Topics discussed:

• Michael Lawley wrapped Ontoserver in a deliberately thin MCP server — no cleverness, just a simpler API for agents — so you can throw clinical free text at it and get back terms mapped to SNOMED codes that are real codes and actually match their labels. His point was blunter than the demo: years of specialist NLP work on segmentation and named entity recognition has been rendered busy work by language models arriving and running it down. Asked by Nikolai about semantic search, he said the MCP does synonym search and you can watch the model guess again in the traces when a term misses, but embeddings stay behind the curtain — CSIRO runs an Ontoserver extension with models it trained itself, and Nikolai's own experience is that off-the-shelf embeddings over SNOMED perform poorly enough that training your own is the only option.
• Michael's second, newer skill probes a value set's scope by querying a terminology server, looking at both the compose and the expansion. It told him one value set referenced by an Australian IG had not been maintained, which he called a wake-up call. Running it twice gave different answers, because past roughly 50-100 codes it samples the expansion — he'd rather raise the thresholds, since most code systems outside the obvious ones are quite small, and it's only SNOMED value sets that get big enough to hurt.
• Gino Canessa's FHIR Augury indexes Zulip, Jira, Confluence, GitHub and FHIR packages into a SQLite database per source, each with full-text search, keyword extraction, staleness scoring and cross-references, so asking about a ticket returns the threads and commits that mention it, ranked so 2020 conversations don't outrank this month's. The design call worth stealing is the interface: a CLI that takes JSON and returns JSON, written for agents rather than humans — it costs fewer tokens than an MCP server because agents are already good at building JSON and calling CLI tools, and they don't stumble on the commands. Two skills sit on top: one preps tickets for working-group review with related threads and proposed dispositions, the other turns a resolution into a file-level implementation plan a human can argue with.
• Nikolai proposed pooling the raw indexed data, perhaps in an S3 bucket, so the community isn't all re-downloading the same sources — Confluence especially, which he found painful to index — and Gino agreed, wanting to finish iterating first. Nikolai also noted that what they are really building is training data, and suggested a licence so that if the model vendors come for it, HL7 gets paid.
• Brian Kaney's Health Skills repo is built to complement spec frameworks like OpenSpec and SpecKit rather than replace them — product discovery, FHIR modelling, refactoring, human factors, compliance and docs — with jurisdiction overlays declared once in a bootstrap file and flowing into every skill, and base skills compiled against agent profiles into a dist branch so the source of truth stays single. Nikolai challenged whether spec-driven development actually works or is an illusion, pointing to evaluation showing heavy specs and documentation can burn tokens without improving the output. Brian's answer was that skills are how you scale subject-matter expertise across a mixed team, with a full EHR rebuild his team is halfway through on a four-month deadline as the proof, and that hands-off use without actively managing context is what produces the bad results.