Skip to content
 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

60 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🏠 agent-homebase

The operating system your AI coding assistant is missing

13 agents · 3 delivery modes · 7 enforced schemas · 4 platforms

CI GitHub stars License: MIT Copilot Claude Cursor


Quick Start · What You Get · Three Modes · Why It Works · Full Docs



Copilot, Claude, and Cursor are powerful. But without structure, they're chaos machines. agent-homebase turns your AI assistant into a professional delivery team.


🔥 The Problem Everyone Ignores

You're using GitHub Copilot or Claude Code. They're fast. They're capable.

They also:

  • Forget your project conventions after every session
  • Ship code without tests, reviews, or documentation
  • Require you to manually coordinate planning → coding → QA
  • Lose all learned patterns when context resets
  • Make inconsistent decisions across identical problems

You have a powerful tool with no process. That's not engineering — that's gambling.


💡 The Solution: A Professional Delivery Team in Your Config

agent-homebase is a portable, multi-agent operating system for software projects.

Skills, instructions, and agents are authored once in plain Markdown, resolved with your project-specific tokens, and emitted as copy-ready artifacts for any compatible coding agent — GitHub Copilot, Claude Code, Cursor, or OpenAI Codex.

One config file transforms chaotic AI assistance into a structured software pipeline:

Your feature request
        ↓
   @planner      → Scopes requirements, estimates complexity
        ↓
   @pm           → Validates against project goals (prevents scope creep)
        ↓
   @sprint-lead  → Orchestrates full implementation cycle
        ↓
   @qa           → Enforces tests, coverage, linting (blocks failures)
        ↓
   @reviewer     → Catches security issues, anti-patterns, code smells
        ↓
   @docs         → Updates documentation automatically
        ↓
   Deployed code — tested, reviewed, documented

📈 Before & After

❌ Without agent-homebase ✅ With agent-homebase
You: "Add user authentication"

Agent: [writes 200 lines immediately]
       • No tests
       • Hardcodes API keys
       • Uses deprecated patterns
       • Forgets your Auth0 setup

You: "Wait, we use Auth0..."

Agent: [rewrites everything]
       • Still no tests
       • Documentation? What docs?
You: "@planner add user auth"

@planner: Scoped 4 tasks (~8 files)
          Using Auth0 per config
          Complexity: Medium

You: "@sprint-lead run Sprint 12"

@sprint-lead:
  ✔ Implementation complete
  ✔ Tests passing (91% coverage)
  ✔ 14 security checks passed, 0 CVEs
  ✔ Docs updated
  ✔ Pushed to feature branch

🎯 What You Get

13 Specialized Agent Roles

Agent What It Does
🎯 @pm Validates features through 5 structured tests — prevents scope creep before code is written
📋 @planner Scopes requirements, drafts sprint plans, estimates complexity
🎬 @sprint-lead Orchestrates complete sprint execution with zero manual coordination
@qa Runs typecheck, lint, tests, coverage — blocks deploys on failures
🔍 @reviewer Reviews for patterns, security, performance — flags issues by severity
🏗️ @architect Designs approaches, writes ADRs, makes structural decisions
📚 @researcher Surfaces prior art with citations, identifies failure modes
🐛 @bug Captures bugs into structured, prioritized backlog
📖 @docs Keeps documentation in sync after every sprint
🔐 @security 14 automated security checks — see details below
@a11y WCAG 2.1 AA accessibility audits
@perf Bundle size, build time, dependency analysis
🚀 @onboarding Ramps up new team members and explains project conventions

🔐 Security: Built In, Not Bolted On

@security runs 14 automated checks as a sprint gate or on demand:

Check What It Catches
Dependency CVE scan Known vulnerabilities in your packages
Active exploit research CISA KEV catalog, proof-of-concept detection
Secret detection API keys, tokens, passwords in source
OWASP pattern matching Injection, XSS, auth failures
File integrity hashes Unauthorized changes to tracked files
SBOM generation Full software bill of materials
SAST scanning Static analysis security findings
Git history secret scan Secrets in old commits
License compliance Denylist/allowlist enforcement
HTTP security headers Missing CSP, HSTS, X-Frame-Options
Container image scan Vulnerabilities in Docker images
IaC scanning Misconfigurations in infrastructure code
Supply chain audit Dependency provenance verification
Security changelog Append-only finding log (SEC-NNN entries)

Every finding gets an OWASP remediation classification: patched, delayed, no-fix, or zero-day — with timelines and effort tags.

7 Enforced Schemas

Every artifact is validated at build time — no more silent drift:

Schema Validates
frontmatter-v1 YAML frontmatter on every skill, instruction, and agent
callable-v1 Callable manifest on every skill
project-v1 A single project entry in a Mode 3 registry
registry-v1 A full Mode 3 choreography registry
subagent-return-tier1 Lightweight subagent return contracts
subagent-return-tier2 Standard subagent return contracts
subagent-return-tier3 Full subagent return contracts

3 Ready-to-Use Profiles

Profile For
react-web-app React + Vite single-page apps
monorepo-fullstack TypeScript monorepo (pnpm + Vite + Expo)
python-api FastAPI / Flask / Django backends

🔀 Three Delivery Modes

agent-homebase ships as three standalone, independently consumable modes on top of a shared protocol layer. Use one, two, or all three:

Mode What It Is When You Need It
Mode 1 — Team A substrate of skills, instructions, and agents for interactive use You want @planner, @qa, @security, etc. in your project
Mode 2 — Orchestration A dispatcher that pulls work from a queue, invokes callables, and verifies results You want non-interactive, issue-driven dispatch
Mode 3 — Choreography Coordinates a program of works across many projects with drift control You manage multiple projects and need cross-repo visibility

Each mode is a standalone install. Adoption order is free. Adding a new mode doesn't modify existing ones.

