AG KIT

AI Agent templates with Skills, Agents, and Workflows — featuring Coordinator Mode, Persistent Memory, and Context Compression.

🇻🇳 Tiếng Việt (Vietnamese Version)

---

⚡ Quick Start

Install and initialize AG Kit to inject the .agents/ configuration folder directly into your local project.

Method 1: On-demand Execution (Recommended)

npx @vudovn/ag-kit init

Method 2: Global Installation

npm install -g @vudovn/ag-kit
ag-kit init

---

🛡️ Safe Updates and Rollback

AG Kit updates are merge-aware by default. Files you created or changed locally are preserved, while clean managed files receive upstream updates.

ag-kit update --dry-run          # Preview the exact update plan
ag-kit update                    # Safe merge with backup and conflict report
ag-kit update --strategy replace # Explicit full replacement, with backup
ag-kit rollback                  # Restore the newest pre-update backup

Update metadata is stored in .agents/.ag-kit/. Backups are stored outside the toolkit at .ag-kit-backups/, so an interrupted or conflicting update can be reviewed or rolled back safely.

---

🌍 Global Shared Setup (Symlinks)

If you work across multiple repositories and want to avoid duplicating the .agents/ folder in every single project, you can centralize AG Kit and use symbolic links.

1. Install centrally (e.g., to a global folder like ~/.ag-kit):

   mkdir -p ~/.ag-kit && cd ~/.ag-kit
   npx @vudovn/ag-kit init

2. Link it locally from inside your project root:

   • macOS / Linux:

     ln -s ~/.ag-kit/.agents .agents

   • Windows (CMD - Run as Administrator):

     mklink /D .agents "%USERPROFILE%\.ag-kit\.agents"

   • Windows (PowerShell - Run as Administrator):

     New-Item -ItemType SymbolicLink -Path ".agents" -Target "$env:USERPROFILE\.ag-kit\.agents"

---

⚠️ Important Note on .gitignore

If you are using AI-native code editors (like Antigravity), adding the .agents/ directory to .gitignore will prevent the editor's language server from indexing the workflows, which disables autocomplete for slash commands (e.g. /plan, /debug).

Recommended Solution:

To keep .agents/ out of your remote repository without losing editor integration:

1. Ensure .agents/ is NOT listed in your project's .gitignore.
2. Add .agents/ to your local Git exclude file: .git/info/exclude instead.

---

📦 What's Included

AG Kit packages domain-specific knowledge, specialized agent personas, and automated workflows optimized for modern AI coding tools.

Component	Count	Description
Agents	20	Specialist AI personas (Frontend, Backend, Security, PM, QA, etc.)
Skills	47	Domain-specific context modules with conditional loading rules
Workflows	13	Pre-configured interactive developer procedures (slash commands)

---

🔐 Versioned Agent Components

Starting with 2026.7.18, every agent, skill, workflow, and rule has a strict SemVer contract. The .agents/manifest.json registry and .agents/manifest.lock.json integrity lock make dependencies machine-readable and prevent documentation/configuration drift.

npm run generate:agents # regenerate registry, lock, and dependency graph
npm run check:agents    # non-mutating release/CI verification

The generated .agents/DEPENDENCY_GRAPH.md shows workflow → agent → skill relationships. Official runtime support remains Gemini CLI and Google Antigravity; the registry format is portable by design.

---

🛠️ Usage

1. Zero-Setup Agent Auto-Routing

You don't need to manually orchestrate agents. The system silently classifies your request, auto-routes to the best domain experts, and applies their rules instantly:

You: "Add JWT authentication to the login API"
Agent: Applying @security-auditor + @backend-specialist...

You: "Align the checkout button to the center and fix dark mode"
Agent: Using @frontend-specialist...

2. Interactive Workflows (Slash Commands)

Execute structured development workflows by typing slash commands in your AI agent chat:

Command	Description
/brainstorm	Structured exploration of options and architecture before coding
/coordinate	Orchestrate multiple agents in parallel for complex reviews
/create	Create new features or full applications from scratch
/debug	Activate evidence-based systematic debugging
/deploy	Execute pre-flight checks and deploy to production
/enhance	Safely add or update features in an existing codebase
/plan	Generate a structured implementation plan and checklist
/preview	Start, stop, or check status of local preview servers
/remember	Save custom project conventions to persistent memory
/status	Generate a clear status report of the agent's progress
/test	Generate and execute comprehensive tests
/verify	Prove code works via execution rather than simple inspection

---

🧠 Core Architectural Concepts

AG Kit is built on production-tested agentic design patterns designed to reduce token usage by 13% to 33% while yielding higher output quality:

• Coordinator Mode: Multi-agent orchestration with parallel workers and synthesis, avoiding expensive sequential retries.
• Persistent Memory: A 4-type taxonomy memory engine index (MEMORY.md) to prevent re-explaining project guidelines across sessions.
• Context Compression: Automated summarization and micro-compaction routines to prevent context degradation in long-lived sessions.
• Conditional Skill Loading: Context-aware loading of rules via custom frontmatter, preventing your context window from bloating with idle instructions.

---

📚 References & Attribution

AG Kit represents an original implementation of markdown-based prompt and rules engineering. It was built by analyzing production agent patterns to distill core agentic behaviors:

• No proprietary code or files were copied.
• All templates, rules, and scripts are rewritten as original, open-source implementations under the MIT license.

---

☕ Support the Project

If AG Kit has made your AI programming sessions more productive, consider supporting the project:

Vietnamese Bank (MBBank QR):

---

CA: Gjpatn3d24dCRhUng7F37K6xJba4R8SDBC18xs1Apump

📄 License

Released under the MIT License © Vudovn.