OpenAI-managed Responses delegation

Configures GPT Live with session.delegation.type: "responses". GPT Live owns backend conversation continuity and result injection; application tool calls execute inside the assistant against isolated application-owned state. Evaluators only observe tool events and final state.

Commands assume the harness directory. See setup for setup and installed-wheel usage; live commands incur API usage.

Everything needed to customize this backend lives in this folder: prompts/backend.txt, tools/restaurant.py, tools/definitions.json, and tools/restaurant_facts.json. Frontend instructions live in ../frontend/prompts/voice.txt.

The bundled restaurant behavior follows the comparison-baseline contract. Keep baseline fixes aligned with the client assistant; document intentional variants.

uv run run-eval --scenario restaurant_booking_complete --assistant responses