Fix reasoning model token exhaustion | Generated by AI

Home 2026.07

Two changes applied. Here’s the summary:

Root cause: deepseek/deepseek-v4-flash is a reasoning model. It outputs chain-of-thought in a reasoning field before producing the actual content. When content is null and finish_reason is "length", the model exhausted its token budget during reasoning and never emitted the answer.

Fixes (both files in /home/xiaoxin/projects/ww/):

  1. ww/llm/openrouter_client.py — when content is null, fall back to the reasoning field instead of crashing. A warning is printed so you know it happened, but the note gets created.

  2. ww/note/create_note_utils.py — bumped max_tokens for title generation from 1024 → 4096, giving reasoning models enough room to both think (reasoning) and produce the 6-word title (content) without exhausting the budget.

If the watcher/processor is already running, you’ll need to restart it for the changes to take effect (Ctrl+C the watcher, then ww note watch again).

Back Donate