feat: add [ LiteLLM as AI gateway ] provider

[RheagalFire]

Features

• Add LiteLLM provider in metagpt/provider/litellm_api.py implementing BaseLLM against the litellm.acompletion() API
• Add LLMType.LITELLM = "litellm" to metagpt/configs/llm_config.py and register in metagpt/provider/init.py
• Add tests/metagpt/provider/test_litellm_api.py covering completion, streaming, and config kwarg mapping
• Add litellm>=1.55,<1.65 to requirements.txt (version range chosen to resolve cleanly against MetaGPT's openai~=1.64 pin).
• Add aask_code / _achat_completion_function / get_choice_function_arguments so Engineer / QAEngineer / DataAnalyst roles work on LiteLLM-routed models.

Feature Docs

Users set api_type: litellm in config2.yaml and route to 100+ providers (OpenAI, Anthropic, Bedrock, Vertex, Ollama, OpenRouter, Groq, DeepSeek, etc.) via a single interface:

llm:
api_type: "litellm"
model: "anthropic/claude-3-5-sonnet-20241022"
api_key: "sk-..."

LiteLLM routes based on the model-name prefix. Full provider list: https://docs.litellm.ai/docs/providers

Influence

• Unblocks MetaGPT on any LiteLLM-supported provider without writing a new provider class per provider, reduces maintenance burden of per-provider files going forward
• Additive only existing providers (OpenAI, Anthropic, Bedrock, etc.) are untouched
• One new required dependency (litellm>=1.55,<1.65). Range pinned to stay compatible with MetaGPT's base install (newer litellm requires openai>=1.66; MetaGPT pins openai~=1.64)
• Matches existing provider conventions (header block, @register_provider decorator, async-first)
• Reuses OpenAI-shaped response parsing since litellm.ModelResponse follows OpenAI format
• Thinking/reasoning-mode passthrough matches anthropic_api.py

Result

$ pytest tests/metagpt/provider/test_litellm_api.py -v
tests/metagpt/provider/test_litellm_api.py::test_litellm_acompletion PASSED
tests/metagpt/provider/test_litellm_api.py::test_cons_kwargs_maps_config PASSED
tests/metagpt/provider/test_litellm_api.py::test_cons_kwargs_forwards_custom_base_url PASSED
tests/metagpt/provider/test_litellm_api.py::test_aask_code_forwards_tools_and_parses_arguments PASSED
tests/metagpt/provider/test_litellm_api.py::test_aask_code_respects_user_supplied_tools PASSED
======================== 5 passed in 5.32s =========================

Existing provider tests (test_openai.py) also pass, no regressions.

E2E smoke test against Azure OpenAI (azure/gpt-4o):

→ Model: azure/gpt-4o
→ Endpoint: https://<redacted>.cognitiveservices.azure.com                                                                                                                                                                                                                                                                                                              
                                                                                                                                                                                                                                                                                                                                                                        
[Test 1] acompletion, prompt 'What is 2+2?'                                                                                                                                                                                                                                                                                                                             
  response: '4'                                                                                                                                                                                                                                                                                                                                                         
                                                                                                                                                                                                                                                                                                                                                                        
[Test 2] aask with system message, prompt 'What is 10 - 3?'                                                                                                                                                                                                                                                                                                             
  response: '7'
                                                                                                                                                                                                                                                                                                                                                                        
MetaGPT LiteLLM live test PASSED.                                                                                                                                        

Exercised the full call path: LLMConfig(api_type=LITELLM) to LLM_REGISTRY[LITELLM] to LiteLLM._cons_kwargs to litellm.acompletion (real Azure endpoint) to response parsing.

Other

• Forwards api_key / base_url / proxy / reasoning config fields through to litellm.acompletion
• Default OpenAI base_url is intentionally not forwarded so LiteLLM can route by model-name prefix
• Cost tracking via _update_costs LiteLLM returns an OpenAI-shaped usage object
• aask_code / tool-use support lets MetaGPT's code-generating roles (Engineer, QAEngineer, DataAnalyst) work unchanged on LiteLLM-routed models. LiteLLM forwards tools=[...] to all providers that support native function calling and returns OpenAI-shape tool_calls regardless of backend, so the same parsing logic from openai_api.py applies.

[RheagalFire]

cc @geekan @garylin2099