Consider allowing boolean tests of arbitrary values to fail

[brandjon]

The spec requires that

any value may be converted to [...] a Boolean truth value using the expression bool(x).

This FR would consider relaxing this to allow truth-checking to fail for certain types, so e.g. if some_application_defined_type: could be an error.

This would enable banning (via a dynamic error) certain patterns that are error-prone, such as Bazel's select, where a boolean could be inadvertently conflated with an object wrapping a computation yielding a boolean. See bazelbuild/bazel#14506.

Downsides: This would effectively prohibit programmatically taking the truthiness of arbitrary values, including where this is done implicitly (without calling bool()), and break existing code that does this. A more minor detail is that it would require implementaitons (particularly the Java interpreter) to propagate an exception where one is not declared today.

Alternatives: Rely on linting, style guides, and documentation to mitigate the damage caused by that kind of confusion.

Personally I'm hesitant to restrict truth-checking in this way.

[laurentlb]

Note that it would also break some patterns, like x or default_value

[brandjon]

See update in aforementioned bug. We want to do this, at least for implicit bool conversion.

[adonovan]

Personally I'm hesitant to restrict truth-checking in this way.

Me too. Truthiness has long been a part of Python style, and revoking it would be an enormously incompatible change.