@mbeijen I'm confused. Do you agree with the suggestion or not? 👀

Sorry for the confusion @Kludex — I went the opposite direction on the squash and should have explained.

Your suggestion (top-of-decode() check with if self.decompressor.eof and data:) handles the multi-frame-across-calls case, but it doesn't cover one scenario: an empty decode(b"") call after EOF. The and data guard skips the swap, then decompressor.decompress(b"") is called on the EOF'd decompressor and raises EOFError: Already at the end of a Zstandard frame. on stdlib compression.zstd (verified on 3.14.4). That happens internally via MultiDecoder.flush(), which calls child.decode(b"") to drain residue across stacked encodings.

The post-decompress swap avoids this because the decompressor is always fresh when the next decode() is entered, regardless of what data is. I also dropped the (now-unreachable) top check in the squashed version, which is why the diff looks different from what you reviewed.

Happy to switch to a top-check + if not data: return b"" early-return instead if you prefer that shape — it's the same number of lines and avoids the seen_eof flag. Let me know.

HAHAHAHA Claude answered for me while I was checking to see if or how we can go each way. Anyway you read the text, we can avoid seen_eof flag but we must handle empty frames AND eof both