Skip to content

Release Notes -- 2026-03-13

Databricks Forge v0.29.0 → v0.31.2

v0.29.0 -- Async Genie Builds from Ask Forge

New Features

Async Fire-and-Forget Genie Space Builds from Ask Forge

Ask Forge Genie Space creation is now fully asynchronous, eliminating SSE/upstream request timeouts. When a user triggers "Create Genie Space" from Ask Forge, the build starts in the background and the user can continue their conversation immediately. A persistent progress toast in the bottom-right shows real-time build status with cancel, retry, and deploy actions. Completed builds can be deployed directly from the toast or from Genie Studio.

Build Progress Toast with Cancel Support

A new persistent sonner toast component tracks Genie Space builds across the application. The toast displays a progress bar, current step message, and percentage. Users can cancel running builds at any time. On completion, the toast transitions to a deploy-ready state with one-click deployment or a link to Genie Studio.

Genie Deploy Dialog

A lightweight deploy dialog extracted from the original GenieBuilderModal provides quality gate enforcement, metric view schema selection, and a streamlined deploy flow. Reuses shared deploy logic for consistent behavior across all build surfaces.

Source Badges on Genie Studio Job Tiles

Build jobs in Genie Studio now display their origin (Ask Forge, Schema Scan, or Requirements) as a compact badge. Completed jobs also show whether they have been deployed, and conversation context is displayed when available.

Build-in-Progress Sidebar Indicator

The Genie Studio navigation item shows a pulsing violet dot when any Genie Space build is actively running, giving users visibility across the entire application.

Page Refresh Resilience

Active build job IDs are persisted in localStorage. When the page reloads, the layout-level provider restores progress toasts for any still-running or completed-but-undeployed builds.

Improvements

AbortSignal Threading in Fast Genie Engine

runFastGenieEngine now correctly destructures and threads AbortSignal through all four async LLM calls (expressions, instructions, example queries, title generation) with step-boundary abort checks. This fixes a bug where cancelling a fast build had no effect.

Shared Deploy Utilities

Deploy logic (API call, polling, quality gate, metric view payload, embeddings backfill) has been extracted into lib/genie/deploy-utils.ts with both a pure function (deployGenieSpace) and a React hook (useGenieDeploy). The PATCH /api/genie-spaces/generate endpoint allows writing deployedSpaceId back to the job store for cross-surface coordination.

Unified Build Job Provider

A new GenieBuildProvider React context replaces local polling in the Genie Studio page with a shared useGenieBuildJobs hook. This eliminates duplicate polling and provides consistent job state across Ask Forge, Genie Studio, and the sidebar nav.

Progress Callbacks in Fast Engine

runFastGenieEngine now reports progress via onProgress callbacks at each step (metadata scraping, expression generation, instruction generation, query generation, assembly, quality checks), enabling real-time progress feedback in the async toast.

v0.29.1 -- Scoped Structured Logging

Improvements

Scoped Structured Logging

Introduced createScopedLogger factory that produces loggers carrying immutable structured context -- origin, task, module, fn, runId, errorCategory, and phase -- through every log entry. Replaced the flat singleton logger across 51 files covering the full stack:

  • Pipeline Engine -- root logger per run with per-step children carrying task and module; automatic lifecycle start/end entries with durationMs via the logStep wrapper.
  • All 10 pipeline steps -- use ctx.logger with fn and errorCategory on every warn/error; removed manual [step-name] prefixes and redundant { runId } metadata.
  • Genie, Comment, Dashboard, and Estate Scan engines -- orchestrators create scoped loggers with origin set to GenieEngine, CommentEngine, DashboardEngine, or EstateScan; sub-passes receive child loggers.
  • 9 high-priority API routes -- use apiLogger(route, method, { runId }) for request-scoped logging.
  • 8 infrastructure modules -- prisma.ts, provision.ts, embed-pipeline.ts, assistant/engine.ts, context-builder.ts, rate-limiter.ts, sql-reviewer.ts, llm-cache.ts.

