Running Qwen Code on a Custom OpenAI-Compatible API: modelProviders, a 41,000-Character First Request and Retries That Keep Going for Minutes