v0.8.1

OntoBricks — Release Notes V0.8.1

Release window: September 2026

Type: Patch release (v0.8.0 → v0.8.1)

Test status: all changes shipped with the non-scenario suite green: 5939 passed, 304 skipped, 6 deselected, 32 warnings.

---

Highlights

• Fail-closed Spark SPARQL (#167) — Explorer warehouse SPARQL is validated before translation. Unsupported algebra is rejected with a capability error instead of a silent or misleading SQL mapping.
• Faster describe_entity / /triples/find (#182) — BFS traversal, distinct payload fetch, ordering, and pagination run in one backend query. Type-wide scans no longer materialize hundreds of thousands of triples in Python. Responses keep exact total / entity_count and add has_more.
• MCP session isolation (#183) — concurrent MCP clients on one server process keep their own selected domain and label/action caches. Resource reads use the same session scope as tool calls.
• Admin-only domain creation — only Databricks App administrators (CAN_MANAGE) can create a new domain. Editors and Builders still save and work on assigned domains.
• SWRL inference filters — rules that use string built-ins such as swrlb:contains now apply SQL predicates during Run Inference. Active-contract “owns meter” style rules return inferred triples instead of empty results.
• Data Quality condition builder (#186) — selecting a property after the target entity is chosen no longer clears the property list or drops the selected value. Guarding a conformance/consistency rule with an IF condition can be completed and saved.
• License alignment (#169) — product, API, and OntoViz surfaces state the Databricks License as the first-party license.
• Dependabot security floors — PyJWT, GitPython, urllib3, anyio, cryptography, MLflow, and sentence-transformers raised to patched releases. oauthlib 4.x and the remaining NLTK pathsec advisory are blocked upstream (see Security).

No registry schema migration. No breaking MCP tool names.

---

SPARQL (Spark / Lakehouse Explorer)

Spark SPARQL is fail-closed. The translator calls SparqlCapabilityValidator before generating warehouse SQL. Local RDFLib execution is not a fallback for warehouse data.

Supported subset (warehouse):

• Basic SELECT graph patterns
• LIMIT
• OPTIONAL / LeftJoin
• Literal BIND
• String FILTER (CONTAINS, STR equality, STRSTARTS, STRENDS)
• FILTER(?predicate | ?p | ?pred IN (…)) with URI lists
• The specialized relationship-UNION shape: projected ?subject ?predicate ?object, branch variables ?subject / ?object, and BIND(… AS ?predicate)

Rejected (examples): GROUP BY, ORDER BY, OFFSET, numeric or arbitrary FILTER, property paths, subqueries, MINUS, VALUES, SERVICE, GRAPH, non-specialized UNION.

Non-specialized UNION fails at capability validation with a UNION error rather than a later unrelated mapping failure.

---

Knowledge Graph (describe_entity / /triples/find)

DigitalTwin.find_triples_bfs now delegates to a folded backend page (find_triples_bfs_page) instead of walking the graph, fetching every triple, then slicing in Python.

• SQL backends use a sargable bidirectional-edge walk and return one deterministic page (DISTINCT, ORDER BY, LIMIT/OFFSET plus one extra row for has_more).
• Neo4j implements the same page contract (walk from structured seeds, de-dup, slice) so metadata stays aligned: seed_count, entity_count, total, has_more.
• Public and internal /triples/find responses keep existing fields and add has_more. Missing has_more defaults to false for older payloads.
• Agent and MCP formatters show exact totals ({count} of {total}) and only hint that more rows exist when has_more is true.
• MCP describe_entity guidance: prefer entity-scoped deep scans via search; keep depth=1 for type-wide scans.

Provenance: originally contributed by Laurent Prat (#182). Live warehouse measurement on a type-wide depth-2 Counterparty scan went from ~263s of Python materialization to ~34s for the first 100-triple page.

---

MCP Server

Per-connection selected domain

The MCP process still binds one MCPServerSession (shared registry, policy, HTTP pool). selected_domain_name, ontology_labels, and class_actions are now keyed by MCP session ID (StreamableHTTP mcp-session-id).

• Every MCP request (tools and resources such as ontobricks://status, ontobricks://stats, ontobricks://graphql-schema) resolves that session’s domain.
• The in-process store keeps at most 512 recently used sessions; an evicted client must call select_domain again.
• Off-request / missing session ID uses a shared compatibility bucket and logs a warning once per process.
• State is process-local. A multi-worker deployment would need a shared store keyed by the same session ID.

Provenance: originally contributed by Laurent Prat (#183).

---

Ontology, Rules, and Inference

Data Quality conditions (#186)

Delegated condition-row handlers stay bound to the latest render options. An empty first render no longer poisons later property selections, so the IF-condition property dropdown keeps the chosen attribute and guarded rules can be saved.

