An autonomous AI Engineering orchestrator CLI. Trully Model-Agnostic: supports Gemini, Claude, and any OpenAI-compatible provider (DeepSeek, GLM, Groq, Ollama, etc.).

🧠 The "Three Brains" Architecture (v1.6.1+)

Ceobe now utilizes a Three Brain Pattern to prevent AI hallucinations and eliminate self-evaluation bias:

Planner (The Architect): Reads your request and generates a strict Blueprint (BRD, Design, Architecture). Best paired with highly intelligent models (e.g., Claude 4.5 Sonnet or GPT-4o).
QA Auditor (The Reviewer): Strictly reviews the Blueprint against the core rules to ensure architectural soundness.
Executor (The Programmer): Follows the validated blueprint to write the code. Best paired with fast/cheap models (e.g., Gemini Exp, GLM, or Llama 3).

⚡ Features (v1.12.0 V3 Engine)

🔥 Hash Convergence Guard (V3 New): Infinite-loop protection! Supervisor tracks document state via SHA-256 hashes and gracefully skips QA audits if no documents were changed, saving thousands of tokens and eliminating loop traps.
🛡️ Token Bleed Protection (V3 New): Executor now detects max_tokens exhaustion and breaks loops safely after 3 retries, preventing run-away token burn.
🎯 Dynamic Tool Injection (V3 New): The system intelligently parses your task and only injects the required LLM skills/tools into the prompt, resulting in massive Input Token reduction.
✨ Prefix Caching Support (V3 New): Fully typed NormalizedContentBlock support enables 100% compatibility with Anthropic's cache_control: true feature, dramatically speeding up audits and reducing prompt token costs.
🌐 Polyglot Architecture Check (v1.12.0 New): Automatically detects and verifies compilations/tests across TypeScript, Go, Rust, PHP/Laravel, and Python after each execution wave.
🛡️ Telegram Interactive HITL (v1.12.0 New): Safe mode mode ask now streams confirmation requests directly to your Telegram with interactive Inline Buttons (Approve/Reject).
✨ Prefix Caching Support (V3 New): Fully typed NormalizedContentBlock support enables 100% compatibility with Anthropic's cache_control: true feature, dramatically speeding up audits and reducing prompt token costs.
🔐 Concurrent File-Lock Safety: Rock-solid thread safety for multi-agent I/O. File writes are protected with proper-lockfile and strict mutex-like write locks, eliminating corruption even when 10 AI subagents write simultaneously!
🌐 Strict Context Isolation (New): Fully supports Multi-Tenancy for the Telegram Daemon. Each Telegram user session has strict project boundaries and dynamic path resolution (executionContext), eliminating path traversal vulnerabilities.
📱 Telegram Remote Daemon: Command and monitor your AI orchestration fully from your mobile phone! Includes multi-project session management (/cd, /addproject) and remote file viewing (/read).
🛡️ Interactive Human-In-The-Loop (HITL): Approve or reject dangerous commands and file deletions directly via Telegram Inline Buttons.
🧩 Dynamic Plugin System: Inject custom behaviors and tools into the agent by simply dropping .ts scripts into .ceobe/plugins/. No core modifications needed!
⚡ Multi-Agent Parallel Execution: The Supervisor intelligently groups independent tasks into "Waves" and executes them concurrently for blazingly fast delivery.
⏪ Git Auto-Snapshot & Rollback: Automatically creates snapshots before AI execution. If post-execution auto-healing fails after max retries, Ceobe automatically rolls back the codebase to safety.
💰 Live Cost Tracking: Real-time token usage and API cost monitoring per session (/cost).
🌐 Provider Agnostic: Switch models instantly (Anthropic, Gemini, OpenAI, GLM).
👁️ Multi-Modal Support: Pass .md files or UI Mockup Images (.png, .webp) as your prompt!

📦 Installation

npm install -g ceobe-mastery-cli

🛠️ Configuration (CLI Wizard)

Configure your API keys and roles easily using the built-in CLI:

# 1. Set your API Keys
ceobe key set glm <YOUR_GLM_API_KEY>
ceobe key set gemini <YOUR_GEMINI_API_KEY>

# 2. Assign The Three Brains
ceobe key set planner-provider anthropic
ceobe key set qa-provider gemini
ceobe key set executor-provider glm

# 3. Verify your setup
ceobe doctor

📖 Usage Scenarios

1. Greenfield (New Projects)

Start a new project from scratch:

ceobe auto "Build a REST API with Go, Fiber, and PostgreSQL"

2. Brownfield (Adding Features)

Add features to an existing codebase without breaking it:

ceobe auto --feature "Add Stripe payment gateway integration to the checkout module"

3. Data-Driven (Documents & UI Mockups)

Pass a PRD or an image mockup:

# From a document
ceobe auto --file requirements.md

# From an image mockup
ceobe auto --file dashboard-mockup.png "Make this using React and Tailwind"

4. Human-in-the-Loop (Safe Mode)

Want to approve actions before Ceobe executes them?

# Set mode globally
ceobe mode ask

# Or bypass temporarily for one run
ceobe auto "Build a portfolio" --ask

5. Telegram Daemon (Remote Mobile Control)

Run Ceobe on a server/laptop and control it securely from your phone!

# 1. Setup your keys
ceobe key set telegram-token <YOUR_BOT_TOKEN>
ceobe key set telegram-allowed-users <YOUR_TELEGRAM_USER_ID>

# 2. Start the Daemon
ceobe daemon --telegram

In Telegram, simply type:

/projects and /addproject <name> <path> to manage workspaces.
/cd <name> to seamlessly switch contexts.
/mode <ask|autonomous> or just /ask and /auto to toggle between HITL confirmation and full autonomy.
/cost and /status to track execution metrics.
/logs to tail the recent terminal logs.
/cancel to clear the task queue.
Use /read <file> to fetch the generated design document or any source code.
Send a prompt: "Tolong buatkan halaman login..."
Approve/Reject dangerous commands (mode ask) directly using Interactive Buttons!