# Ajarbot

A lightweight, cost-effective AI agent framework for building proactive bots with Claude and other LLMs. Features intelligent memory management, multi-platform messaging support, and efficient monitoring with the Pulse & Brain architecture.

## Table of Contents

- [Features](#features)
- [Quick Start](#quick-start)
- [Installation](#installation)
- [Core Concepts](#core-concepts)
- [Usage Examples](#usage-examples)
- [Architecture](#architecture)
- [Documentation](#documentation)
- [License](#license)

## Features

- **Flexible Claude Integration**: Use Pro subscription OR pay-per-token API via Agent SDK (no server needed)
- **Cost-Optimized AI**: Default Haiku 4.5 model (12x cheaper), auto-caching on Sonnet (90% savings), dynamic model switching
- **Smart Memory System**: SQLite-based memory with automatic context retrieval and hybrid vector search
- **Multi-Platform Adapters**: Run on Slack, Telegram, and more simultaneously
- **15 Integrated Tools**: File ops, shell commands, Gmail, Google Calendar, Contacts
- **Pulse & Brain Monitoring**: 92% cost savings with intelligent conditional monitoring (recommended)
- **Task Scheduling**: Cron-like scheduled tasks with flexible cadences
- **Multi-LLM Support**: Claude (Anthropic) primary, GLM (z.ai) optional

## Quick Start

### Option 1: Agent SDK (Recommended - Uses Pro Subscription)

```bash
# Clone and install
git clone https://vulcan.apophisnetworking.net/jramos/ajarbot.git
cd ajarbot
pip install -r requirements.txt

# Authenticate with Claude CLI (one-time setup)
claude auth login

# Configure adapters
cp .env.example .env
cp config/adapters.example.yaml config/adapters.local.yaml
# Edit config/adapters.local.yaml with your Slack/Telegram tokens

# Run
run.bat              # Windows
python ajarbot.py    # Linux/Mac
```

### Option 2: API Mode (Pay-per-token)

```bash
# Clone and install
git clone https://vulcan.apophisnetworking.net/jramos/ajarbot.git
cd ajarbot
pip install -r requirements.txt

# Configure
cp .env.example .env
# Edit .env and add:
#   AJARBOT_LLM_MODE=api
#   ANTHROPIC_API_KEY=sk-ant-...

# Run
run.bat              # Windows
python ajarbot.py    # Linux/Mac
```

**See [CLAUDE_CODE_SETUP.md](CLAUDE_CODE_SETUP.md)** for detailed setup and mode comparison.

### Model Switching Commands

Send these to your bot:
- `/haiku` - Fast, cheap (default)
- `/sonnet` - Smart, caching enabled (auto 90% cost savings)
- `/status` - Check current model and settings

## Core Concepts

### Agent

The central component that handles LLM interactions with automatic context loading:

- Loads personality from `SOUL.md`
- Retrieves user preferences from `users/{username}.md`
- Searches relevant memory chunks
- Maintains conversation history

```python
from agent import Agent

agent = Agent(provider="claude")
response = agent.chat("Tell me about Python", username="alice")
```

### Memory System

SQLite-based memory with full-text search:

```python
# Write to memory
agent.memory.write_memory("Completed task X", daily=True)

# Update user preferences
agent.memory.update_user("alice", "## Preference\n- Likes Python")

# Search memory
results = agent.memory.search("python")
```

### Task Management

Built-in task tracking:

```python
# Add task
task_id = agent.memory.add_task(
    "Implement API endpoint",
    "Details: REST API for user auth"
)

# Update status
agent.memory.update_task(task_id, status="in_progress")

# Get tasks
pending = agent.memory.get_tasks(status="pending")
```

### Pulse & Brain Architecture

The most cost-effective way to run proactive monitoring:

```python
from agent import Agent
from pulse_brain import PulseBrain

agent = Agent(provider="claude", enable_heartbeat=False)

# Pulse runs pure Python checks (zero cost)
# Brain only invoked when needed (92% cost savings)
pb = PulseBrain(agent, pulse_interval=60)
pb.start()
```

