yeachan-heo/gajae-code

Gajae-Code is a command-line tool that wraps an AI coding agent in a structured workflow, aimed at making AI-assisted programming more reliable and less prone to going off course. You install it as a package and run it from a terminal. It is early-stage and described as an experimental beta project. The core idea is a four-step loop: first, the agent asks clarifying questions before touching any code, then it builds and critiques a plan, then it runs the actual implementation, and finally it verifies the result against a stated goal. These four steps are packaged as named workflow skills called deep-interview, ralplan, team, and ultragoal. The intention is that the agent does not jump straight to writing code without understanding what you actually want. The tool also ships four role-based agents that specialize in different tasks: an executor for writing and refactoring code, an architect for read-only review of how the codebase is structured, a planner for sequencing work and defining acceptance criteria, and a critic for reviewing whether a plan is actionable before execution starts. For running multiple tasks in parallel, Gajae-Code can orchestrate sessions using tmux, a terminal multiplexer that lets you split a terminal into multiple panes running independently. It also supports git worktrees, which create an isolated copy of a repository so you can work on a branch without touching your main checkout. The project is published on the npm registry and runs on Bun as the recommended runtime. It is licensed under MIT and described by its author as a distillation of lessons from earlier agent harness projects they built for OpenAI and Anthropic APIs.