- backend/ai_chat.py: chat_completion provider-agnostique (OpenAI-compat tools/tool_calls + Gemini functionDeclarations/functionCall), retry sans tools si rejete - backend/agent/loop.py: run_agent multi-etapes (limite 10, truncation, confirmation two-step), LLM injectable - endpoint opt-in POST /api/ai/bookslm/agent (events SSE tool/message/confirmation), extraction _resolve_system_prompt - tests: test_ai_chat.py, test_agent_loop.py + 3 tests endpoint (728 passed au total) - ROADMAP B1/B2/B3/B7 livres ; B4/B5/B6 restants