**Cost comparison:**
- Traditional polling: ~$0.48/day
- Pulse & Brain: ~$0.04/day
- **Savings: 92%**

### Multi-Platform Adapters

Run your bot on multiple messaging platforms simultaneously:

```python
from adapters.runtime import AdapterRuntime
from adapters.slack.adapter import SlackAdapter
from adapters.telegram.adapter import TelegramAdapter

runtime = AdapterRuntime(agent)
runtime.add_adapter(slack_adapter)
runtime.add_adapter(telegram_adapter)

await runtime.start()
```

### Task Scheduling

Cron-like scheduled tasks:

```python
from scheduled_tasks import TaskScheduler, ScheduledTask

scheduler = TaskScheduler(agent)

task = ScheduledTask(
    "morning-brief",
    "What are today's priorities?",
    schedule="08:00",
    username="alice"
)

scheduler.add_task(task)
scheduler.start()
```

## Usage Examples

### Basic Chat with Memory

```python
from agent import Agent

agent = Agent(provider="claude")

# First conversation
agent.chat("I'm working on a Python API", username="bob")

# Later conversation - agent remembers
response = agent.chat("How's the API coming?", username="bob")
# Agent retrieves context about Bob's Python API work
```

### Model Switching

```python
agent = Agent(provider="claude")

# Use Claude for complex reasoning
response = agent.chat("Explain quantum computing")

# Switch to GLM for faster responses
agent.switch_model("glm")
response = agent.chat("What's 2+2?")
```

### Custom Pulse Checks

```python
from pulse_brain import PulseBrain, PulseCheck, BrainTask, CheckType

def check_disk_space():
    import shutil
    usage = shutil.disk_usage("/")
    percent = (usage.used / usage.total) * 100
    return {
        "status": "error" if percent > 90 else "ok",
        "percent": percent
    }

pulse_check = PulseCheck("disk", check_disk_space, interval_seconds=300)

brain_task = BrainTask(
    name="disk-advisor",
    check_type=CheckType.CONDITIONAL,
    prompt_template="Disk is {percent:.1f}% full. Suggest cleanup.",
    condition_func=lambda data: data.get("percent", 0) > 90
)

pb = PulseBrain(agent)
pb.add_pulse_check(pulse_check)
pb.add_brain_task(brain_task)
pb.start()
```

### Skills from Messaging Platforms

```python
from adapters.skill_integration import SkillInvoker

skill_invoker = SkillInvoker()

def skill_preprocessor(message):
    if message.text.startswith("/"):
        parts = message.text.split(maxsplit=1)
        skill_name = parts[0][1:]
        args = parts[1] if len(parts) > 1 else ""

        if skill_name in skill_invoker.list_available_skills():
            skill_info = skill_invoker.get_skill_info(skill_name)
            message.text = skill_info["body"].replace("$ARGUMENTS", args)

    return message

runtime.add_preprocessor(skill_preprocessor)
```

Then from Slack/Telegram:
```
@bot /code-review adapters/slack/adapter.py
@bot /deploy --env prod --version v1.2.3
```

## Architecture

```
┌──────────────────────────────────────────────────────┐
│                    Ajarbot Core                      │
│                                                      │
│  ┌────────────┐  ┌────────────┐  ┌──────────────┐  │
│  │   Agent    │  │   Memory   │  │ LLM Interface│  │
│  │            │──│   System   │──│(Claude/GLM)  │  │
│  └─────┬──────┘  └────────────┘  └──────────────┘  │
│        │                                            │
│        │         ┌────────────────┐                 │
│        └─────────│  Pulse & Brain │                 │
│                  │   Monitoring   │                 │
│                  └────────────────┘                 │
└──────────────────────┬───────────────────────────────┘
                       │
         ┌─────────────┴─────────────┐
         │                           │
    ┌────▼─────┐              ┌──────▼──────┐
    │  Slack   │              │  Telegram   │
    │ Adapter  │              │   Adapter   │
    └──────────┘              └─────────────┘
```

### Key Components

1. **agent.py** - Main agent class with automatic context loading
2. **memory_system.py** - SQLite-based memory with FTS5 search
3. **llm_interface.py** - Unified interface for Claude and GLM
4. **pulse_brain.py** - Cost-effective monitoring system
5. **scheduled_tasks.py** - Cron-like task scheduler
6. **adapters/** - Multi-platform messaging support
   - **base.py** - Abstract adapter interface
   - **runtime.py** - Message routing and processing
   - **slack/**, **telegram/** - Platform implementations
7. **config/** - Configuration management

