Fix reasoning model JSON parsing | Generated by AI

Home 2026.07 English 中文

---

Fixed and pushed (7afc39b).

Two problems:

1. max_tokens=100 was too small — deepseek/deepseek-v4-flash is a reasoning model, so it spent all 100 tokens on internal reasoning (finish_reason=length) and never produced visible output. Bumped to 2000.
2. json.loads(response) failed when the model wrapped JSON in markdown fences or reasoning text. Now uses regex to extract the {...} JSON object from the response.

Try ww x unfollow again — the llm_error cases should be gone.

---

Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate