feat: add LiteLLM as AI gateway engine

[RheagalFire]

Summary

• Adds LMMEngineLiteLLM as a new engine type across all Agent-S versions (s2, s2_5, s3)
• Gives users access to 100+ LLM providers (OpenAI, Anthropic, Azure, Bedrock, Vertex, Groq, Cohere, Ollama, etc.) through a single unified interface
• Users switch providers by changing the model string, no code changes needed
• Follows the exact same LMMEngine pattern used by the existing 8 engines

---

• What kind of change does this PR introduce? Feature, adds LiteLLM as a new LLM engine

• Why was this change needed? Agent-S currently has 8 individual LLM engines (OpenAI, Anthropic, Gemini, Azure, OpenRouter, vLLM, HuggingFace, Parasail), each maintained separately. LiteLLM provides a
  unified interface to 100+ LLM providers through a single completion() call. Users switch providers by changing the model string, no code changes needed.

• Other information: See sections below.

---

What Changed

File	Change
gui_agents/s2/core/engine.py	Added LMMEngineLiteLLM class
gui_agents/s2/core/mllm.py	Registered litellm engine type + isinstance check
gui_agents/s2_5/core/engine.py	Added LMMEngineLiteLLM class
gui_agents/s2_5/core/mllm.py	Registered litellm engine type + isinstance check
gui_agents/s3/core/engine.py	Added LMMEngineLiteLLM class
gui_agents/s3/core/mllm.py	Registered litellm engine type + isinstance check
setup.py	Added litellm>=1.60.0,<2.0.0
requirements.txt	Added litellm>=1.60.0,<2.0.0
tests/test_litellm_engine.py	15 unit tests across all versions

---

Usage

Quick Start

from gui_agents.s3.core.mllm import LMMAgent

agent = LMMAgent(
    engine_params={
        "engine_type": "litellm",
        "model": "anthropic/claude-3-haiku",  # any LiteLLM model string
    },                                                                                                                                                                                                             
    system_prompt="You are a helpful assistant.",
)                                                                                                                                                                                                                  
                
response = agent.get_response(user_message="What is 2+2?")
print(response)

Switching Providers: Just Change the Model String                                                                                                                                                                  
 
# OpenAI                                                                                                                                                                                                           
engine_params = {"engine_type": "litellm", "model": "openai/gpt-4o"}

# Anthropic
engine_params = {"engine_type": "litellm", "model": "anthropic/claude-3-haiku"}
                                                                                                                                                                                                                   
# AWS Bedrock
engine_params = {"engine_type": "litellm", "model": "bedrock/anthropic.claude-v2"}                                                                                                                                 
                
# Google Vertex AI
engine_params = {"engine_type": "litellm", "model": "vertex_ai/gemini-pro"}
                                                                                                                                                                                                                   
# Groq
engine_params = {"engine_type": "litellm", "model": "groq/llama3-70b-8192"}                                                                                                                                        
                
# Local Ollama
engine_params = {"engine_type": "litellm", "model": "ollama/llama3"}

Provider API keys are read from standard environment variables automatically (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.), or pass api_key in engine_params.

Any of the https://docs.litellm.ai/docs/providers works.

Direct Engine Usage (Without LMMAgent)

from gui_agents.s3.core.engine import LMMEngineLiteLLM
                                                                                                                                                                                                                   
engine = LMMEngineLiteLLM(model="openai/gpt-4o", api_key="sk-...")
                                                                                                                                                                                                                   
response = engine.generate(
    messages=[
        {"role": "system", "content": [{"type": "text", "text": "You are helpful."}]},
        {"role": "user", "content": [{"type": "text", "text": "Hello!"}]},                                                                                                                                         
    ],
    temperature=0.7,                                                                                                                                                                                               
    max_new_tokens=100,
)                                                                                                                                                                                                                  
print(response)

---

Testing

Unit Tests (15/15 Passing)

  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_generate_returns_content PASSED
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_generate_passes_drop_params PASSED
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_generate_passes_model PASSED                                                                                                                               
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_generate_forwards_api_key PASSED                                                                                                                           
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_generate_omits_api_key_when_none PASSED                                                                                                                    
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_instance_temperature_overrides_param PASSED                                                                                                                
  tests/test_litellm_engine.py::TestS3EngineLiteLLM::test_default_max_tokens PASSED
  tests/test_litellm_engine.py::TestS2EngineLiteLLM::test_generate_returns_content PASSED                                                                                                                            
  tests/test_litellm_engine.py::TestS2EngineLiteLLM::test_drop_params_default PASSED
  tests/test_litellm_engine.py::TestS25EngineLiteLLM::test_generate_returns_content PASSED                                                                                                                           
  tests/test_litellm_engine.py::TestLMMAgentRegistration::test_s3_agent_creates_litellm_engine PASSED
  tests/test_litellm_engine.py::TestLMMAgentRegistration::test_s3_agent_get_response PASSED                                                                                                                          
  tests/test_litellm_engine.py::TestLMMAgentRegistration::test_s2_agent_creates_litellm_engine PASSED
  tests/test_litellm_engine.py::TestLMMAgentRegistration::test_s2_5_agent_creates_litellm_engine PASSED                                                                                                              
  tests/test_litellm_engine.py::TestLMMAgentRegistration::test_unsupported_engine_raises PASSED                                                                                                                      
  ============================== 15 passed in 0.99s ==============================                                                                                                                                   

Live E2E (Azure AI Foundry, Claude Sonnet via LiteLLM)

Model: azure_ai/claude-sonnet-4-6

LMMEngineLiteLLM.generate() -> "4"
LMMAgent.get_response() -> "1, 2, 3"

---

Risk / Compatibility

• Additive only, no existing engine code changed
• drop_params=True set by default so provider-unsupported kwargs are silently dropped
• LiteLLM lazily imported inside generate(), users who don't use this engine are unaffected
• Added to all 3 versions (s2, s2_5, s3) for consistency

[RheagalFire]

cc @alckasoc @Richard-Simular

[RheagalFire]

@eric-xw do you have any update on this PR?