coding-team

Inspired by Stavros's How I write software with LLMs.

A Claude Code plugin that runs a multi-agent development team: an architect plans the work, a developer implements it, and two independent reviewers (Claude + OpenAI Codex) approve it before moving on.

Prerequisites

• Claude Code
• OpenAI Codex CLI installed and authenticated (used by the codex-reviewer agent)

Install

claude
/plugin marketplace add maxdubakov/multi-agent-development-claude-plugin
/plugin install coding-team@multi-agent-development
/reload-plugins

Exit current claude session. Plugin should be active on the next sessin

Permissions

Plugins cannot ship pre-configured permissions, so you'll need to allow these yourself on first use.

How it works

The plugin sets architect as the default agent. When you describe what you want to build, the architect goes in this loop:

1. Planning -- this step will take MOST of your time. It is a back and forth conversation about implementation with the architect agent. It will ask clarifying questions, suggest specific implementation path, and ask your opinion.
2. Task Definition -- once implementation path is accepted, architect splits it into multiple tasks. Each is written as the implementation progresses in .coding-team/<feature> folder.
3. Implementation -- developer implements one task at a time, following the Task Brief.
4. Parallel review -- code-reviewer (Opus) and codex-reviewer (Codex CLI) review the diff independently.
5. Feedback loop -- if either reviewer requests changes, architect verifies if the ask is valid, and decides if they are needed. If yes - it sends developer back with to implement them.
6. Next task -- only after both reviewers approve does the architect move to the next task.

Example of prompt:

Hey, I want to add an action to fetch an entity/view by URL.

Actions are stored here @ai-answer/actions. An action can be later called by our mcp server (which is not in that directory), or by a tool of an agent (for example, @ai-answer/agents/schema/tools/createSpace/createSpace.d.ts).

We had somewhere in @ai-answer/context/ logic to fetch a single entity by id.

We can get entity id by url (for example, provided such url, we can get public id, and from it, we can do basically the same): htt ps://the.fibery.io/Product_Management/feature/Add-an-action-to-fetch-an-entity-view-by-URL-7585

Agents

Agent	Model	Role
architect	Opus	Orchestrates the workflow, writes Task Briefs, never touches code
developer	Sonnet	Implements tasks, runs validation, reports back
code-reviewer	Opus	Reviews diffs for correctness, security, simplicity
codex-reviewer	Haiku (proxy)	Relays review to OpenAI Codex CLI and returns its output
repo-scout	Sonnet	Scans the repo and maintains ARCHITECTURE.md

License

MIT