feat: add LiteLLM as LLM provider

[RheagalFire]

Summary

• Adds LiteLLMProvider as a new LLM adapter in cc_kernel/runner/llm/, enabling access to 100+ providers (OpenAI, Anthropic, Google, Azure, Bedrock, Ollama, etc.) via a single unified SDK
• Implements the same Provider protocol as AnthropicProvider: __call__(LlmRequest) -> LlmResponse + stream(LlmRequest, on_delta) -> LlmResponse
• Added as optional dependency: pip install cheetahclaws[litellm]

Changes

• cc_kernel/runner/llm/litellm_provider.py - new LiteLLMProvider with:
  • litellm.completion() with drop_params=True
  • stream() with on_delta callback for streaming text deltas
  • Multi-turn messages + system prompt support
  • Tool calls parsed into canonical {id, name, input} format
  • Token usage tracking
• pyproject.toml - added litellm optional dependency (pip install cheetahclaws[litellm]), also included in all extra
• requirements.txt - added litellm>=1.60.0,<2.0.0
• tests/test_litellm_provider.py - 12 unit tests (all passing)

Tests

Unit tests (12/12 passing):

$ PYTHONPATH=. python -m pytest tests/test_litellm_provider.py -v
test_default_timeout PASSED                                                                                                                                                                                        
test_custom_api_key PASSED
test_no_lazy_state PASSED                                                                                                                                                                                          
test_calls_litellm_completion PASSED                                                                                                                                                                               
test_omits_api_key_when_none PASSED
test_system_prompt_included PASSED                                                                                                                                                                                 
test_multi_turn_messages PASSED
test_returns_llm_response PASSED
test_rejects_non_llm_request PASSED                                                                                                                                                                                
test_api_error_raises_provider_unavailable PASSED
test_stream_rejects_non_callable_on_delta PASSED                                                                                                                                                                   
test_litellm_in_requirements PASSED
12 passed in 1.18s                                                                                                                                                                                                 

Live E2E tests (3/3 passing against real API):

from cc_kernel.runner.llm.litellm_provider import LiteLLMProvider                                                                                                                                                  
from cc_kernel.runner.llm.provider import LlmRequest
                                                                                                                                                                                                                   
provider = LiteLLMProvider(api_key=os.environ["ANTHROPIC_FOUNDRY_API_KEY"])
                                                                                                                                                                                                                   
req = LlmRequest(model="anthropic/claude-sonnet-4-6", user="What is 2+2?", max_tokens=10)
resp = provider(req)                                                                                                                                                                                               
print(resp.text, resp.tokens_input, resp.tokens_output)

Test 1 (basic): text="4", tokens_in=20, tokens_out=5
Test 2 (stream): text="OK", chunks=1                                                                                                                                                                               
Test 3 (system): text="Ahoy!"
ALL 3 E2E TESTS PASSED                                                                                                                                                                                             

Example usage

from cc_kernel.runner.llm.litellm_provider import LiteLLMProvider
from cc_kernel.runner.llm.provider import LlmRequest

# LiteLLM reads provider keys from env automatically
provider = LiteLLMProvider()                                                                                                                                                                                       

# Basic call                                                                                                                                                                                                       
req = LlmRequest(
    model="anthropic/claude-sonnet-4-20250514",                                                                                                                                                                    
    system="You are a security expert.",
    user="Find SQL injection vectors in this query...",                                                                                                                                                            
)                                                                                                                                                                                                                  
resp = provider(req)
print(resp.text)                                                                                                                                                                                                   
                
# Streaming
def on_token(delta):
    print(delta, end="", flush=True)                                                                                                                                                                               

resp = provider.stream(req, on_delta=on_token)                                                                                                                                                                     

See https://docs.litellm.ai/docs/providers for 100+ supported model strings.

Impact

• Additive only, existing AnthropicProvider untouched
• Follows the same Provider protocol (__call__ + stream)
• Same error types (ProviderUnavailable, ProviderInvalidRequest)
• litellm is an optional dependency (pip install cheetahclaws[litellm])
• drop_params=True silently drops provider-unsupported kwargs

[RheagalFire]

cc @chauncygu

[chauncygu]

Hi @RheagalFire, thanks for the PR. A few blockers before this can merge:

1. Description vs. diff mismatch. Body says litellm is an optional dep, but the third commit moved it into [project] dependencies in pyproject.toml and the core block of requirements.txt. There's no litellm extra and it's not in all. Please move it back to [project.optional-dependencies] and add it to all.

2. Not wired into either user path. cc_kernel/runner/llm/__main__.py:_select_provider() only knows mock / scripted / anthropic, and the top-level providers.py (the registry the CLI / Web UI consult for --model X) has no litellm entry. So no end-to-end caller can reach the new class — the 12 mocked tests pass by constructing it directly. Please add a litellm branch to _select_provider() and a "litellm" entry to PROVIDERS in providers.py.

3. stream() drops usage + tool_calls. It always returns tokens_input=0, tokens_output=0, tool_calls=(). The runner emits ledger charge messages from those fields and decides tool dispatch via response.is_tool_use, so streamed calls bypass quota and silently lose tool calls. Please request stream_options={"include_usage": True} and read usage off the final chunk; assemble tool_calls from streaming deltas, or raise ProviderInvalidRequest when request.tools and request.stream are both set.

4. cost_micro hardcoded to 0 in both paths. AnthropicProvider keeps a pricing table with a fallback — please mirror that, or set metadata={"cost_unknown": True} so callers can detect it.

Smaller items: live E2E results in the PR body aren't in the test suite (please add a skipif-gated tests/e2e_litellm_provider.py); test_litellm_in_requirements reads Path("requirements.txt") and fails outside the repo root; __call__ and stream() duplicate ~30 lines of message-building (factor like anthropic_provider._build_kwargs); no doc update under docs/guides/ — would help to call out when to prefer litellm/<provider>/<model> over the existing custom/ / ollama/ / anthropic/ adapters (Bedrock SigV4, Azure deployment routing, Vertex auth are the real value-add).

Direction is good — fix blockers 1–4 and I'm happy to merge. Thanks!

[chauncygu]

Hi @RheagalFire, Thanks for your great contributions. I have merged and fixed the issues.

Thanks.