v1.7.0

Architecture-review release (minor). No MCP tool surface changes — still 26 tools, same signatures.

Added

• Driver/dialect registry — single source of truth for db_type → driver + sqlglot dialect
• Parser-based SQL safety gate (analyze_sql_statement) as the primary read-only/single-statement check
• Typed HandlerContext for handlers
• CI enforcing ruff/black/isort/pytest (mypy advisory)

Changed

• OBQC dialect parity — bigquery/duckdb/databricks/mysql now validated with their own dialect instead of the Postgres fallback
• SQL validation catches DML hidden after a CTE and no longer mis-flags ;-in-strings as multiple statements
• Internal refactor: split main.py; decomposed handlers/ontology.py; repo formatted with black + isort

Fixed

• AUTO_ONTOLOGY background generation (was silently producing nothing)
• Artifact downloads honor an explicit schema_name (per-schema state)
• RDF-store auto-restore no longer AttributeErrors on RDF-enabled workspaces
• Startup banner reports the real tool count

PyPI: https://pypi.org/project/orionbelt-analytics/1.7.0/

Full details in CHANGELOG.md.

v1.6.0 — Graph reasoning surface over generated ontology

Added

• Graph reasoning surface over the generated ontology (design/PLAN_graph_reasoning.md):
  • oba:joinsTo shared join predicate. Every declared or inferred many-to-one foreign key now also emits a single directed oba:joinsTo edge between the table classes (finer grain → coarser grain). A single SPARQL property path ?from oba:joinsTo+ ?to answers directed reachability across all FKs without enumerating each per-FK object property. Declared in ontology/oba.ttl (vocabulary bumped 0.1 → 0.2).
  • reachable_from MCP tool — the dimension-capable tables for a query anchored on a table (many-to-one closure): coarser-grain tables joinable without row multiplication, safe to GROUP BY / filter on.
  • measurable_from MCP tool — the measure-capable tables (one-to-many closure, the inverse): finer-grain tables that fan out the anchor and must only be aggregated, never used as dimensions at this grain. Both closures are cycle-safe.
  • plan_composite_query MCP tool — advisory Composite Fact Layer (CFL) decomposition: detects whether the requested facts are independent grains (disjoint siblings) requiring a UNION ALL composite, and returns the leg roots, conformed (shared) GROUP BY dimensions, and per-leg NULL-pad sets. Advisory only — OBA does not compile SQL; defer compilation to OrionBelt Semantic Layer when connected.
  • Optional in-process SHACL validation. When pyshacl is installed, generate_ontology validates the generated ontology against ontology/oba-shacl.ttl and surfaces violations as a non-blocking warning (gated by OBA_SHACL_VALIDATE, default on). Gracefully no-ops when the dependency or shapes file is absent.

Changed

• OBQC fan-trap detection is now grounded in the ontology's own owl:disjointWith axioms (sibling facts sharing a dimension) instead of only re-deriving the risk from the relationship heuristic. When disjoint facts appear together in an aggregating query, OBQC cites the actual disjoint tables and recommends a Composite Fact Layer (UNION ALL). The relationship heuristic is retained as a fallback when no disjointness axioms are present.

v1.5.3

Changed

• Constrained MCP tool string parameters at the input boundary. Enumerated parameters now use Literal types (db_type, cache_type, artifact_type, source, chart_type, chart_style, sort_order, output_format, element_type), free-text and identifier parameters carry max_length bounds, and filename/model-name parameters reject path separators — so invalid, oversized, or path-traversing arguments are rejected before reaching a handler, and the constraints are published in each tool's JSON schema. Handlers keep their existing runtime validation as defense-in-depth. save_semantic_model additionally reduces model_name to a bare filename component so it can never escape the models directory.

Fixed

• Aligned all documentation with the registered tool set. Removed references to four tools that were advertised but never registered (validate_sql_syntax, download_ontology, list_tables_sparql, diagnose_connection_issue) across the README, docs/, Claude skills, and all 8 integration guides, replacing them with registered equivalents (execute_sql_query's built-in validation, download_artifact, query_sparql). Completed docs/tools-reference.md to cover all 23 tools (was 16). Corrected stale tool counts (README 32, startup banner 22 → 23).
• Resolved all pre-existing lint findings (ruff: unused imports, f-strings without placeholders, unused locals, redundant imports; intentional late imports annotated with # noqa: E402). ruff check . is now clean.

