← Back to context

Comment by mashlol

1 hour ago

If you try to just instruct them they might get it wrong. If the surrounding software forces them to do it, it'll always work. E.g. it can check for merge conflict markers like <<<<<<< and re-invoke codex/claude to merge again if the previous resolution failed for whatever reason (e.g. AI hallucinated, threw up, whatever).

Also you'll need a wrapper to actually detect when merge conflicts will occur and when rebasing is necessary.

Generally, the less you rely on the AI the better. Make the AI write the code, sure, but don't make the AI be the process.

I think that's being pretty uncharitable to the model... I use Opus and it has zero issue handling this