Day 20: Started Phase 3 of the project. Built retrieval into the agent and executed first live voice test which failed.
Greeting exchange worked ("Hi, am I audible?" → answered), but llm_ttft was already 7.4s because the call needed two retries against a "503 UNAVAILABLE" - model currently experiencing high demand , before succeeding.
The very next turn - "Introduce yourself. In short." — which the grounding prompt correctly routes through search_my_background, since "introduce yourself" is a factual claim from the corpus — hit a "429 RESOURCE_EXHAUSTED": generativelanguage.googleapis.com/generate_content_free_tier_requests capped at
"5 requests/minute" for gemini-3.7-flash, the model `GEMINI_MODEL=gemini-flash-
latest currently resolves to. The plugin retried three more times internally and then raised APIConnectionError: failed to generate LLM
completion after 4 attempts. agent_state dropped from thinking back to listening` with nothing spoken, which lead to a infinite hang, because nothing tells the visitor anything went wrong.
So now I need to choose a different model which has atleast 10 RPM. Also implement a fallback message, so when same error occurs it can be identified instead of dead silence for 5 mins which I experienced a few mins ago.