Praxis Graph - Core Runtime

The core runtime for Praxis AI agents, implementing the Graph orchestrator with Node abstraction and React agent pattern.

Features

• Node Abstraction: LLMNode, ToolNode - composable units of computation
• Router: SimpleRouter implementing React pattern (LLM → Tool → LLM → END)
• Event Streaming: Real-time token-by-token streaming via bounded channels
• Tool Execution: Mock tools (calculator, weather, search) with graceful error handling
• Guardrails: Max iterations, timeouts, cancellation support

Architecture

Graph Orchestrator
    ↓
LLMNode → Router → ToolNode → Router → LLMNode → END
    ↓                ↓
  Events          Events
    ↓                ↓
  Client          Client

Quick Start

Prerequisites

export OPENAI_API_KEY=your_openai_api_key_here

Run Interactive Demo

cargo run --example react_loop

This will start an interactive CLI where you can chat with the React agent.

Example Interactions

Simple Question (no tools):

You: What is the capital of France?
Assistant: The capital of France is Paris.

Using Tools:

You: What's the weather like in San Francisco?
Assistant:
💭 Reasoning: I should use the get_weather tool...
🔧 Calling tool: get_weather ({"location": "San Francisco, CA"})
✓ Tool result (112ms): {"temperature": 22, "condition": "sunny", "location": "San Francisco"}
The weather in San Francisco is currently sunny with a temperature of 22°C.

Calculator Example:

You: What's 1234 multiplied by 5678?
Assistant:
💭 Reasoning: I need to use the calculator tool...
🔧 Calling tool: calculator ({"expression": "1234 * 5678"})
✓ Tool result (105ms): {"result": 42, "expression": "1234 * 5678"}
The result of 1234 multiplied by 5678 is 7,006,652.

Testing

Build all crates

cargo build --all

Run tests

cargo test --all

Check lints

cargo clippy --all

Code Structure

src/
├── node.rs              # Node trait, NodeType enum
├── router.rs            # Router trait, SimpleRouter
├── graph.rs             # Graph orchestrator
├── tools.rs             # ToolExecutor trait, MockToolExecutor
└── nodes/
    ├── llm_node.rs      # LLM interaction node
    └── tool_node.rs     # Tool execution node

Next Steps

• Add real MCP integration (replace MockToolExecutor)
• Add persistence layer (MongoDB)
• Add more sophisticated routers (conditional, parallel)
• Add observability (metrics, tracing)
• Performance benchmarks

See Also

• Architecture Documentation
• Plan & Philosophy
• praxis-types - Shared type definitions
• praxis-llm - LLM client implementations