fix: strip trailing newline from GROQ_API_KEY before using in HTTP header

fix: use httpx directly for Groq API (bypass SDK connection issues in HF Spaces)

fix: route LLM to Groq (deepseek-r1) instead of hardcoded gpt-4o/openai

fix: auto-load .env in llm_client; add SWEBenchLoader class to loader

fix: normalise LangGraph dict output back to AgentState in _run_with_langgraph

Initial commit