[0z5a]

Unpacker(file_like) in the C extension checks if file_like: before it looks up file_like.read. A stream that defines __bool__ or __len__ and is falsy is therefore stored as the input source, but its read method is never saved. The first unpack(), next() or read_bytes() that needs data then fails with TypeError: 'NoneType' object is not callable.

Such objects are common file wrappers. For example, werkzeug's FileStorage is falsy when it has no filename, so Unpacker(FileStorage(io.BytesIO(data), filename="")) fails with the C extension. The pure Python Unpacker checks file_like is None and unpacks the same stream correctly, and the documentation only asks for an object with a .read(n) method.

import io
import msgpack

class FalsyBytesIO(io.BytesIO):
    def __bool__(self):
        return False

list(msgpack.Unpacker(FalsyBytesIO(msgpack.packb([1, 2]))))
# C extension:  TypeError: 'NoneType' object is not callable
# fallback:     [[1, 2]]

This PR changes the check to if file_like is not None:, which matches the fallback and the rest of the class (feed(), read_bytes() and _unpack() already test self.file_like is not None).

Nothing that worked before changes behaviour. A falsy non-None file_like could not be used at all: feed() raised AssertionError and reading raised the TypeError above. A falsy object without a read attribute, such as b"", now raises AttributeError in the constructor, as the fallback already does.

A regression test in test/test_sequnpack.py unpacks two objects and calls read_bytes() on a falsy BytesIO subclass. It fails with the C extension before this change and passes after it; it passes with MSGPACK_PUREPYTHON=1 both before and after. The full test suite passes with the C extension (179 passed, 1 skipped) and with the pure Python implementation (176 passed, 4 skipped), and ruff check / ruff format --check are clean.

This change was drafted with Claude Code (claude-opus-5-5); the test results above are from the runs described.