Skip to content

reasoning leaks into the content channel on live traffic — known issue, not force-reproducible; boundary tracing added to catch and fix #250

Description

@warelik

中文 TL;DR

已知问题,记录在案:thinking 模型偶尔把 reasoning 写进 CONTENT 通道。只能在真实会话中偶发复现,受控探针无法复现。已有多代防御:#238(空答案救援)、#241(救援精度)、#243(前导 <think> 块从 CONTENT 重定向到 thinking);另有 #247(出口逐字去重)、#248(会话连续性)、#249(边界结构化日志,门控默认关)。下一步:开着 tracing 抓现场,抓到后另开 PR 提修复。

Что известно

Что дальше

Если у кого-то есть пойманный живой пример (сырой ответ, где reasoning виден в content) — это самый короткий путь к репродукции; приложите, пожалуйста, сюда.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething is broken or behaves incorrectly

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions