DSH | dsh-auto-review | second-model approval review with an audit trail

[PerryLink]

Project: https://github.com/PerryLink/dsh-auto-review

What it does. A second model reviews actions that cross the sandbox boundary and returns a structured allow/deny verdict with a reason. It is fail-closed by default and writes a full audit trail.

How it integrates with DSH. It registers on the approval answerer chain, reads the tool call and its evidence, and returns a verdict before the human is asked; it also exposes /auto-review and a session-projection review panel.

Install. dsh plugin --profile web add dsh-auto-review

Unofficial community project; not endorsed by the DSH maintainers.

[lloydzhou]

是否考虑使用 jev？

[PerryLink]

不同项目应对不同场景嘛。
可以看看我的模型和项目，
https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1
质量比jev好，效率快10倍。本地部署很方便，CPU也能跑。
还有对应的DSH的项目，https://github.com/PerryLink/dsh-phocinae

[lloydzhou]

dsh-auto-review 走的是你的这个 Phocinae-Largha-150M-v1 的模型吗？我看项目写的是 second-model 看起来只是另外一个LLM？

[PerryLink]

不是。
dsh-auto-review你可以自己配一个自己喜欢的第三方的任意模型。
如果你觉得需要，后面我可以在dsh-auto-review的更新中，加上可以选配本机运行的Phocinae-Largha-150M-v1。

[lloydzhou]

我希望加的是 类 jev的这种判别模型（可以是Phocinae-Largha-150M-v1）

[PerryLink]

dsh-auto-review后面我找时间加上。