Roadmap

Candidate features, scored. Two axes, because they disagree often: importance is how much daily pain a feature removes for someone modelling an ontology and publishing documentation; creativity is how novel it is against Protégé, WebVOWL, Chowlk, Graffoo, yEd, draw.io and Google Slides. Effort is S ≈ a day, M ≈ a week, L ≈ a milestone, XL ≈ a project.

Nothing here is a commitment. Issues and opinions are welcome on GitHub.

Shipped in 0.7.7

Struck from the tables below:

• Explain this edge (imp 7 · cre 6.5) — right-click an inferred edge for a proof tree down to asserted facts. Built exactly where this roadmap said it should be: the single add(s,p,o) funnel in the OWL-RL materializer. The reasoner-parity risk noted at the bottom of this page turned out to be the design — backends that cannot justify a triple say so.
• Select by quality finding (imp 7 · cre 6.5) — smell chips and competency-question results select their terms on the canvas.
• Competency questions by example (cre 8 · imp 6) — partly. The ontoink-cq fence and the CI gate shipped; the novel half, deriving the SPARQL from a selection so you never write it by hand, did not. It stays below, rescoped.

Also shipped, though never on this roadmap: shape recommendation with per-constraint evidence, shape-drift detection, the build report and badges, and the ontology citation panel.

By importance

Feature	imp	cre	effort	Why
Layout lockfile — .ontoink-layout.json + layout: fence key	8	6.5	M–L	Makes a hand-arranged diagram durable and reviewable. Today an arrangement lives in localStorage: private to one browser, colliding between fences built from the same file, invisible to CI.
Real figures — caption: / id: / alt:, numbering, cross-references	7	2	M	A documentation plugin currently emits a bare <div> and authors hand-maintain "Figure 3".
Select by quality finding — OntoSniff report → graph selection	7	6.5	S	The detectors already carry the IRIs; today they render as text next to a selection API that could act on them.
Explain this edge — justify an inferred, SHACL or restriction edge	7	6.5	M	Every OWL-RL rule funnels through one add(s,p,o); widening that turns a black box into a glass box. An unexplained purple edge costs trust.
Extract module — closure policies over a selection	7	5.5	L	Signature, +superclasses, +domain/range, +restrictions — and show the delta: "you picked 6 classes, a usable module needs 11".
Assert axiom from selection	7	6	L	The missing half of the loop: OntoInk can parse, reason, validate and score an ontology but cannot change one axiom from the picture.
Export parity — shared legend geometry, export: key, provenance manifest	6.5	4	M	PNG and SVG take different paths with duplicated legend renderers.
Accessible figure — generated long description + ontology outline	6.5	7	L	The canvas is opaque to assistive technology; generation from the parsed RDF is the only design that ships, and it buys Ctrl+F and search indexing.
Keyboard focus for the canvas — tabindex, focus ring, arrow-walk	6.5	3.5	M	The context menu is keyboard-operable; the canvas itself still cannot take focus.
Selection-scoped SPARQL and SHACL	6	5.5	M	Turn a selection into a VALUES block or a scoped validation run instead of hunting through a 400-entry dropdown.

By creativity

Feature	cre	imp	effort	What makes it new
Change heat — per-IRI git provenance on the diagram	8.5	5	L	Needs the repository, an RDF parser and the renderer in one process — structurally unavailable to browser-only tools.
Layout rules — alignment as a re-derivable claim over a predicate	8	6	L	yEd's layer constraints know nothing about what a class is.
Competency questions by example — selection → SPARQL → CI gate	8	6	L	Themis, CQChecker and SAMOD all require you to write the SPARQL first — the step that stops teams.
Accessible figure / ontology outline	7	6.5	L	A long description derived from the RDF itself; draw.io's outline is a thumbnail, not a semantic tree.
Tidy as taxonomy, continuously	7	5.5	M	The one-shot version ships in 0.7.5; keeping it true as the ontology changes is the interesting part.
ODP capture — mint a pattern by generalising a selection	7	5	XL	XD and CoModIDE assume the pattern already exists.
Ghost axioms — LLM proposals gated by the reasoner and SHACL	7	4.5	XL	Acceptance is mechanically blocked unless the ontology stays consistent and conformant.

Quick wins

Highest value per line of code, all feasible in a day or two:

1. Smell modes in selectBy, plus a "select these" action on each OntoSniff finding
2. prefers-reduced-motion gates across the remaining animation sites
3. tabindex="0" and a focus ring on the canvas wrapper
4. caption: / id: / alt: fence keys with figure numbering (Python only)
5. Manchester, OWL Functional and ROBOT template serialisers next to "copy as TTL"

Known risks

Recorded so they are not rediscovered:

• Position cache — _positionsSave rebuilds the map from the visible nodes only, so an LOD sweep drops coordinates for hidden ones; and the cache key hashes content alone, so two fences built from the same file collide. Both must be fixed before a lockfile is worth shipping.
• Style presets — cy.style() replaces the whole stylesheet, so any overlay (inferred edges, selection halo, a future diff layer) has to be re-appended on every preset switch or it silently disappears.
• Compound nodes — dagre drops edges to parents, positions() has no ancestor guard, and no preset declares the category-box selector. Anything that groups a selection into a container inherits all three.
• Reasoner parity — only the JavaScript OWL-RL path can produce justifications; HermiT, Konclude and the server route return proof-free triples. Any "explain" feature has to show that degradation rather than hide it.