SWRL Run Inference

build_inference_sql no longer treats swrlb:* atoms as graph properties. Built-ins are excluded from triple joins and applied as SQL filters after variables are bound (for example object LIKE CONCAT('%', 'active', '%') for swrlb:contains).

Re-run affected active rules after upgrade; previously empty inference results for status-contains filters should now populate.

---

Registry, Permissions, and UI

• New Domain (Home, Registry, empty state) is gated to administrators.
• The API distinguishes a new domain from an existing-domain save and rejects non-admin creation.
• Help Center and the development role matrix document Admin-only creation and the assigned-domain workflow for Editors and Builders.

---

License and metadata

GitHub issue #169 closed conflicting MIT / Apache 2.0 / Databricks License claims on first-party surfaces.

• README, product, architecture, development, and deployment docs identify OntoBricks (and OntoViz) as source-available under the Databricks License.
• OpenAPI schemas (main and external) publish matching license metadata.
• OntoViz header and About license link point at the canonical Databricks License.
• Third-party and vendored notices are unchanged.

---

Security (Dependabot)

Raised uv constraint floors and relocked uv.lock / src/mcp-server/uv.lock against patched releases:

Package	Floor	Notes
PyJWT	≥2.15.0	HMAC/JWK confusion series, RecursionError, options-dict mutation
GitPython	≥3.1.62	submodule path traversal, ReDoS, git-dir impersonation
urllib3	≥2.8.0	proxy TLS override, unbounded chunk-size, Deflate loop
anyio	≥4.14.2	TLSStream IDNA 2003 spoofing, process-pool stderr stall
cryptography	≥50.0.0	PKCS#7 oracle + path-building; now allowed by MLflow 3.16.1 (cryptography<51)
MLflow	≥3.16.1	outside AI Gateway SSRF range (3.13.0–3.15.2)
sentence-transformers	≥5.6.0	trust_remote_code bypass on local model load (pitfalls extra)

Not patched (upstream):

• oauthlib 4.0.0 — databricks-sql-connector 4.4–4.6 still requires oauthlib>=3.1.0,<4.0.0.
• NLTK GHSA-8mgp-746c-j5xp — latest published is 3.10.3; no patched release. Pitfalls extra only.

---

Documentation and operator notes

• Architecture and user-guide Explorer sections describe the Spark SPARQL support boundary.
• API, MCP, and Graph Explorer query-card docs describe folded BFS paging, exact totals, additive has_more, and depth=1 for broad scans.
• MCP docs replace the old “one selected domain per process” limitation with concurrent-session behaviour and the 512-session LRU cap.
• Contributor Guide (docs/contributing.md) is the canonical setup / tests / PR workflow. Repo-root CONTRIBUTING.md is a short pointer. README, Help Center index, development, and code-organization link there.
• Agent/git conventions: there is no develop branch. Active work uses a version-named branch matching pyproject.toml (next line is 0.9.0). Feature branches merge into that version branch; it ships to master.

---

Upgrade notes

• SPARQL: queries that previously appeared to “work” with unsupported Spark constructs now fail closed. Rewrite them to the supported subset or run them outside warehouse SPARQL.
• /triples/find and describe_entity: clients can keep using total / entity_count. Treat has_more as additive. Broad type-wide scans should start at depth=1; use search for entity-scoped deeper walks.
• MCP: two clients on one mcp-ontobricks process can now select different domains. After LRU eviction (512 sessions) the client must select_domain again. Multi-worker MCP still needs an external session store.
• Domain creation: grant CAN_MANAGE on the Databricks App to users who must create domains. Existing Editor/Builder assignments are enough to save work on domains they already have.
• Inference: re-run SWRL rules that use swrlb:contains (and other registered string/comparison built-ins). No ontology rewrite is required if the rule text is already valid SWRL.
• License: downstream packaging and legal review should treat the Databricks License as authoritative for OntoBricks first-party code.
• Lakebase / MCP policy: the v0.8.0 domains.mcp_policy migration remains a prerequisite if you are still on 0.7.x. This patch does not add columns.

---

Bug Fixes (selected)

• Fixed SWRL inference joining swrlb:contains as predicate = '…#contains' (no triples matched).
• Fixed Data Quality IF-condition property dropdown emptying after entity selection (#186).
• Fixed non-admin users being able to persist a new domain through save/create paths.
• Fixed Spark SPARQL accepting or mis-mapping unsupported UNION / FILTER / GROUP BY shapes.
• Fixed first-party license strings disagreeing across README, OpenAPI, and OntoViz.
• Fixed concurrent MCP clients clobbering each other’s selected domain (markets vs trading entity counts / ontology).
• Fixed MCP resource reads (ontobricks://status, stats, GraphQL schema) still sharing the default domain bucket after tool-call isolation.
• Fixed describe_entity / /triples/find type-wide scans timing out or returning 502s from unbounded Python triple materialization.
• Fixed missing has_more on older find payloads defaulting unsafely in internal/public route mapping.