Clamp task scores to open interval

Prevent direct OpenAI bypass during inference

Support both validator proxy key env vars

Add local .env loading for inference

Use validator API_KEY for LiteLLM proxy calls

Handle fallback clients without connect method

Use OpenEnv websocket client sessions for inference

Wrap HTTP step payloads in OpenEnv StepRequest format

Support both old and new OpenAI Python clients

Ignore local egg-info artifacts

Harden inference bootstrap and container startup

Clean submission history