Feat/add litellm provider

[RheagalFire]

Summary

Adds LiteLLM as a new LLM provider (oxy.LiteLLM), enabling access to 100+ LLM backends (OpenAI, Anthropic, Google, Azure, AWS Bedrock, Ollama, etc.) through a single unified interface.

Changes

• Feature: oxy.LiteLLM provider added

  New files:

  • oxygent/oxy/llms/litellm_llm.py -- LiteLLM(BaseLLM) class using litellm.acompletion()
  • tests/unittest/test_litellm_llm.py -- 8 unit tests
  • examples/llms/demo_litellm.py -- usage example

  Modified files:

  • oxygent/oxy/llms/__init__.py -- export LiteLLM
  • oxygent/oxy/__init__.py -- export LiteLLM
  • oxygent/oxy_factory.py -- register in OxyFactory._creators
  • docs/docs_en/introduction/agents/select-llm.md -- added LiteLLM to provider table + usage section

  Design decisions:

  • Extends BaseLLM directly (not RemoteLLM) because RemoteLLM's validator requires non-empty base_url, but LiteLLM routes by model prefix and doesn't always need one
  • drop_params=True by default -- silently drops provider-unsupported kwargs (e.g. strict:true on Anthropic, seed on Gemini)
  • api_key is optional -- when unset, LiteLLM falls back to provider-specific env vars (ANTHROPIC_API_KEY, OPENAI_API_KEY, etc.)
  • Lazy-imports litellm inside _execute() so the package stays optional
  • Streaming + reasoning_content (<think> tags) handled identically to OpenAILLM

Tests

Unit tests (8 passing):

tests/unittest/test_litellm_llm.py::test_streaming_success PASSED
tests/unittest/test_litellm_llm.py::test_non_streaming PASSED
tests/unittest/test_litellm_llm.py::test_api_key_forwarded PASSED
tests/unittest/test_litellm_llm.py::test_api_key_omitted_when_unset PASSED
tests/unittest/test_litellm_llm.py::test_base_url_forwarded PASSED
tests/unittest/test_litellm_llm.py::test_drop_params_default_true PASSED
tests/unittest/test_litellm_llm.py::test_llm_params_merged PASSED
tests/unittest/test_litellm_llm.py::test_factory_registration PASSED
8 passed in 0.85s

Existing tests -- no regressions (38 passed):

test_litellm_llm (8) + test_remote_llm (5) + test_http_llm (2) + test_base_llm (1) +
test_mock_llm (3) + test_oxy_factory (12) = 38 passed in 1.88s

Live e2e against Anthropic via LiteLLM:

# Non-streaming
State: OxyState.COMPLETED
Output: 4
Usage: input_tokens=20 output_tokens=5 model_name='anthropic/claude-sonnet-4-20250514' estimation_method='exact'

# Streaming
State: OxyState.COMPLETED
Output: Hello there, friend!
Usage: input_tokens=15 output_tokens=8 model_name='anthropic/claude-sonnet-4-20250514' estimation_method='exact'

Format check: ruff format clean on all new/modified files.

Example usage

import asyncio
from oxygent import MAS, oxy

oxy_space = [
    oxy.LiteLLM(
        name="default_llm",
        model_name="anthropic/claude-sonnet-4-20250514",
        # api_key="sk-...",  # or set ANTHROPIC_API_KEY env var
    ),
    oxy.ReActAgent(
        name="agent",
        llm_model="default_llm",
        system_prompt="You are a helpful assistant.",
    ),
]

async def main():
    async with MAS(oxy_space=oxy_space) as mas:
        result = await mas.call(
            callee="agent",
            arguments={"messages": [{"role": "user", "content": "What is 2 + 2?"}]},
        )
        print("Result:", result.output)

asyncio.run(main())

Checklist

• been self-reviewed.
  • Code is formatted
  • Tests added/updated
  • All CI checks passed
  • Documentation updated
• added documentation for new or modified features or behaviors.
• added new features, such as new agents, new flows, etc.
• added or updated version, license, or notice information.
• added or updated demo, integration tests, unit tests, or others.
• added or updated ui, or had changes in frontend.

[RheagalFire]

cc @JDOxygen

[RheagalFire]

@JDOxygen do you have any update on this PR?