Visual reference: open docs/command-centre-visual.html in a browser for an interactive cheat sheet covering all three modes.


🚀 Quick Start

# 1. Clone
git clone https://github.com/j78f88/agent-homebase.git
cd agent-homebase

# 2. Install
pip install -r requirements.txt

# 3. Choose your profile and generate resolved files
python init.py --config profiles/python-api.config.yml
# → resolved/skills, resolved/instructions, resolved/agents

# 4. Copy to your project
cp -r resolved/skills/* ../.github/agents/
cp -r resolved/instructions/* ../.github/instructions/
cp -r resolved/agents/* ../.github/agents/

# 5. Initialize planning files (first time only)
cp starters/SPRINTS.md ../
cp starters/BACKLOG_LEDGER.md ../docs/planning/

Then use naturally:

@planner scope the dark mode feature
@sprint-lead run Sprint 3
@qa check coverage for auth module
@reviewer look at the last 3 commits

📖 Complete setup guide → · Quickstart →


🧠 Why It Works

Thin Orchestration Architecture

@sprint-lead (coordinator)
    │
    ├── Reads plans
    ├── Tracks state
    ├── Manages workflow
    │
    └── Delegates ALL implementation to:
            ├── Unnamed subagents → write code
            ├── @qa → validate quality
            ├── @reviewer → check patterns
            └── @docs → update documentation

The secret: @sprint-lead never reads source code. It coordinates. This keeps context windows clear for what actually matters.

Contract-First Design

Every agent returns structured data following JSON schemas:

{
  "tier": 1,
  "agent": "qa",
  "status": "blocked",
  "summary": "2/5 quality gates failed",
  "blockerReason": "Coverage at 72% (threshold: 85%)",
  "findings": [...]
}

When @qa returns status: blocked, @sprint-lead knows exactly what to do. No guessing. No hallucinating.

Protocol-v1: Frozen Contracts

protocol-v1 is frozen at 2.0.0. Every skill, instruction, and agent gets validated against JSON Schemas at build time. Frontmatter validation is strict by default. No more drift — if it builds, it conforms.

Works Everywhere

author once → resolve with tokens → deploy to any platform:

Platform Status
GitHub Copilot ✅ Ready
Claude Code ✅ Ready
Cursor ✅ Ready
OpenAI Codex ✅ Ready

🗂️ Architecture (1 minute)

skills/           author once, token-templated, callable-v1 manifests
instructions/     shared rules, generic + configurable
agents/           per-agent body that wraps a skill
schemas/          7 JSON Schemas that gate the build
config/           project-specific token values
profiles/         pre-built configs (python-api, react-web-app, monorepo-fullstack)
command-centre/   contracts (protocol-v1, mode contracts, ADRs, ref impls)
        │
        ▼  python init.py --config <profile>
resolved/         deploy artifacts (skills/, instructions/, agents/)

init.py is the single source of truth for the build. It runs security validation, frontmatter validation (strict by default), resolves {{tokens}}, and writes deterministic output.

resolved/ is build output — never edit it directly, never commit it.


🤔 Is This For You?

✅ Perfect fit ❌ Not for you
  • You use Copilot/Claude/Cursor for real development
  • You're tired of re-explaining context every session
  • You want consistent quality without babysitting
  • You value process but don't want to build from scratch
  • You manage multiple projects and need cross-repo consistency
  • You only use AI for quick code snippets
  • You prefer building your own agent framework from scratch
  • You don't use any AI coding assistants

📚 Documentation

Guide What You'll Learn
Onboarding Step-by-step setup (start here)
Quickstart Fastest path to running
Architecture Design decisions & rationale
Personas Who this is for (evidence-tagged)
Skill Flow How agents orchestrate work
Example Sprint Complete sprint walkthrough
Three Modes Team, Orchestration, Choreography explained
Customization Adapting skills for your needs
Extension Guide Authoring new skills
Troubleshooting Common issues & fixes

🤝 Contributing

Contributions welcome. See CONTRIBUTING.md and AGENTS.md.

  • Conventional Commits (feat:, fix:, docs:, chore: …).
  • Every PR must keep python init.py green and the test suite passing.
  • New skills must include a callable-v1 manifest and follow EXTENSION_GUIDE.md.

📄 License

MIT © 2026



Your AI assistant has the power.

Give it the process.


⭐ Star on GitHub · 📖 Get Started · 🐛 Report Issue



Extracted from real production teams. Battle-tested on TypeScript, Python, React, and monorepo projects.
protocol-v1 frozen at 2.0.0 — contracts enforced by 7 JSON Schemas.

About

skills, agents, instructions to drop in my agentic framework for reuse

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages