Claude Code Orchestration Agents

Production-grade workflow agents for Claude Code. Drop-in sub-agent set that adds automated planning, review, audit, and test-coverage to every change you make.

What this is

A pre-built set of 5 Claude Code sub-agents plus a /change slash command that turns any project into a disciplined development environment:

• cm-planner — Architectural authority. Plans every change. Surfaces decisions that need your input before any code is touched.
• cm-decider — Routes individual questions to "decide now" vs "auto-resolve" based on 14 trigger criteria.
• reviewer — Reviews every staged diff before commit. Categorizes findings as blocking, medium, or nits.
• conventions-audit — Enforces decision-doc conventions when sub-phases close or decision files are touched.
• test-author — Adds tests for new behavior in monitored source paths. Surfaces source bugs that tests reveal.
• /change — Universal slash command that orchestrates the full pipeline.

The problem this solves

Claude Code is fast. Without discipline, that speed becomes risk:

• The AI changes 12 files when you asked for one
• Important decisions get baked in silently
• "Why did we do this?" mysteries surface months later
• Mistakes compound because there's no paper trail

This agent set adds checkpoints throughout the workflow. Every change goes through plan → verify → implement → review → audit → test → commit. You stay in control of architectural decisions while routine work flows automatically.

How it works

Your request
     ↓
/change <user request>
     ↓
cm-planner (plans scope, surfaces ratifications)
     ↓
[ if STOP_FOR_HUMAN: wait for your decision ]
     ↓
executor (Edit / Write / Bash with announced operations)
     ↓
git add (stage changes)
     ↓
reviewer (HIGH / MEDIUM / LOW findings on staged diff)
     ↓
conventions-audit (if docs/decisions/** touched OR sub-phase closing)
     ↓
test-author (if monitored source paths added new behavior)
     ↓
git commit (with clean conventional commit message)
     ↓
... continue or wrap sub-phase ...
     ↓
cm-planner "review commits for push" → git push

Installation

See INSTALL.md for the complete walkthrough. Quick version:

1. Clone or copy this repo's .claude/, docs/, and other files into your project root
2. Update your project's CLAUDE.md with the routing rules (full text in INSTALL.md)
3. Restart Claude Code (exit, then claude again)
4. Verify with /agents — all 5 should appear
5. Try /change update the README tagline as a practice run

Total time: ~15 minutes.

Requirements

• Claude Code installed (claude command available in your terminal)
• An existing project with at least an initial git commit
• Familiarity with the basic concepts of sub-phases, gates, and decision docs

If you haven't worked with that mental model before, start with the universal framework — same patterns but tool-agnostic. This repo extends it with automation.

What you get

After installation, your project has:

your-project/
├── .claude/
│   ├── agents/
│   │   ├── cm-planner.md         ← Plans every change
│   │   ├── cm-decider.md         ← Routes questions
│   │   ├── reviewer.md           ← Reviews diffs
│   │   ├── conventions-audit.md  ← Enforces conventions
│   │   └── test-author.md        ← Fills coverage gaps
│   └── commands/
│       └── change.md             ← /change slash command
├── docs/
│   └── conventions/
│       └── decision-docs.md      ← Decision-doc format spec
└── CLAUDE.md                     ← Updated with routing rules

Plus an AI assistant that uses all of them automatically.

Customization

Each agent file is plain markdown. Open and edit any of them to match your project's specific patterns:

• Add project-specific trigger phrases to cm-planner
• Adjust reviewer's blocking criteria
• Change conventions-audit's length budgets
• Update test-author's monitored paths

The files are intentionally generic — you make them yours.

What this is NOT

• ❌ Not a Claude Code installer (you bring your own)
• ❌ Not appropriate for tools other than Claude Code (use the universal framework for Kimi/Cursor/ChatGPT)
• ❌ Not zero-config — you'll edit the agents to match your project over time
• ❌ Not magic — the discipline works because YOU follow the pipeline, agents enforce it

Companion tools

This is part of a small ecosystem:

Tool	Audience	Where
Universal AI Orchestration Framework	Any AI coding tool, any project type	github.com/Aleksacom/ai-orchestration-framework
Claude Code Agent Set (this repo)	Claude Code users wanting full automation	You are here

Use whichever matches your tool. They share the same mental model.

Why this exists

I'm a business analyst in banking-sector IT — not a developer. When AI coding tools made it possible for me to build my own projects, I loved the speed but kept losing control. The AI would change things I didn't understand. Decisions got re-debated because nothing was written down. Mistakes compounded.

This agent set is what I built for myself, refined through real production project work. Sharing it freely so others can skip the painful learning curve.

License

MIT License — free to use, fork, modify, redistribute. See LICENSE for details.

Issues and feedback

Found a bug? Want to suggest an improvement? Open an issue.

Or just star the repo if it helped you. ⭐

---

Star this repo if it helped you. ⭐
Found a problem? Open an issue.
Want updates? Watch this repo for releases.