Skip to content

Release Notes -- 2026-04-30

Databricks Forge v0.40.0 → v0.41.0

v0.40.1 -- Opus 4.7 compatibility + max_tokens defaults

Bug Fixes

  • Opus 4.7 temperature 400 -- databricks-claude-opus-4-7 now rejects requests that include the temperature parameter with HTTP 400. Added a per-model supportsTemperature capability (only Opus 4.7 is false) and gated the wire body in lib/dbx/model-serving.ts so all callers can keep passing temperature without changes; the field is simply omitted for models that disallow it.
  • Claude Sonnet 4 silent 1K truncation -- Databricks defaults max_tokens to 1,000 for Claude Sonnet 4 when the field is omitted, causing silent finish_reason: length truncation. The wire layer now always emits max_tokens, falling back to a per-model defaultMaxTokens when the caller doesn't specify one.

Improvements

  • Gemini Flash output ceiling raised -- databricks-gemini-3-flash and databricks-gemini-3-1-flash-lite maxOutputTokens raised from 8,192 to 32,768 to match Databricks' published per-request envelope.
  • Visible clamp warnings -- The maxTokens clamp log was promoted from info to warn so unintended truncation (e.g. the 128K asks in lib/pipeline/steps/scoring.ts against a 32K Claude cap) surfaces in normal log views.
  • Per-model defaultMaxTokens -- Added a sensible default per model: GPT-5.4 = 16,384; Claude Opus/Sonnet 4.x = 8,192; Gemini Flash + Llama Maverick = 4,096. Surfaced via getModelCapabilities() and logged in the pool startup summary.

Other Changes

  • Extended ModelTemplate, ModelEndpoint, getModelCapabilities(), and UNKNOWN_CAPS with the two new fields.
  • Updated __tests__/model-pool/model-registry.test.ts (now 27 tests in the model-pool suite, 922 across the full repo) and the __tests__/model-serving/content-parsing.test.ts mock to match the new getModelCapabilities shape.

v0.41.0 -- WAF Assessment

New Features

Self-service Databricks Well-Architected Framework assessment

A new /assessment page lets customers score their workspace against the Databricks Well-Architected Framework (WAF) without leaving Forge. Forge runs deterministic SQL over system.* (lineage, audit, billing, compute, information_schema) on behalf of the logged-in user (OBO) and produces:

  1. Per-pillar scores -- Governance, Reliability, Cost Optimisation, Performance Efficiency, plus an overall score derived from the available pillars.
  2. Drill-down per pillar -- Each pillar has its own tab with a score header (Mature / Progressing / At Risk / Critical) and a control table sorted failing-first.
  3. Failing controls tab -- A cross-pillar shortcut to the controls that did not meet their threshold, ordered by score ascending.
  4. History -- Every run is persisted in Lakebase (forge_waf_assessments) with timestamps, scores, and triggering metadata.

The catalog of 30 controls (sourced from Databricks-WAF-Light-Tooling) is bundled as a CSV, upserted into forge_waf_controls on first boot, and re-synced on every run.

Each failing control surfaces a one-click action mapped to the right destination:

  • Forge engines (Wrench icon) for controls genuinely automatable inside Forge -- Comment Engine for missing column comments, Estate Scan for lineage / tag / format / managed-table inventory.
  • Open docs (Book icon, opens in new tab) for admin- or workload-level controls (Photon, autoscaling, cluster policies, model serving, audit-log enablement, etc.) where the canonical Databricks doc is the right next step.
  • Ask Forge fallback when no mapping is curated.

All 30 controls now have at least one curated next step.

CSV export

"Export CSV" button on the assessment toolbar generates a client-side CSV (waf-assessment-YYYY-MM-DD.csv) with waf_id, pillar, principle, best_practice, score, threshold, met, recommendation -- ready to drop into a customer report.

Drift comparison between runs

A new /assessment/compare?from=X&to=Y view diffs two assessments and groups controls by drift category:

  • Regressed -- was met, now not met (highlighted in red, top of page).
  • Improved -- was not met, now met.
  • Score moved -- same status, score delta beyond a 5-point noise floor.
  • Catalog changes -- controls that entered or left the catalog between runs.
  • Unchanged -- collapsed by default.

A "vs. latest" button on each row of the History tab launches the comparison.

API

  • POST /api/assessment/run -- triggers a full assessment run, persists results, returns the summary.
  • GET /api/assessment -- returns latest assessment (with control results) plus history list.
  • GET /api/assessment/[assessmentId] -- returns a single assessment with its per-control results.
  • GET /api/assessment/controls -- returns the bundled controls catalog (no run required).

Schema

Three new Lakebase tables (auto-migrated via prisma db push on boot):

  • forge_waf_controls -- catalog of WAF controls with thresholds, metric definitions, and curated remediation text + fix-action mapping.
  • forge_waf_assessments -- one row per assessment run with status, per-pillar scores, overall score, met/total counts, error message, timestamps.
  • forge_waf_control_results -- per-control result rows joined to a parent assessment via FK.

All Commits

Hash Summary
3447c1b fix(model-serving): make Opus 4.7 work + close Sonnet 4 max_tokens footgun
fill at merge feat: WAF Assessment module with drill-down, CSV export, and drift compare