Verification command

[AndreyMalykhin]

Hi, what do you think about having a verification command based on Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification https://arxiv.org/abs/2601.15808 ?

[LeoVS09]

Good suggestion. It is actually already implemented. In the sadd plugin and named as /judge. The main agent will launch the Judge agent, which uses predefined rubrics to judge the results of code implementation. I will update the README to describe it properly.

[LeoVS09]

In general, there may be meaning in creating a separate /verify command that does not use a sub-agent or is more generic. That can work similarly to /reflect. But I still struggle to find proper "generic rubrics" that can be used for any use case. For example current rubrics include: Code Quality, Usability, Maintainability... They are not suitable for other use cases, such as research on a better database. On the other hand, when I write more specific rubrics tailored to the task, I get much better results.

So this approach has a tradeoff: quality vs generalization. As a result, I do not see a way to implement it as generic as the /reflect command, which can be used to reflect on code and on an answer about some "rare" topic. Mainly, it is because LLMs are currently too bad at creating proper rubrics tailored to the use case. But if you have any prompts or ideas on how to make them more predictable in it, then great, I will be happy to implement this command or adjust /judge.