Repository navigation
[Bug] DeepSeek DSML tool-call markers leak into assistant content and streamlit UI #804
Closed as not planned
Description
Summary
When using DeepSeek with tool calls, raw DSML markers such as <‖DSML‖tool_calls> and <‖DSML‖invoke ...> are leaking into the assistant content field. Downstream consumers (UI/TUI) then display or speak these internal markers.
Suggested fix
- Add DSML marker detection before emitting
content. - Strip / normalize Unicode variants (
‖,｜, etc.). - Keep tool-call data only in structured
tool_calls. - Add regression tests with streaming chunks split across markers.
Activity
ZhulongNT commentedon Sep 22, 2026
补充一组实际
model_responses扫描结果（2026-09-22）：DSML命中。<‖DSML‖...>，还包括双全角竖线形式，例如：<｜｜DSML｜｜parameter ...></｜｜DSML｜｜parameter><｜｜DSML｜｜tool...>hello<‖DSML‖tool_calls>world如果过滤器逐 chunk 检查完整字符串
DSML，上述分块会原样泄漏。建议修复时使用跨 chunk 的小型状态机/尾部缓冲，并覆盖‖...‖、｜｜...｜｜、|...|的开闭标签及分块边界回归测试。lsdefine commentedon Sep 29, 2026
Closing as won't fix, see the discussion in #805: DSML leaking into
contentis a serving/parser-side issue, and we keep model output faithful on the agent side.