feat: add LiteLLM as a unified LLM gateway provider

[RheagalFire]

Summary

Adds LiteLLM as a new LLM provider across all of kit's LLM call paths (Summarizer, PR Review, Agentic Review, Commit Generation, Deep Research, and CLI). LiteLLM acts as a unified gateway that routes to 100+ LLM providers (OpenAI, Anthropic, Azure, AWS Bedrock, Vertex AI, Ollama, Groq, etc.) through a single litellm.completion() interface.

Changes

11 files changed, +458 / -11 lines

File	What changed
pyproject.toml	Added litellm>=1.80.0,<1.87.0 optional dep group, included in all
src/kit/summaries.py	Added LiteLLMConfig dataclass + litellm.completion() branches in summarize_file, summarize_function, summarize_class
src/kit/__init__.py	Exported LiteLLMConfig
src/kit/llm_client_factory.py	Added create_litellm_client() factory + config/review-config dispatch
src/kit/pr_review/config.py	Added LLMProvider.LITELLM enum, litellm/ prefix detection, config loading, example config
src/kit/pr_review/reviewer.py	Added _analyze_with_litellm_enhanced() with cost tracking
src/kit/pr_review/agentic_reviewer.py	Added _run_agentic_analysis_litellm() with OpenAI-compatible tool calling
src/kit/pr_review/commit_generator.py	Added _generate_with_litellm()
src/kit/deep_research.py	Added LiteLLMConfig branch in _init_llm_client() and research()
src/kit/cli.py	Added LiteLLM API key handling in 3 model-switching locations
tests/test_llm_client_factory.py	Added 5 unit tests for LiteLLM client creation

Usage

Python SDK (Summarizer)

from kit import Repository, Summarizer, LiteLLMConfig

config = LiteLLMConfig(
    model="anthropic/claude-sonnet-4-6",  # any litellm-supported model
    api_key="sk-...",                       # optional if set in env
    api_base="http://localhost:4000",       # optional: LiteLLM proxy URL
)

repo = Repository("/path/to/repo")
summarizer = Summarizer(repo, config)
summary = summarizer.summarize_file("src/main.py")

PR Review (config file)

llm:
  provider: litellm
  model: bedrock/anthropic.claude-3-sonnet  # Bedrock via LiteLLM
  # api_key: optional
  # api_base_url: http://localhost:4000     # optional proxy

CLI

# Direct provider routing
kit summarize file src/main.py --model litellm/anthropic/claude-sonnet-4-6

# Via LiteLLM proxy
LITELLM_API_KEY=sk-proxy-key kit summarize file src/main.py --model litellm/openai/gpt-4

Design decisions

• Lazy import: import litellm happens at call time, not module load. Kit imports cleanly without the optional dep installed
• drop_params=True: Silently drops provider-unsupported kwargs (e.g., seed on Anthropic). Critical for cross-provider compatibility without per-provider param filtering
• OpenAI-compatible response parsing: LiteLLM returns ModelResponse with the same .choices[0].message.content shape as OpenAI, so response handling mirrors the existing OpenAI provider
• Optional dep group: pip install 'cased-kit[litellm]' keeps the base install lightweight
• Version pin >=1.80.0,<1.87.0: Floor covers all features used (drop_params, tool calling, async). Ceiling avoids surprise breakage from LiteLLM's weekly releases

Tests

Unit tests (25 tests touching LiteLLM paths, all pass)

tests/test_llm_client_factory.py  - 20 passed (5 new LiteLLM tests + 15 existing)
tests/test_summaries.py           - 46 passed
tests/test_pr_review.py           - 63 passed
tests/test_cli_parsing.py         - 27 passed
tests/test_ollama_client.py       - 8 passed

Ruff lint

ruff check src/ tests/ -- All checks passed

Live E2E (Summarizer path)

Model: anthropic/claude-sonnet-4-6 via Azure Foundry endpoint
Input: src/kit/llm_client_factory.py
Output: 810-char structured summary (## Summary, ### Purpose, ### Key Components)
Chain: LiteLLMConfig -> Summarizer.summarize_file() -> litellm.completion() -> Azure Anthropic -> parsed response

Live E2E (DeepResearch path)

Model: anthropic/claude-sonnet-4-6 via Azure Foundry endpoint
Query: "What is the factory design pattern in 2 sentences?"
Output: 684-char research result
Time: 5.88s
Chain: LiteLLMConfig -> DeepResearch.research() -> litellm.completion() -> Azure Anthropic -> ResearchResult

Testing instructions

pip install -e '.[litellm]'

# Summarizer
python -c "
from kit import Repository, Summarizer, LiteLLMConfig
config = LiteLLMConfig(model='openai/gpt-4o-mini', api_key='sk-...')
repo = Repository('.')
print(repo.summarize_file('README.md', config=config))
"

# PR Review
kit review pr --model litellm/openai/gpt-4o-mini <pr-url>