Key capabilities: - child(extra) -- nest loggers by adding task/module context without losing parent context. - timed(message, work) -- auto-emit paired start/end entries with durationMs. - Error categories -- standardised errorCategory values (llm_timeout, sql_hallucination, schema_validation, auth, db, network, etc.) for pattern matching. - Dev breadcrumbs -- human-readable origin > task > fn | message (phase) format with correlation ID tags. - Production JSON -- every entry includes origin, task, module, fn, runId, phase, errorCategory for structured querying. - Backward compatible -- the singleton logger continues to work; unmigrated files function unchanged.

Updated Logger Port Interface

ScopedLogger extends Logger with child(), timed(), and context -- any code typed Logger accepts a ScopedLogger without changes.

Other Changes

  • Added optional logger field to PipelineContext for scoped logger injection.
  • Updated 9 test files to include createScopedLogger in logger mocks.

v0.29.2 -- Metric View Quality & Review Performance

Bug Fixes

  • Metric view placeholder catalog references -- The LLM sometimes copied the literal catalog.schema placeholder from the prompt examples into generated DDL, causing NO_SUCH_CATALOG_EXCEPTION dry-run failures. The prompt now injects the real catalog.schema scope, and a post-processing safety net replaces any residual placeholders.
  • Trusted-asset batch review parse failures -- Batch review responses for trusted assets failed to parse 100% of the time due to output truncation and minimal JSON parsing. Replaced JSON.parse with parseLLMJson (handles truncation recovery, preamble stripping, malformed JSON) and doubled maxTokens per item when requestFix is true.
  • Metric view dry-run timeout risk -- dryRunMetricViewDdl used default SQL timeouts allowing up to ~10 minutes of polling per proposal. Now uses explicit 30s waitTimeout and 35s submitTimeoutMs to prevent domain-blocking hangs.

Improvements

SQL Review 429 Overflow Routing

When the sql-tier review endpoint (databricks-gpt-5-4) enters 429 backoff, the task router now overflows to any unblocked pool endpoint instead of queuing behind the 60s circuit breaker. This allows Claude models to absorb review load during rate-limit storms.

Global SQL Review Concurrency Limiter

Added a cross-domain concurrency cap (4 concurrent) on all reviewChatCompletion calls. With 8+ domains running reviews in parallel, this prevents burst pressure on the review endpoint and reduces 429 activations.

Post-Dry-Run LLM Repair Loop

Metric view proposals that pass static validation but fail dry-run with repairable errors (e.g. METRIC_VIEW_INVALID_VIEW_DEFINITION, UNRESOLVED_COLUMN) now get a second-chance LLM repair attempt. The repaired DDL is re-validated and re-dry-run before acceptance.

v0.30.0 -- Genie Engine Performance Overhaul

New Features

Hierarchical Domain Progress UI

Added a rich, collapsible per-domain progress panel to the Genie Engine workbench. Each domain shows its current phase (expressions, joins, assets, assembly), a thin progress bar, status icon, and elapsed time. The top-level bar remains for overall progress while the detail panel provides granular visibility into long-running generation jobs.

Improvements

Genie Engine Performance Overhaul

Dramatic runtime reduction targeting sub-5 minutes (from ~17 minutes) through six optimization strategies: - Parallel subdomain processing: Metric view generation now runs up to 3 subdomains concurrently via mapWithConcurrency instead of sequentially. - Parallel dry-run validation: All metric view dry-runs execute concurrently via Promise.all instead of one-by-one. - Smarter model routing: Opus models restricted to reasoning-only tier; semantic expressions downgraded to classification tier for faster inference. - Reduced SQL review overhead: Metric view and join inference reviews disabled by default (validated by YAML parser + dry-run instead); SQL_REVIEW_CONCURRENCY bumped from 4 to 6. - Lower token budgets: maxTokens reduced across all passes -- semantic expressions (8K to 4K), trusted assets (32K to 12K), benchmarks (32K to 8K), metric views (32K to 16K), column intelligence (32K to 12K). - Skip metric view planning: New skipMetricViewPlanning config option (default: true) eliminates an LLM call per domain.

