feat(openontology): Phase 2 + 3 — storage, REST/SSE, MCP, RDF/SHACL, adapters, TUI, explorer

[ralyodio]

Everything the two shipped PRD phases deferred, minus what's called out at the bottom. 53 files, ~6.9k lines, 76 new tests.

Storage — Turso/libSQL

createLibsqlStore hydrates the read model at open, serves reads synchronously (a query evaluator that awaits per triple pattern is unusable), and buffers mutations as SQL that flush() writes in one transaction. Versioned idempotent migrations; indexes over subject, predicate, entity-valued object, status, both time axes, aliases, external ids; FTS5 for label/alias search.

The append-only status log is replayed on open, so retractions, supersessions, and merge redirects survive a reopen — tested by closing and reopening a real database file.

REST + SSE + OpenAPI

16 paths under /api/ontologies, described at /api/ontologies/openapi, referencing the published JSON Schemas rather than restating them.

• No token → read-only. Curator token → can apply. Agent token → can propose, cannot apply.
• Idempotency-Key on mutations, revision ETags, 409 on a stale base revision
• SSE stream emitting the same event objects as the JSON endpoint

MCP

OpenOntology and OpenPRD on the standards server: 4 + 2 resources, 11 + 6 tools, 5 + 2 prompts. Read-only by default; OPENONTOLOGY_MCP_WRITABLE=1 buys proposals and never applies — the denial is the shared policy layer, not a second rule that could drift.

Interoperability

RDF/Turtle export and import of the reified profile, plus the plain triple for asserted relationships so a consumer wanting only the accepted graph gets one.

SHACL for 5 of 7 constraint kinds. unique and query are reported as unmapped in both the return value and the generated Turtle — a shape that quietly means something narrower is worse than no shape.

Source adapters

CSV, JSON, YAML, NDJSON, Markdown, generic JSON HTTP, GitHub. All produce proposed operations with source, evidence selector, run id, and confidence attached; fetch is injected so ingestion is offline and testable. Each declares its capabilities — so "nothing was deleted upstream" is never confused with "this adapter cannot see deletions." None of the seven can.

TUI + explorer

Keyboard-first panels as plain strings that survive SSH and 60 columns; status is a glyph and a word, never colour alone; the key bar wraps rather than truncating. Wired as logicsrc ontology tui.

Read-only web explorer at /openontology/explore with entity and claim views showing status, both clocks, confidence, sources, evidence, and append-only history. Plus an /openprd page for the companion standard.

Two real bugs the tests caught

1. The API built a new engine per request, so explain could never find a resultId from a prior request — every explain was a 404. Engines are now cached per role over the shared store.
2. The TUI status bar called engine.validateOntologyPackage(), appending a package.validated event on every repaint. A read-only view must not write to the audit log; it now uses the pure validator.

Verification

• 76 new tests; 527 total across the monorepo, all passing; full build green
• libSQL exercised against real database files, the API through its route handlers, MCP over an in-memory transport

Not included

PWA review/approval write flows (they need an auth story this deployment doesn't have), OWL/RDFS mappings, SPARQL/Cypher/Datalog query adapters, and Phase 4 governed actions. The compatibility matrix marks those "planned", not "supported".

🤖 Generated with Claude Code

[github-actions]

vu1nz Security Review

0 finding(s) in PR #?

No security issues found.