<div align="center">

🌌 VibeGravityKit

<br>

<img src="https://raw.githubusercontent.com/Nhqvu2005/VibeGravityKit/main/Web.png" alt="VibeGravityKit Documentation Site" width="100%">

The AI-Native Software House in a Box. <br> Build enterprise-grade software with a team of 18 AI Agents — with parallel delegation for maximum speed and minimum token costs.

</div>

---

🎩 What is VibeGravityKit?

Imagine having a full-stack engineering team living inside your IDE. VibeGravityKit turns your IDE into a coordinated squad of 18 specialized agents, from the Architect who designs your database, to the Researcher who searches the web with DuckDuckGo.

But here's the killer feature: We hate wasting tokens.

• Context Manager: Minifies your code before the AI sees it. (Saves ~50% tokens).
• Context Router: Queries only relevant data from 34+ data sources. (Saves ~70% tokens).
• Diff Applier: Applies surgical patches instead of rewriting files. (Saves ~90% tokens).

---

🚀 How It Works — Two Ways to Build

VibeGravityKit gives you two powerful work modes to build software with your AI team:

Mode 1: @[/quickstart] — Instant Noodle 🍜

Describe your idea → Get a working product. Fast, automatic, delicious. Like instant noodles — just add water (your idea) and it's ready to eat.

You → Leader confirms plan → Auto-build → ♻️ Verify Loop → Deploy → Done!

Perfect for: MVPs, prototypes, demos, hackathons, or non-tech users who just want results.

How it works:

1. Describe your idea — even vague is fine ("I want an online store").
2. Leader auto-detects tech stack + checks template marketplace (saves ~70% tokens if match).
3. Leader shows you a simple checklist — you approve/edit before building.
4. Build runs fully automatic with parallel agents.
5. ♻️ Completion Loop (max 5): Leader scans codebase to verify EVERY feature in your checklist actually works. Missing or broken? → dispatches sub-agent to fix → loops until all ✅.
6. Auto-deploys via Cloudflare Tunnel → you get a live link immediately.

Example:

You: "Build me a simple URL shortener"
Quickstart:
  📋 Plan: [Home page ✓] [Shorten URL ✓] [Redirect ✓] [Copy link ✓]
  "Want to add or remove anything?"
You: "Looks good ✅"
Quickstart:
  🔥 Designing... → 💻 Building... → ♻️ Verifying features (2/5)...
  🚀 Done! https://xxx.trycloudflare.com

---

Mode 2: @[/leader] — Slow & Steady 🍲

You are the Chef. The Leader is your sous-chef. Full control at every step. Like a slow-cooked stew — takes time, but the result is production-grade quality.

You → Leader → Agents → Report back per phase → You approve → Next phase

Perfect for: Production apps, enterprise projects, or when quality matters more than speed.

How it works:

1. Tell the Leader what you want to build.
2. Leader analyzes, brainstorms, and presents a plan.
3. You approve the plan ✅
4. Leader auto-delegates to the right agents:

| Phase | Agent | What Happens | Mode | |-------|-------|-------------|------| | 📋 Planning | @[/planner] | PRD, user stories, timeline | Sequential | | 🏗️ Architecture + 🎨 Design | @[/architect] + @[/designer] | DB schema + UI/UX system | ⚡ PARALLEL | | 💻 Development | @[/frontend-dev] + @[/backend-dev] | Build frontend + backend simultaneously | ⚡ PARALLEL | | 🧪 QA & Fix | @[/qa-engineer] | Test → Find bugs → Fix → Retest | Sequential | | 🚀 Launch | @[/devops] + @[/security] + @[/seo] + @[/docs] | Deploy, audit, SEO, docs — all at once | ⚡ PARALLEL |

5. After each phase, Leader reports results and waits for your approval.
6. ⚡ Parallel Delegation: Architecture + Design run at the same time. Frontend + Backend run at the same time. Up to 4x faster.
7. QA Smart Loop: If a bug can't be fixed, Leader calls @[/meta-thinker] + @[/planner] to rethink the approach. Max 3 retries.

