I think this is controversial 😏

I spent ages trying to figure out which one is right.

It turns out that this is in fact not a bug, even though the result is not intuitive.
The root cause for the "counter-intuitive" behaviour is that a [?…] bracket-specifier actually ends a projection. So the trailing [?starts_with(@, 'one') ends the upstream projection and ends up operating on an array-of-arrays, which is not the correct type as per the specification.

Source: from the Python mother-of-all implementations :

        # Everything above stops a projection.
        'star': 20,
        'filter': 21,

I also think that the language design that led to this difference is unfortunate and should definitely be addressed.
We need to try and think about a proposal that would fix this unambiguously.

I think this is controversial 😏

I spent ages trying to figure out which one is right.

It turns out that this is in fact not a bug, even though the result is not intuitive. The root cause for the "counter-intuitive" behaviour is that a [?…] bracket-specifier actually ends a projection. So the trailing [?starts_with(@, 'one') ends the upstream projection and ends up operating on an array-of-arrays, which is not the correct type as per the specification.

Source: from the Python mother-of-all implementations :

        # Everything above stops a projection.
        'star': 20,
        'filter': 21,

I also think that the language design that led to this difference is unfortunate and should definitely be addressed. We need to try and think about a proposal that would fix this unambiguously.

Potential workaround / design decision that should be documented?

Flatten operator

[?stack==''].branch[] | [?starts_with(@, 'one')]
// Result: ["one/"]