Story

I built a small framework that turns the OpenAI Codex plugin into a swarm of specialized “workers” that all live inside a single repo. Instead of just chatting with a model in your IDE, you get an orchestrator, planner, coder and reviewer that share a JSON task board, touch only the files in your project and are forced to finish every step with a clean git commit.

The whole thing is defined in prompts and JSON (AGENTS.md + .AGENTS/*.json), with git and tasks.json as the shared memory layer. You describe a goal, the orchestrator decomposes it into atomic tasks, the planner writes them into the backlog, the coder makes small diffs and commits them, and the reviewer/doc agent keeps the repo and docs consistent. There’s no separate runtime or web UI — just your IDE, the Codex plugin and a strict one-task-one-commit workflow inside your local repo.

Repo: https://github.com/basilisk-labs/codex-swarm

Would love feedback on whether this “agents as versioned workers” model fits how you’d actually want to use LLMs in your dev workflow, and what’s missing for you to adopt or fork it.