Striking how so many of the points in this article are also reasons mentat is better than Codex/Claude Code

It's weird to open a PR where the author can't contribute which is essentially the codex/claude code model

It's weird to have the agent figure out the tests when they're defined in code by CI.