native multi-agent tools
- Skill library: review-work, remove-ai-slops, frontend-ui-ux, programming, LSP, AST-grep, rules, and comment-checker skills activate by domain
- Model routing: task categories and fallback chains pick a cheaper or stronger model per subtask
- Diagnostics: a doctor command reports plugin cache, hooks, MCP servers, agent roles, and config state
