Question for the provenance-minded here. I've been building FlowProof, an MCP server that lets an AI assistant run Nextflow pipelines and get back a Workflow Run RO-Crate for each run (tool versions, params, SHA-256 of every input and output). One design choice I keep going back and forth on: nf-prov serialises the crate from inside the engine, which is clean and battle-tested. I've instead put the provenance emitter above the backend interface, so it's engine-agnostic and the crate is retrievable through the same MCP call that launched the run. The tradeoff is I'm reconstructing some of what the engine already knows. For people who run provenance in anger: is that separation actually useful to you (e.g. wanting one provenance shape across engines), or is in-engine (nf-prov style) just strictly better in practice? Genuinely trying to figure out if the engine-agnostic layer earns its keep. Repo if useful: github.com/ajibadedapo/flowproof-mcp

wrong channel, this is one is for nf-core/tools 😉