dev-loop

dev-loop

A virtuous development loop for Rails, as a Claude Code plugin. One command chains six phases with deliberate model tiers and an independent quality gate:

Phase	Who	Model
1 · Plan & orchestrate	the session	Opus, max reasoning — writes PLAN.md (the grading contract) + worktree
2 · Implement	specialist army	Sonnet, parallel, via /rails-feature (roundhouse)
3 · Adversarial review	Codex	/codex:adversarial-review — challenges approach & assumptions
4 · Rate against the plan	independent reviewer	Opus — scores each axis + overall plan-fidelity
5 · Fix (gated)	specialist army	Sonnet — only if overall < 8.5 or any axis < 7; ≤ 2 rounds
6 · Learnings & ship	the session	Opus — appends durable learnings, opens the PR

The point is the separation of concerns: a planner, an independent implementation army, an adversarial reviewer, an independent rater (not the implementer grading itself), a gated fixer, and a learning sink — so quality is enforced by structure, not willpower.

Install

/plugin marketplace add kurenn/dev-loop   # if not already on the kurenn marketplace
/plugin install dev-loop@kurenn

Prerequisites (recommended)

dev-loop wraps two other plugins. It degrades gracefully without them, but it's designed to run with both:

/plugin install roundhouse@kurenn      # Phase 2: the Sonnet specialist army
/plugin install codex@openai-codex     # Phase 3: the adversarial review

Without roundhouse, Phase 2 falls back to directly-spawned Sonnet agents. Without codex, Phase 3 falls back to an Opus adversarial reviewer agent.

Usage

/dev-loop-setup            # once per repo: checks prereqs, scaffolds config + learnings file
/dev-loop <feature or bug> # run the full loop

Per-project configuration

The loop reads a ## Dev-loop config block from the repo's CLAUDE.md (scaffolded by /dev-loop-setup). Override any of:

• Gate — default overall ≥ 8.5 AND every axis ≥ 7
• Base / extra rating axes — e.g. add design-system fidelity for UI-heavy apps
• Critical paths — e.g. money / auth / KYC; those axes also require ≥ 8.5
• Learnings file — default docs/dev-loop-learnings.md
• Fix-round cap — default 2
• PR conventions — repo-specific screenshot/embed/gh quirks

Defaults apply when the block is absent, so /dev-loop works before setup — but /dev-loop-setup tailors the gate and rigor tiers to the project.

Why model tiers

• Opus plans and rates because planning and judging quality are the high-leverage, reasoning-heavy steps.
• Sonnet implements and fixes because the work is parallelizable and well-specified once the plan exists — an army of cheaper agents covers more ground.
• A separate Opus rates the work so the grader isn't the implementer. Self-rating inflates; independent rating against a written plan does not.
• Codex challenges from outside the Claude family for a genuinely adversarial second opinion on the approach, not just defect-spotting.

License

MIT © Abraham Kuri