fix: add score= to END output, clamp to (0.01, 0.99)

fix: use HF_TOKEN as api_key per hackathon spec, fix END format

fix: use os.environ for API_KEY and API_BASE_URL, remove dummy fallback

fix: use GenericAction(**kwargs) instead of positional dict arg

fix: emit structured START/END output even when env connection fails

fix: run all 7 tasks with per-task START/END logging, match reference format

fix: test LLM proxy call before env connection to ensure proxy usage

fix: prioritize API_KEY over HF_TOKEN for LLM proxy, fail loudly on first call

fix graders

fix inference final

fix get model_messages

fix: api key as first priority

fix graders

fix: inference format

fix: inference

fix inference format

minor changes

inference format fix

minor changes

Major fixes

feat: harden Task 5 grader + fix inference.py for hackathon submission

fix: add demo and inference.py