Lazycodex

LazyCodex installs an agent harness inside Codex, OpenAI's coding agent, to help it handle complex codebases. It is described as a light port of the Hephaestus deep worker agent from the OmO project, built around goals rather than step by step recipes and around verified completion, meaning work counts as done only when observable evidence confirms it. The harness adds project memory, planning, parallel agents, skills, hooks and routing. Including the keyword ultrawork, or the short form ulw, anywhere in a prompt switches Codex into an evidence driven orchestration mode that begins with a research wave: an Explorer agent reads the local implementation while a Librarian agent checks documentation and contracts, and their findings feed a planning agent. Several commands structure the work. The ulw loop command runs until a verification step by an Oracle agent confirms completion, with an iteration cap of 500 in ultrawork mode and 100 otherwise. The ulw plan command interviews the user and produces a written work plan without writing product code. The start work command executes that plan with parallel subagents, test driven development and evidence gates, keeping durable state across turns. It targets developers who already use Codex on large repositories. Installation is a single npx command, which wraps the oh my openagent installer for the Codex platform. The site shows no paid plans and publishes no pricing.