hub-gif be0215bfe2 fix(ai_crawler): cap max_tokens to remaining context window
LiteLLM rejects when max_tokens exceeds context minus measured input
tokens. Estimate input from prompt length and clamp max_tokens; allow
LLM_CONTEXT_WINDOW override. Prevents ContextWindowExceededError on long
Chinese/JSON prompts (e.g. regenerate-report §8.5).

Made-with: Cursor
2026-04-14 10:16:38 +08:00
..