English | 简体中文
An iterative development-loop skill for coding agents: first-principles build → commit & deploy (staging/local only, never production) → adversarial black-box review by a subagent → repeat until 100% → cleanup & report.
Instead of write-code-and-hope, devloop runs a closed loop. Your agent restates the problem from first principles, builds the minimal correct implementation, commits and deploys it to a safe target — staging or local, never production — then hands the running artifact to a fresh-eyes subagent whose only job is to break it. Whatever the reviewer finds feeds the next iteration. The loop only exits at a 100% review score, and finishes by cleaning up dev artifacts and reporting a full iteration table (version, git SHA, review score, deploy state).
┌──────────────────────────────────────────────────────────┐
│ 0. Clarify: restate the problem from first principles — │
│ essence, success criteria, invariants │
│ 1. Develop: minimal correct implementation, fix root │
│ causes, verify locally with evidence │
│ 2. Commit: commit + push, record the exact SHA │
│ Deploy: server-side changes → staging; │
│ frontend-only / no staging → local verification │
│ ❌ Production release commands are never executed │
│ 3. Review: a subagent treats the work as untrusted and │
│ probes the running artifact adversarially │
│ 4. Decide: score <100% → new fix plan, back to step 1 │
│ 5. Cleanup: remove dev artifacts, report the iteration │
│ table (version / git SHA / score / deploy state) │
└──────────────────────────────────────────────────────────┘
The repo is a plugin for every agent that supports plugin manifests — and skills/devloop/ is an ordinary skill directory you can copy anywhere.
| Agent | Install | Invoke |
|---|---|---|
| Devin CLI / Desktop | devin plugins install 598934321/devloop |
/devloop:devloop |
| Devin CLI / Desktop (bare) | cp -r skills/devloop ~/.config/devin/skills/ |
/devloop |
| Claude Code | cp -r skills/devloop ~/.claude/skills/ |
/devloop |
| Codex | cp -r skills/devloop ~/.codex/skills/ |
/devloop |
| Cursor | cp -r skills/devloop ~/.cursor/skills/ |
/devloop |
| Kimi Code | plugin manifest .kimi-plugin/ (ships tool-name mapping) |
/devloop |
| Gemini CLI | gemini extensions install https://github.com/598934321/devloop |
/devloop |
| agentskills-compatible | cp -r skills/devloop .agents/skills/ (project) or ~/.agents/skills/ (global) |
/devloop |
| Any other agent | copy skills/devloop/ into that agent's skills directory |
/devloop |
Tool references inside SKILL.md are harness-neutral (subagent / task-list / question tools); per-agent manifests carry skillInstructions mappings where the format supports them.
Two skill variants ship side by side — skills/devloop/ (简体中文, /devloop) and skills/devloop-en/ (English, /devloop-en); pick one.
CI (scripts/validate.mjs + GitHub Action) checks every manifest parses, version is semver, declared skills dirs exist, and each SKILL.md frontmatter name/description matches its directory.
/devloop add faceted filtering to the search page, mobile-friendly
The skill iterates autonomously until the adversarial review scores 100%, or stops and reports when it hits a decision only you can make.
- Never ships to production — staging or local deploy is the hard stop; production release stays a human decision
- Adversarial review — the subagent probes the running artifact against the requirement, not the code's intent; it reports findings but never edits
- Auditable — every iteration is logged with version, git SHA, review score, and deploy state; the final report is a full iteration table
- Project-agnostic — reads each project's own conventions (AGENTS.md/CLAUDE.md, deploy scripts, test commands) instead of hardcoding any stack
devloop/
├── .devin-plugin/plugin.json # Devin plugin manifest
├── .claude-plugin/plugin.json # Claude Code plugin manifest
├── .codex-plugin/plugin.json # Codex plugin manifest
├── .cursor-plugin/plugin.json # Cursor plugin manifest
├── .kimi-plugin/plugin.json # Kimi Code manifest (tool-name mapping)
├── .agents/plugins/marketplace.json # agentskills marketplace
├── gemini-extension.json # Gemini CLI extension
├── GEMINI.md # context for Gemini CLI
├── skills/
│ ├── devloop/SKILL.md # the skill itself — 简体中文 (agent-neutral)
│ └── devloop-en/SKILL.md # English variant
├── scripts/validate.mjs # structural validation
├── .github/workflows/validate.yml # CI
├── README.md / README.zh-CN.md
└── LICENSE
MIT