add eval for security_review prompt

[elfedy]

Adds the first eval so that we have a skeleton to evaluate and improve our flows. These are meant to be run locally for now with own credentials for LLM models.

An eval has runs which consist on a task (something that the agent performs using the mcp server) and an eval which performs different evaluations of the task. Results are stored locally in an .evals folder under a randomly generated id.

Also adds a clean script to delete run results.

[snowmead]

Awesome job! 💯

[snowmead]

Did you include things like zombienet, node, etc. to potentially have other types of evals like having the agent run the node and test transactions?

[elfedy]

Yes I didn't think that deep but I just left the template as is because any part of it could be part of a potential eval.

[snowmead]

proposal: we could probably one shot a very simple run+eval framework so we don't need to duplicate any code when creating new evals

[elfedy]

proposal: we could probably one shot a very simple run+eval framework so we don't need to duplicate any code when creating new evals

Yes. I think this is better done at the time we create the second eval so we have good info about what needs to be reused and where it makes more sense to just duplicate.