Hi all,

User skills are unclear for me now. Could you provide some examples?
Looks like AI agents already can search docs, forums and even source
code if it's required.

Next, about contributing skills.

Firstly, I think we have to declare an AI usage policy for the Apache
Ignite project. Many projects reject AI-generated PRs, such as Apache
CouchDB and PouchDB (see [1] as an example of projects that prohibit
AI usage). Some other projects do not encourage AI, but do not
prohibit it either. For example, Apache DataFusion allows AI-assisted
PRs, but requires the author to understand the implementation
end-to-end and to be able to justify the design and the code during
review [2].

Secondly, from my experience, it is challenging to review vibe-coded
PRs from new contributors for several reasons:

- Agents do not follow coding guidelines. Sometimes they do not even
start the code style check.
- Agents do much more than required for the ticket.
- Agents generate many more lines than needed, without trying to
reduce code lines.
- Agents sometimes do not see connections between different parts of
the code. As an example from one of the PRs: the agent generated a new
test, but did not add it to the test suites.

So, I suggest two action items:

1) Start a discussion/voting on the rules of AI usage.
2) If we agree to allow AI-generated code, then create AGENTS.md as a
root file (to have more chances for AI to read it) and write the rules
there:

- Check CONTRIBUTING.md before start
- Make the patch as small as possible and do only one goal
- Make the code similar to the neighbouring files
- If new test files were generated, check that they are connected to test suites

[1] https://codeberg.org/brib/slopfree-software-index
[2] 
https://datafusion.apache.org/contributor-guide/index.html#ai-assisted-contributions

Best regards,
Aleksandr Chesnokov

Reply via email to