Added

• scripts/bump-version.sh — bumps the version across all files, inserts a CHANGELOG stub, and runs uv lock in one step.
• scripts/publish-docker.sh accepts a generic DOCKERHUB_PAT (legacy DOCKERHUB_RALFORION_PAT still works).

📦 PyPI: https://pypi.org/project/orionbelt-analytics/1.5.3/
🐳 Docker: ralforion/orionbelt-analytics:1.5.3 (multi-arch)

v1.5.2

Removed

• get_server_info MCP tool — redundant with the MCP initialize handshake, server instructions, and tools/list. Unique capability descriptions folded into the instructions.

Changed

• Condensed server instructions (~58 → ~30 lines) — dropped the duplicated database list and Version: footer, merged overlapping sections.
• Slimmed generate_chart from 13 → 11 parameters — removed width/height. Interactive charts are responsive; static PNG export uses fixed 800×600. Docstring condensed.

Full Changelog: v1.5.1...v1.5.2

v1.5.1

Fixed

• MCP handshake advertised the wrong version — FastMCP() was constructed without an explicit version, so the initialize response's serverInfo.version fell back to the FastMCP package version (e.g. 3.2.4) instead of the application version. The constructor now receives version=__version__.

Changed

• Upgraded fastmcp[apps] from >=3.2.4 to >=3.3.1,<3.4.

Full Changelog: v1.5.0...v1.5.1

v1.5.0 — MCP sampling for semantic-rename suggestions

Added

• MCP sampling for suggest_semantic_names -- when the connected client advertises the sampling capability, the server now calls back through the host LLM via ctx.sample() to pre-fill rename suggestions for cryptic identifiers, returning a suggestions dict alongside the cryptic-name lists in a single tool call.
  • Gated on a new ENABLE_SAMPLING env flag (default true). Set to false to force the legacy manual-review path everywhere.
  • Clients without sampling support (e.g. Claude Desktop) silently fall back to the legacy response shape — no breaking change.
  • Sampling requests, results, and failures are logged with elapsed time and item counts for observability.

Fixed

• MCP session crash on client disconnect during suggest_semantic_names -- a notification (ctx.info) write that hit anyio.ClosedResourceError because the client had already closed the streamable-HTTP session was caught by the handler's outer except and triggered a second doomed write, bringing down the entire FastMCP TaskGroup. Notifications are now sent through safe_ctx_info (failures swallowed at debug level), and ClosedResourceError-class disconnects re-raise cleanly so the framework tears the session down instead of writing into a dead transport.
• Sampling response parsing -- replaced pydantic-ai's result_type=Dict[str, str] (which forces the model to call an injected final_response tool, fragile on large responses) with explicit JSON parsing that accepts bare JSON, ```json fences, and prose-embedded JSON.

v1.4.4

What's Changed

• Chart storage cleanup: renamed legacy tmp/tmp/ to tmp/charts/ and removed duplicate chart saves — charts are now stored once in tmp/<connectionId>/charts/
• FastMCP structured output fix: create_error_response now returns a dict instead of a JSON string, fixing ValueError: structured_content must be a dict crashes in graphrag_search and other Dict-returning tools
• Improved query_sparql tool description: better guidance for LLM clients on when and how to use the SPARQL tool

Full Changelog: v1.4.3...v1.4.4

v1.4.3

What's Changed

• Inline ontology upload: support loading ontologies directly via chat upload
• Database compatibility check: validate database type before ontology import
• Version sync fix: aligned version across pyproject.toml, src/__init__.py, and README badge

Full Changelog: v1.4.2...v1.4.3

v1.4.2

What's New

Chart Improvements

• JSON chart resource: Registers a chart-json resource alongside the HTML resource, allowing OB Chat to render Plotly charts directly without HTML parsing overhead
• Multi-y bar charts: Support for multiple y_columns in bar charts (grouped/stacked), not just line charts
• CHART_RETURN_BINARY env setting: Control whether inline binary images are included in generate_chart responses (default: false)
• Responsive layout: Charts use responsive sizing with vertical right-side legend and horizontal modebar toolbar with all buttons

Full Changelog

v1.4.1...v1.4.2

v1.4.1

What's Changed

• Renamed analyze_schema → discover_schema across the entire codebase to prevent LLMs from confusing "analyze my data" with schema discovery
• discover_schema returns early when workspace is already restored — no redundant re-discovery
• get_table_details uses cached data from session and discourages bulk usage
• sample_table_data auto-fills schema_name from session when omitted
• execute_sql_query recommends semantic layer (OBML) when available
• Updated all integration examples, docs, skills, and tests