What shipped, and when.

2026-06-18 — SQL Server types, AI cost accuracy, row-count integrity

• 01SQL Server DATETIMEOFFSET and TIME columns now read correctly. ODBC output converters are registered on every MSSQL read, so those columns come back as timezone-aware datetimes and times. If a still-unhandled ODBC type is hit, the error names the column and tells you how to cast it in query mode.
• 02AI cost estimates are no longer overstated for cheap models: cost resolves the per-model rate rather than the provider tier, including dated and namespaced model ids.
• 03Fewer tokens per Copilot turn: tool schemas are narrowed per request on cloud providers as well as local ones.
• 04New Steward detector, ROW_COUNT_DELTA: a step whose contract is strictly one-to-one but silently changed its row count now raises a node-level finding. Observe-only and conservative — filter, join and aggregate steps are never flagged.

2026-06-17 — AI keys in Credentials, Needs Attention clarity

• 01An AI provider's API key can now live in the central Credentials store, with the same governance as every other secret: encryption at rest, expiry, vault source, audit and used-by. The AI Provider form offers a saved-credential picker; the key is resolved at request time and never copied into the AI config.
• 02Schema v31 → v32: an additive credential_id column on user and workspace AI config, with idempotent upgrade.
• 03Dashboard → Needs Attention now shows current state only — a pipeline appears when its most recent run failed and drops off when it runs clean — with a last-failed timestamp, a consecutive-failure badge, a per-row Clear action and a deep link to the actual failed run.

2026-06-08 — External-review follow-ups

• 01Steward foreseer: automatic volume-anomaly detection (VOLUME_ANOMALY). Learns each source's normal volume from its own history using a robust median/MAD baseline and flags drops and spikes against that baseline, never against an absolute threshold.
• 02CI runs the backend test suite in parallel with pytest-xdist: roughly twelve minutes serial became about seventy seconds across fourteen workers.
• 03Frontend bundle split: React and the canvas library moved into long-cached vendor chunks, cutting the entry chunk by roughly half.

2026-06-08 — 1.2 reliability foundations: lineage, executor, backfill

• 01Lineage: runtime step-run lineage events with an API; an OpenLineage formatter and append-only exporter; an HTTP exporter for Marquez and DataHub with bounded retry that never fails a run; and opt-in consumer self-attestation.
• 02Executor: a failure-class taxonomy stamped on every failed step; an optional per-step retry policy that short-circuits non-retryable failure classes; and a cancellation-token foundation.
• 03Backfill: a lookback window on incremental sync for late-arriving data; a merge write mode with merge key and tombstone column on the warehouse sink; resume-from-window for interrupted backfills; and soft-delete propagation foundations.
• 04Around 260 new unit and integration tests, with source-grep guards pinning each executor and sink wire-in so an integration point cannot be silently removed.

2026-06-06 — UX polish, real-data proof, gap-closure pack

• 01Five workstreams shipped together after reviewer feedback converged on them: interface polish, proof against real data, and closure of the remaining V1 gaps.