A repo-centered agent loop

architect-loop is a pair of Claude Code skills for running a manager and builder pattern. The README describes Claude Fable as the planning and review model, Codex as the builder and researcher, and the repository as the shared memory. The idea is to write specs and acceptance gates first, dispatch fresh Codex contexts into worktrees, then let the architect review evidence before integration.

The project is very young, created on June 12, 2026, so treat it as an emerging workflow rather than proven infrastructure. Still, it is worth indexing because it captures a pattern many agent users are now trying by hand: one model decides scope, another executes, and the repo records the state.

Install and commands

The README installation is simple:

git clone https://github.com/DanMcInerney/architect-loop
cd architect-loop && ./install.sh
npm i -g @openai/codex@latest

Windows uses install.ps1. The README says ./install.sh --project installs only into the current repository. You also need Claude Code on a paid plan and Codex CLI signed into a ChatGPT plan.

After install, the two commands are:

/architect
/architect-research <what you're considering>

What is different from a normal prompt

The workflow writes gates before builders start. Builders run in isolated worktrees and should only touch declared files. The architect runs gate commands itself, reads the diff, and decides whether to integrate. Repository memory lives in files such as docs/HANDOFF.md, docs/gates/, and docs/lanes/.

The useful discipline is that builder claims are treated as evidence to verify, not as truth. That matches the failure mode seen in real coding-agent work: tests can pass while a diff still misses the intent.

Where to be careful

This is not a free automation button. It consumes Claude and ChatGPT plan capacity, creates worktrees, writes process files, and depends on local CLI behavior. Use it first in a small repo where extra branches and docs are acceptable. Also inspect the installed skills before letting them drive a real branch.

For methodology-first agent workflows, compare obra/superpowers. For skill collections that feed agent behavior, see mattpocock/skills and phuryn/pm-skills.

FAQ

Does architect-loop require API keys? The README says no by default. It relies on Claude Code and Codex CLI sessions tied to paid plans.

What does /architect do? It runs a build loop: judge the last run, write the next slice, dispatch builders, and review evidence.

Why use worktrees? Worktrees isolate builder lanes so file ownership and diffs are easier to check.

Is it mature? No. It is a very new repo as of 2026-06, so treat it as a workflow experiment.