bob-runtime

Runtime orchestration layer for the Bob Agent Framework.

Overview

bob-runtime provides the orchestration layer that coordinates agent execution:

• Scheduler: Finite state machine for agent turn execution
• Action Parser: Parses LLM responses into structured actions
• Prompt Builder: Constructs prompts with tool definitions and context
• Composite Tool Port: Aggregates multiple tool sources

This crate depends only on bob-core port traits — never on concrete adapters.

Architecture

┌─────────────────────────────────────────┐
│         AgentRuntime (trait)            │
├─────────────────────────────────────────┤
│  ┌──────────┐  ┌──────────┐  ┌───────┐ │
│  │Scheduler │→ │Prompt    │→ │Action │ │
│  │  FSM     │  │Builder   │  │Parser │ │
│  └──────────┘  └──────────┘  └───────┘ │
└─────────────────────────────────────────┘
         ↓ uses ports from bob-core

Usage

Add this to your Cargo.toml:

[dependencies]
bob-runtime = "0.2.1"

Example: Creating a Runtime

use bob_runtime::{AgentBootstrap, AgentRuntime, RuntimeBuilder};
use bob_core::{
    ports::{LlmPort, ToolPort, SessionStore, EventSink},
    types::TurnPolicy,
};
use std::sync::Arc;

fn create_runtime(
    llm: Arc<dyn LlmPort>,
    tools: Arc<dyn ToolPort>,
    store: Arc<dyn SessionStore>,
    events: Arc<dyn EventSink>,
) -> Result<Arc<dyn AgentRuntime>, bob_core::error::AgentError> {
    RuntimeBuilder::new()
        .with_llm(llm)
        .with_tools(tools)
        .with_store(store)
        .with_events(events)
        .with_default_model("openai:gpt-4o-mini")
        .with_policy(TurnPolicy::default())
        .build()
}

Running an Agent Turn

use bob_core::types::AgentRequest;

let request = AgentRequest {
    input: "Summarize this document in three bullet points".to_string(),
    session_id: "session-123".to_string(),
    model: None,
    context: Default::default(),
    cancel_token: None,
};

let result = runtime.run(request).await?;

Features

• Finite State Machine: Robust turn execution with state tracking
• Streaming Support: Real-time event streaming via run_stream()
• Tool Composition: Aggregate multiple MCP servers or tool sources
• Turn Policies: Configurable limits for steps, timeouts, and retries
• Health Monitoring: Built-in health check endpoints
• AgentLoop Facade: Unified slash-command + runtime orchestration for interactive clients

Modules

• scheduler: Core FSM implementation for agent execution
• action: Action types and parser for LLM responses
• prompt: Prompt construction and tool definition formatting
• composite: Multi-source tool aggregation

Documentation

Full API documentation is available at docs.rs/bob-runtime.

Related Crates

• bob-core - Domain types and ports
• bob-adapters - Concrete implementations
• bob-cli - CLI application

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.