The project is pre-alpha, in its foundation phase: scaffold, site-profile schema, persona spec, and the governance approvals that gate publication.

That label is doing real work. Documentation tends to describe a project in the present tense whether or not the present tense is earned, so this page separates three things a reader otherwise cannot tell apart: what runs, what is decided but unbuilt, and what is parked on purpose.

What runs today

Capability State
/setup interview and edit modes Working; see known gaps below
/triage clarification and routing Working
/kb-draft in manual mode Working — paste a worklog, get a draft
/downtime with generic templates Working
/comms-tune template shaping Working
/explain and the onboarding path Working

Everything above operates against a site profile you build locally. With no profile, skills fall back to generic templates labelled not yet tuned to this site rather than guessing at your environment.

Decided but unbuilt

These are settled design decisions with no implementation behind them yet. They are not open questions.

Connected knowledge mode. /kb-draft currently drafts for you to paste. Reading resolved incidents and submitting articles directly requires a ServiceNow MCP connection, which is blocked on access rather than on design. Gap detection and stale-article flagging are blocked on the same thing — in manual mode they work by asking you rather than by scanning a knowledge base the skill cannot see.

ADR backfill. Eight architectural decisions are recorded in the index; individual records are still being written up. The decisions were made in design, not retrofitted — but only ADR-0008 exists as a full record so far.

Parked on purpose

Message forensics — HL7 message parsing — is a separate plugin, not a missing feature here.

It waits on two things: data access, and a symbolic parse layer underneath it. Shipping a model-interpreted version in the meantime would be the exact behaviour this project argues against — a language model freestyle-parsing pipe-delimited clinical data is the failure mode the whole design exists to avoid.

The reasoning is in ADR-0008; parked work is tracked in FORENSICS-BACKLOG.md.

This is the one place where “not yet” and “not ever, here” are easy to confuse. Forensics will exist. It will not exist inside this plugin.

Known gaps

Found by running the plugin against a real environment rather than by design review:

  • The systems schema is single-site. A practice supporting many facilities across several PACS platforms cannot be represented cleanly, and /setup will ask you how to model it rather than deciding. See site profile setup.
  • The interview does not checkpoint. A long /setup session that ends early writes nothing. Answers given are lost.
  • Conflicting answers are not cross-checked. If a late answer contradicts something recorded earlier in the same interview, the contradiction is not currently surfaced.
  • Deferred questions are not tracked. Items postponed during an interview live in the conversation, not in the profile, so edit mode cannot resurface them.

How this is developed

Developed in collaboration with Claude, and that process is published rather than hidden. Decision records capture why it is built this way, including rejected alternatives. Guardrails documents the specific failure modes the collaboration is checked against — design and drafting are co-developed; behaviour rules and published claims are validated by a human before release.