Empty LLM Response Resilience

  • Empty content from Model Serving is now detected inside the retry loop, triggering genuine retries with exponential backoff (up to 3 attempts).
  • Empty responses are never cached, preventing cache poisoning that caused downstream parse failures.
  • Diagnostic logging added in model-serving.ts and sql-reviewer.ts for traceability.
  • Review surface env var is restored in a finally block to prevent cross-feature side effects.

v0.31.0 -- Quality Presets + Performance Model Bundle

New Features

Quality Preset System

User-configurable Speed / Balanced / Premium presets that control the Genie Engine's speed-vs-richness trade-off. Each preset defines a GenerationBudget governing target counts for measures, filters, dimensions, trusted assets, and benchmarks, along with domain concurrency, review surface enablement, and maxTokens per LLM call. The default preset is Balanced. Accessible from the Settings page under the Genie Engine section.

Performance Model Bundle

Three new high-throughput models added to the model registry: databricks-gemini-3-1-flash-lite (priority 0, classification + lightweight), databricks-llama-4-maverick (priority 0, generation + classification), and databricks-gemini-3-flash (priority 1, generation + classification + lightweight). A new DATABRICKS_SERVING_ENDPOINT_LIGHTWEIGHT env var and --lightweight-endpoint deploy flag bind the 7th model slot.

Improvements

Budget-Driven Generation

All Genie Engine passes (semantic expressions, trusted assets, benchmarks) now accept budget parameters instead of using hardcoded target counts and maxTokens. Worker A/B/C prompts, use case caps, and benchmarks-per-batch are all driven by the resolved budget.

Right-Sized Output

Speed preset generates lean spaces (4 measures A / 2 measures B / 4 filters / 4 dimensions, 3 concurrent domains, metric views disabled). Premium produces maximum richness (15/10/12/12, 10 concurrent domains, all review surfaces enabled). Balanced sits in the middle as the recommended default.

Cross-Surface Wiring

Quality preset flows through the pipeline Genie workbench, ad-hoc Genie engine (Scan Schema, Upload Requirements), and Ask Forge conversational builder. All surfaces respect the user's chosen preset.

Bug Fixes

  • Metric view dry-run collision -- Parallel dry-runs could create TABLE_OR_VIEW_ALREADY_EXISTS errors due to identical temp table names. Fixed by appending a UUID suffix to the temp table name.
  • SQL review maxTokens cap -- Batch review calls were capped at 16,384 tokens, causing truncation on large batches. Increased to 32,768.

Other Changes

  • Deploy script updated with --lightweight-endpoint flag, resource binding, env var injection, and success banner.
  • Model registry tests expanded with 5 new test cases for performance bundle models (15/15 passing).
  • AGENTS.md updated to document quality presets, GenerationBudget, performance bundle, and the lightweight endpoint deploy flag.

v0.31.1 -- Model Compatibility Guards

Bug Fixes

  • Reasoning model content parsing -- Fixed e.replace is not a function and e.trim is not a function crashes when Gemini 3 models return array-of-blocks content (reasoning + text blocks) instead of a plain string. New extractContentText() safely extracts text from both formats.

  • JSON mode 400 errors -- Replaced string-matching supportsJsonResponseFormat() with a registry-backed lookup. Only GPT-5.x models (verified supportsJsonMode: true) now receive response_format: json_object. Gemini, Llama, and Claude models no longer trigger 400 BAD_REQUEST from unsupported JSON mode requests.

  • maxTokens exceeds model limit -- Added automatic token clamping in both chatCompletion and chatCompletionStream. Requests for 12288 tokens to Gemini Flash Lite (max 8192) are now silently clamped instead of returning 400 BAD_REQUEST.

  • GPT-5.4 empty responses (finishReason=length) -- Doubled the SQL reviewer per-item token budget (1024->2048 review, 2048->4096 fix) and added a 16384 token floor. GPT-5.4's hidden reasoning tokens no longer exhaust the budget before producing visible output.

  • Unknown model rejection -- templateFor() now returns null for unrecognised models, and buildPool() skips them with a warning log. Prevents undefined-capability models from entering the pool with permissive defaults.