## Documentation

Comprehensive documentation is available in the [docs/](docs/) directory:

### Getting Started
- [Quick Start Guide](docs/QUICKSTART.md) - 30-second setup and basic usage
- [Windows 11 Deployment](docs/WINDOWS_DEPLOYMENT.md) - Complete Windows deployment and testing guide
- [Pulse & Brain Quick Start](docs/QUICK_START_PULSE.md) - Efficient monitoring setup

### Core Systems
- [Pulse & Brain Architecture](docs/PULSE_BRAIN.md) - Cost-effective monitoring (92% savings)
- [Memory System](docs/README_MEMORY.md) - SQLite-based memory management
- [Scheduled Tasks](docs/SCHEDULED_TASKS.md) - Cron-like task scheduling
- [Heartbeat Hooks](docs/HEARTBEAT_HOOKS.md) - Proactive health monitoring

### Platform Integration
- [Adapters Guide](docs/README_ADAPTERS.md) - Multi-platform messaging (Slack, Telegram)
- [Skills Integration](docs/SKILLS_INTEGRATION.md) - Claude Code skills from messaging platforms

### Advanced Topics
- [Control & Configuration](docs/CONTROL_AND_CONFIGURATION.md) - Configuration management
- [Monitoring Comparison](docs/MONITORING_COMPARISON.md) - Choosing the right monitoring approach

## Project Structure

```
ajarbot/
├── agent.py                      # Main agent class
├── memory_system.py              # Memory management
├── llm_interface.py              # LLM provider interface
├── pulse_brain.py                # Pulse & Brain monitoring
├── scheduled_tasks.py            # Task scheduler
├── heartbeat.py                  # Legacy heartbeat system
├── hooks.py                      # Event hooks
├── bot_runner.py                 # Multi-platform bot runner
├── adapters/                     # Platform adapters
│   ├── base.py                   # Base adapter interface
│   ├── runtime.py                # Adapter runtime
│   ├── skill_integration.py      # Skills system
│   ├── slack/                    # Slack adapter
│   └── telegram/                 # Telegram adapter
├── config/                       # Configuration files
│   ├── config_loader.py
│   └── adapters.yaml
├── docs/                         # Documentation
├── memory_workspace/             # Memory storage
└── examples/                     # Example scripts
    ├── example_usage.py
    ├── example_bot_with_pulse_brain.py
    ├── example_bot_with_scheduler.py
    └── example_bot_with_skills.py
```

## Configuration

### Environment Variables

```bash
# LLM Mode (optional - defaults to agent-sdk)
export AJARBOT_LLM_MODE="agent-sdk"  # Use Pro subscription
# OR
export AJARBOT_LLM_MODE="api"        # Use pay-per-token API

# Required for API mode only
export ANTHROPIC_API_KEY="sk-ant-..."

# Optional: Alternative LLM
export GLM_API_KEY="..."

# Adapter credentials (stored in config/adapters.local.yaml)
export AJARBOT_SLACK_BOT_TOKEN="xoxb-..."
export AJARBOT_SLACK_APP_TOKEN="xapp-..."
export AJARBOT_TELEGRAM_BOT_TOKEN="123456:ABC..."
```

### Adapter Configuration

Generate configuration template:
```bash
python bot_runner.py --init
```

Edit `config/adapters.local.yaml`:
```yaml
adapters:
  slack:
    enabled: true
    credentials:
      bot_token: "xoxb-..."
      app_token: "xapp-..."

  telegram:
    enabled: true
    credentials:
      bot_token: "123456:ABC..."
```

## Testing

Run tests to verify installation:

```bash
# Test memory system
python test_skills.py

# Test task scheduler
python test_scheduler.py
```

## Contributing

Contributions are welcome! Please:

1. Follow PEP 8 style guidelines
2. Add tests for new features
3. Update documentation
4. Keep code concise and maintainable

## Credits

- Adapter architecture inspired by [OpenClaw](https://github.com/chloebt/openclaw)
- Built with [Anthropic Claude](https://www.anthropic.com/claude)
- Alternative LLM support via [z.ai](https://z.ai)

## License

MIT License - See LICENSE file for details

---

**Need Help?**
- Check the [documentation](docs/)
- Review the [examples](example_usage.py)
- Open an issue on GitHub