Ray Agent

AI Agent for Ray.

Principles

• Reflection: Iterative reasoning within and outside the LLM.
• Tool Calling: Diverse integrations with Ray environments.
• Planning: Planning through the reasoning system, with optional support for todo tool.
• Multi-agent Collaboration: Coordinated workflows among specialized agents.

Quick Start

# Install
uv sync

# Set API key
export OPENAI_API_KEY="sk-..."

# Run
uv run ray-agent

# Commands
/analyze @train.py          # Predict GPU requirements
/refactor @app.py           # Fix Ray anti-patterns with suggestions
/deploy @serve.py           # Generate KubeRay manifests
/debug discover             # Debug Ray workloads on GKE and Anyscale
/ask How to scale Ray?      # Search Ray knowledge base across docs, gihub, slack
/cluster show pods          # Manage k8s and Anyscale clusters

Features

• OpenAI Responses API: Native tool-calling with structured outputs
• Multi-source Input: Files (@file.py), inline code
• Specialized Agents: Analyze, Refactor, Deploy, Debug, Ask, Cluster
• Production Ready: Connection pooling, retry logic, async throughout

Installation

git clone https://github.com/pradeepiyer/ray-agent.git
cd ray-agent
uv sync

Configuration

Required:

export OPENAI_API_KEY="sk-..."

Optional:

export ANYSCALE_API_KEY="esecret_..."  # Anyscale platform
export SLACK_API_KEY="xoxb-..."        # Slack search
export PYLON_API_KEY="xoxb-..."        # Pylon search
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"  # GCP

Development

make help
General:
  test         - Run test suite with coverage (excludes integration)
  test-integration - Run integration tests (requires API keys)
  install      - Install ray-agent in development mode
  clean        - Remove build artifacts and cache files
  lint         - Run linter and fix issues with ruff and mypy
  sync-deps    - Sync uv dependencies and lock file
  console      - Run interactive chat console (default mode)
  serve        - Start HTTP server with REST and MCP protocols
  help         - Show this help message

Evaluation:
  eval         - Run full evaluation suite (all agents)
  eval-ask     - Run Ask evaluation
  eval-ask-quick - Quick Ask eval, single query
  eval-analyze - Run WorkloadAnalysis evaluation only
  eval-refactor - Run Refactor evaluation only
  eval-refactor-quick - Quick Refactor eval (subset of tests)
  eval-debug   - Run Debug evaluation only
  eval-debug-quick - Quick Debug eval (first scenario)
  eval-debug-record - Record a debug scenario for evaluation
  eval-quick   - Run quick evaluation (subset of tests)

Package Distribution:
  build        - Build wheel package (default)
  build-wheel  - Build wheel package only
  build-sdist  - Build source distribution only
  build-all    - Build both wheel and source distribution
  package-info - Show package information and files

Dependency Management:
  lock-deps    - Update dependency lock file
  check-deps   - Check if dependencies are up to date
  ci           - Run the same checks as GitHub CI (lint, format check, pyright, tests)

Documentation

Core Components

• Agents - Core orchestration
• Console - CLI interface
• API - Agent API, REST and MCP
• Config - Configuration system
• Prompts - YAML templates
• Clients - API clients

Agent Documentation

• Analyze Agent - GPU requirement prediction
• Refactor Agent - Ray anti-pattern detection and automatic refactoring
• Debug Agent - Ray workload debugging
• Ask Agent - Ray knowledge base search
• Cluster Agent - K8s and Anyscale cluster management
• Deploy Agent - KubeRay manifest generation (In Progress)

Requirements

• Python 3.13+
• uv package manager
• OpenAI API key
• kubectl (for cluster operations)
• ripgrep (for source code analysis)