  • Codex model removal -- Removed databricks-gpt-5-3-codex from KNOWN_MODELS. All Codex models require the Responses API (AI Gateway beta only) and are not supported until GA.

Improvements

Per-call model observability

Every chatCompletion call now logs the endpoint, requested maxTokens, JSON mode, and model capabilities. resolveEndpoint logs tier, chosen endpoint, and resolution source at debug level. All Genie Engine passes (semantic expressions, trusted assets, benchmarks, column intelligence, instructions) include the endpoint in their status logs.

Verified model capability registry

KNOWN_MODELS in model-registry.ts now includes supportsJsonMode and maxOutputTokens for every supported model, sourced from Databricks documentation. New getModelCapabilities() public API enables any module to query model limits. Pool startup summary now includes JSON mode support and max output tokens per endpoint.

v0.31.2 -- Progressive 429 Backoff + 8K Token Optimization

Improvements

Progressive 429 Backoff (10s / 20s / 30s)

Replaced the flat 60-second rate-limit backoff with a progressive escalation ladder: first 429 triggers a 10-second pause, second escalates to 20 seconds, and third+ caps at 30 seconds. Successful calls immediately reset the counter and lift the block. Saves ~50 seconds per rate-limit incident.

Gemini 3.1 Flash Lite Promoted to Generation Tier

Flash Lite is now the preferred model for all generation tasks (trusted assets, metric views, benchmarks, example queries) in addition to classification and lightweight tiers. Llama Maverick is demoted to overflow/backup (priority 1). Based on testing, Flash Lite is both faster and higher quality.

Trusted Asset Prompt Rework for 8K Models

Condensed the trusted assets system prompt by ~50%, reduced batch size from 3 to 2 use cases per LLM call, and lowered default maxTokens from 12,288 to 6,144. Each batch now produces 1-2 complete parameterized queries that fit comfortably within 8K model output limits, eliminating truncated JSON.

Metric View YAML-Only Output

The LLM now returns only the YAML body for metric view proposals; the DDL wrapper (CREATE OR REPLACE VIEW ... WITH METRICS LANGUAGE YAML AS $$ ... $$) is built programmatically via the new buildMetricViewDdl() helper. This halves output token usage per proposal. The YAML spec reference embedded in the prompt was condensed from ~140 lines to ~30 lines. Max proposals reduced from 3 to 2 per subdomain.

Lower Concurrency Caps for Rate-Limit-Prone Models

Reduced max concurrent requests: Llama Maverick 10→6, GPT-5.4 6→4, Gemini Flash 8→8 (unchanged). Prevents 429 storms on pay-per-token endpoints with tight per-second output token limits.

Quality Preset Token Budget Reductions

All three quality presets now use token budgets that fit within the 8,192-token output ceiling of the majority of pool models: - Balanced: trustedAssets 12,288→6,144, benchmarks 6,144→4,096 - Speed: trustedAssets 8,192→4,096, benchmarks 4,096→3,072 - Premium: trustedAssets 32,768→8,192, benchmarks 8,192→6,144

Column Intelligence and Benchmark Default maxTokens Lowered

Hardcoded maxTokens in column intelligence (12,288→6,144) and benchmark generation (8,192→6,144) now leave headroom within 8K model limits. These defaults are used by the ad-hoc Genie Engine which bypasses quality preset budgets.

429 Diagnostics: Raw Retry-After Header Logging

Both regular and streaming LLM call paths now log the raw Retry-After header value (or (none)) on 429 responses, confirming whether Databricks sends one and enabling future tuning of backoff parameters.

All Commits

Hash Summary
f4be01c feat: async fire-and-forget Genie Space builds from Ask Forge
2e3f70d refactor: introduce scoped logging with origin, module, and error category tracking
21eb93d fix: Genie Engine metric view quality and review performance
ea69642 feat: Genie Engine performance overhaul + hierarchical progress UI + empty response resilience
94f6448 chore: format codebase with Prettier
2cff418 feat: quality preset system with performance model bundle
a95199e fix: model compatibility guards -- content parsing, JSON mode gate, token clamping, and per-call logging
(pending) fix: progressive 429 backoff, Flash Lite generation tier, prompt rework for 8K models