[Bug][multi-agent-shared-context]: LLM_MODEL in .env is ignored - read at import time, before load_dotenv() #483

Affected Component

statewave-multi-agent-shared-context — main.py

Bug Description

LLM_MODEL set in .env is ignored. The model is read at import time, before .env is loaded:

# main.py:34  (module level, runs on import)
_LLM_MODEL = os.getenv("LLM_MODEL", "groq/llama-3.3-70b-versatile")

# main.py:91  (inside main(), runs later)
load_dotenv()

LLM_API_KEY is picked up from .env, because it's read inside _build_llm_client() after load_dotenv() has run. So a user who follows .env.example and sets LLM_MODEL=openai/gpt-4o-mini with an OpenAI key gets the Groq default with an OpenAI key, and an authentication error that looks like a bad key. The README troubleshooting section (line 288) even describes that symptom.

timeline_inspector.py gets this right: it calls load_dotenv() first (line 199).

Expected Behavior

LLM_MODEL from .env is honoured.

Suggested fix

Move load_dotenv() to module level, right after the imports in main.py (before line 34). Alternatively, read LLM_MODEL inside main().

Steps to Reproduce

1. In .env: LLM_MODEL=openai/gpt-4o-mini, LLM_API_KEY=<openai key>.
2. python main.py --mode naive → LiteLLM calls groq/llama-3.3-70b-versatile and fails auth.
3. export LLM_MODEL=openai/gpt-4o-mini in the shell instead → works.

Environment

main @ eb66ddc.

Severity

Medium for a first run — the documented way to choose a provider doesn't work.