Example:

You: "Build me a Spotify clone with Next.js"
Leader: [Analyzes → Plans → Reports] "Here's the plan: 6 phases, 3 weeks..."
You: "Approved ✅"
Leader: [Auto-delegates to Planner → Architect → Designer → Dev → QA → Deploy]
        [Reports after each phase for your review]

---

Mode Comparison

| | 🍜 @[/quickstart] | 🍲 @[/leader] | |---|---|---| | Philosophy | Instant noodle — fast & easy | Slow-cooked — careful & thorough | | User involvement | Approve plan once | Approve each phase | | Parallel agents | ⚡ Yes | ⚡ Yes | | Completion verification | ♻️ Auto-loop (max 5) | Manual per phase | | Auto-deploy | ✅ Cloudflare Tunnel | Manual | | Template-first | ✅ Auto-detect | Manual | | Best for | MVPs, demos, non-tech users | Production apps, critical projects |

🧬 Team Profiles — Carry Your Style Across Projects (v2.9.0)

Problem: Every vibegravity init starts fresh — agents forget your coding style, tech preferences, and hard-won bug fixes. Solution: Persistent team profiles that learn from you automatically as you work, and carry that knowledge to every new project.

Quick Start

# Step 1: Create an empty team
vibegravity team create my-team

# Step 2: Init your project with that team
vibegravity init antigravity --team my-team

# Step 3: Just work normally with @[/leader] or @[/quickstart]
# → The agents AUTO-LEARN your coding style every time you:
#    ✅ Confirm a plan  → code scanned, DNA updated
#    ✅ Complete a phase → directives you said become rules
#    ✅ Fix a bug        → journal entry synced to team

That's it. No config files, no manual setup. The team learns passively.

How Auto-Learn Actually Works

The leader/quickstart agent acts as the observer. At each trigger point, it calls team_learner.py:

| Trigger | What Happens | Command Agent Runs | |---------|-------------|-------------------| | 🔵 Plan confirmed | Scans project source code → detects stack, naming style, architecture → generates/updates Team DNA | team_learner.py --scan-project . | | 🟢 Phase completed | Leader observed your directives (e.g. "write in English") → passes each one as a rule | team_learner.py --directive "write in English" | | 🔴 Bug fixed | Journal entry auto-syncs to team profile | team_manager.py save-back | | 🟡 Manual scan | You force-scan an existing codebase (optional) | vibegravity team scan my-team --path ./project | | 🟣 Manual learn | You ask team to learn from current project | vibegravity team learn |

Data Storage

All team data is stored globally (survives across projects):

~/.vibegravity/teams/<name>/
├── team.json               ← Main metadata (name, created_at, stack, code_style)
├── hot/                     ← ALWAYS loaded (~50 tokens)
│   ├── team.dna             ← 1-line DNA string (auto-generated)
│   └── top_rules.md         ← Auto-promoted rules (frequency ≥ 3)
├── warm/                    ← Loaded on demand (TF-IDF search)
│   ├── rules.json           ← ALL rules with dedup tracking
│   └── journal/
│       ├── index.json       ← Bug fix entries (title, tags, frequency)
│       └── entries/*.md     ← Full journal entries
└── cold/                    ← Archived (0 tokens unless requested)
    └── history/             ← Old DNA versions for rollback

When injected into a project (init --team), DNA and rules are copied to .agent/brain/:

.agent/brain/
├── team_dna.txt             ← DNA string for agents to read
├── team_rules.md            ← Hot rules (always applied)
├── team_rules/              ← Per-agent rules
│   ├── global.md
│   └── frontend-dev.md
└── team_meta.json           ← Which team, injected when

Rule Deduplication (Prevents File Bloat)

When a directive is added (manually or by the leader), the system checks if a similar rule already exists before creating a new one:

1. Normalize — strips filler words ("please", "always", "must", "should"...)
2. Stem — reduces suffixes: "documentation" → "document", "writing" → "writ"
3. Abbreviation expand — "docs" → "document", "ts" → "typescript"
4. Jaccard similarity — compares token overlap (threshold ≥ 50%)
5. If match found → increments frequency instead of creating duplicate
6. If frequency ≥ 3 → auto-promoted to Hot tier (loaded every session)

Example:
  Existing rule: "write docs in English"           (freq: 2)
  New directive:  "always write doc