- Replace backup model stepfun-ai/step-3.5-flash with meta/llama-3.1-8b-instruct
(stepfun is a thinking model that uses all tokens on reasoning and
never outputs content, causing all 3 fallthroughs to fail)
- Add retry with doubled max_tokens when primary model returns empty
content (deepseek-v4-flash thinking can exhaust token budget)
- Increase backup timeout to 120s and max_tokens to min 2048
- Move callApi error handling to return null instead of throw for
cleaner fallthrough logic with timeout logging
yuzhiran
wlt