I really like the IR-based approach, it solves something that's always bothered me about shell pipelines: you're forced to think in terms of serializing bytes, even when both ends of the pipe are the same program and could just share memory. Flash makes that optimization explicit and easy to compose with the rest of the pipeline. One question, though: have you run into any issues with the "opportunistic" binary format substitution (the .flatgfa fallback) when scripts are shared across machines where some files have already been converted and others haven't?