Roadmap¶
The roadmap is ordered by safety and integration value, not by a fixed delivery date.
Implemented For 0.7.0¶
- Lightweight base installation with explicit
parquet,mcp,trino, andallextras. - Review-first allowlisted Trino planning through MCP without raw-SQL access on the default tool surface.
- HMAC-authenticated, metadata-only MCP audit records with integrity verification and bounded fail-closed storage.
- An explicit compatibility and deprecation policy for future DatasetSpec
schema_versionrevisions.
Implemented For 0.6.0¶
- One reviewed
DatasetSpeccontract across CLI and Python generation and validation workflows. - Removal of the deprecated parallel specification API and conversion layer.
- Consistent
dataset_spec.jsonanddataset_spec.yamlartifact names. - A smaller project README backed by the published task-oriented documentation site.
Implemented For 0.5.1¶
- Tokenless PyPI publication with post-publish digest comparison and a clean public-index installation smoke test.
- OpenSSF Scorecard reporting and expanded standard project links on PyPI.
Implemented For 0.5.0¶
- Structured generator MCP business-rule inputs with strict contract checks, sensitive-literal rejection, bounded formulas, manifest fingerprints, and compact validation summaries.
- Typed package metadata, strict core/rules/MCP type checks, installed-wheel smoke coverage, and pull-request dependency review.
Completed In 0.4.0¶
- Content-aware PII and secret detection across CSV, Trino, and imported profile trust boundaries.
- Configurable input, output-size, disk-reserve, execution-time, and Trino server-side query budgets.
- Locked dependencies, vulnerability auditing, CodeQL, full-history secret scanning, and live Trino integration coverage.
- Reproducible release artifacts with CycloneDX SBOMs, checksums, provenance, and SBOM attestations.
Completed In 0.3.0¶
- Full MCP workflow for safe CSV profiling, spec inference, deterministic generation, validation, and export.
- Direct safe-profile handoff between Trino and generator MCP tools.
- Versioned DatasetSpec contract and auditable generation manifests.
- Runtime raw-profile and source-row reuse protections.
- End-to-end AI integration demo.
- Review-first agent orchestration with
agent-planandagent-approve.
MVP Freeze¶
Keep the current MVP focused on the reliable golden path:
- CSV file or CSV folder input.
- Safe profile metadata, not source rows.
- Reviewable
DatasetSpec. - Deterministic generation by explicit seed.
- CSV, JSON, and Parquet export.
- Validation report and generation manifest.
- MCP responses that return summaries and artifact paths, not dataset rows.
Treat these as non-negotiable release gates rather than new feature work:
ruff,compileall, and the full pytest suite pass.- Quickstart commands in README run against checked-in fixtures.
- Generated manifests report
synthetic: trueandsource_rows_copied: false. - OpenSpec baseline stays aligned with implemented behavior.
Later¶
- Pluggable synthetic providers for organization-specific semantic types.
- More cross-table aggregate constraints and controlled negative scenarios.
- Packaging and deployment examples for isolated MCP workers.