Use AI to generate tests for AI-generated code?

[HermanMol]

Can we use AI to create the tests for AI-generated code? Looks like "We at company XYZ recommend to use the products of company XYZ". It gives the impression of a conflict of interest.

[jgebal]

Hi @HermanMol

I believe that having AI generate tests for code, or even better, to have AI generate tests from requirements and then generate code from requirements and tests, will be better than generation of code from prompt.

Automated tests always stood as a guardrail and safety harness for the code, preventing changes from breaking existing functionality.

If your AI model can generate and commit high quality tests then you get something magical.
You tell your AI agent to use tests as specification and if there are failing tests it is most likely a place to stop and rethink what is getting implemented.

It seems like a great topic for a deeper discussion.

[HermanMol]

Thank you Jacek,
I have tried with Copilot to generate a UT-package and although it has some improvements compared to the SQL Developer Generate utPLSQL test, it's not optimum (I'm lazy ;-) ).

I'm also not convinced that each time you let AI generate a new version of a UT-package for the same source package of an application, it will be the same. Verifying that the UT-package is correct would then also need a "re-testing of the UT-package".
I.m.o. the UT_package needs to be rather stable and only changed/add when&where the package to be tested changes.

I created my own generate-UT-package package which includes templates of the methods to be tested. It should provide the initial version and be changed manually. The package body is in HMX_UTPLSQL_UTL. The spec in the package_specs folder.

Along with annotations in the source (DB annotations and/or SQLcl annotations - I don't understand why Oracle has two approaches) the generation of the UT-package can be further completed.