This is due to #5000. The motivation of that enhancement was making it more convenient for listeners and model modifiers to set keyword arguments, and it didn't occur to me that it would also affect using BuiltIn.run_keyword programmatically. Or possibly it did occur, but I thought it's so unlikely that it would cause problems that we don't need to care.

Fixing this issue so that BuiltIn.run_keyword works like earlier, without special cases, when used programmatically is easy. The problem is that then libraries using it cannot benefit from the more convenient API provided by #5000. That's unfortunate, but I believe that behavior is safer. We also already have various issues related to making it easier for libraries to call other keywords. I hope we could enhance that situation in general in RF 7.1 or latest in RF 8.0.

This problem can be reproduced also without a library by passing keyword name and arguments as a list variable to Run Keyword so that it needs to resolve also arguments:

${name and args} =    Evaluate    'Log Many', ['a', 'b', 'c'], {'d': 'e'}
Run Keyword    @{name and args}