Release Notes -- 2026-03-14¶

Databricks Forge v0.32.0 → v0.37.3

v0.32.0 -- Demo Mode¶

New Features¶

Demo Mode -- Synthetic Data Generator for Field Engineering & Sales¶

A complete demo data generation system that creates customer-specific synthetic datasets in Unity Catalog for live demonstrations.

Research Engine -- LLM-powered company research with three depth presets (Quick / Balanced / Full) that scrapes company websites, discovers investor relations documents, processes uploaded PDFs, auto-detects industry, and runs multi-pass strategic analysis to identify priorities, data assets, and customer-specific nomenclature.

Data Engine -- SQL-first synthetic data generation that designs relational schemas from industry outcome maps, generates dimension and fact tables with embedded data narratives (trends, spikes, anomalies, seasonal patterns), and writes directly to Unity Catalog managed Delta tables (2,000--10,000 rows).

Key capabilities: - 6-step wizard accessible from Settings (Company Info, Research Results, Catalog Selection, Schema Review, Generation Progress, Complete) - Auto-industry detection with on-the-fly outcome map generation for industries not in the built-in registry - Demo scoping by division, department, or functional area - Per-session cleanup (DROP TABLE/SCHEMA + Lakebase record deletion) - Real-time progress tracking with per-table phase visibility

Enablement: ./deploy.sh --enable-demo-mode or FORGE_DEMO_MODE_ENABLED=true in .env.local.

See docs/DEMO_MODE.md for the full team guide.

Other Changes¶

Updated AGENTS.md with Demo Mode architecture documentation
Added ForgeDemoSession Prisma model and Lakebase CRUD operations
Extended ForgeOutcomeMap with enrichmentJson for custom LLM-generated industry outcome maps
Added async getMasterRepoEnrichmentAsync() fallback to custom enrichment
Integrated demo_research, demo_generate, demo_cleanup activity actions
Added prisma.forgeDemoSession.deleteMany() to factory reset

v0.32.1 -- Demo Mode Comprehensive Fix¶

Improvements¶

Customer Insight Page¶

New dedicated page at /demo/sessions/{sessionId} with a polished, presentation-ready view of all research output -- company overview, SWOT analysis, industry landscape, data strategy, demo flow, killer moments, and competitive positioning. Each section renders conditionally based on the research preset used.

PPTX and PDF Export¶

Research results can now be exported as a branded PPTX slide deck (10--12 slides) or a PDF document from the Customer Insight Page or the sessions table. Both formats skip sections where data is unavailable.

Demo Sessions Listing Page¶

New /demo page with a full sessions table replacing the previous inline list in Settings. Supports row-click navigation, export actions, and delete with confirmation.

Demo section appears in the sidebar when FORGE_DEMO_MODE_ENABLED=true, with a useDemoModeEnabled() hook and requiresDemoMode flag for consistent feature gating.

Research Progress Timeline¶

Step-by-step timeline in the wizard with completed/active/pending phase indicators, elapsed timer, and human-readable detail messages at every micro-boundary.

Replaced the free-text industry input with a Select dropdown fetching from /api/industries, preventing ID normalization mismatches.

Enrichment-Only Generation¶

Outcome map logic now has three cases: skip if both outcome and enrichment exist, generate enrichment only if outcome exists but enrichment is missing, and full generation only when neither exists. Prevents unnecessary regeneration.

Bug Fixes¶

FK constraint violation -- removed upsertJobStatus calls from both demo engine-status modules; demo jobs now use in-memory tracking only (no writes to ForgeBackgroundJob which has an FK to ForgeRun).
Missing table descriptions -- seed and fact generation now apply COMMENT ON TABLE and ALTER TABLE ... ALTER COLUMN ... COMMENT via buildTableCommentDDL/buildColumnCommentDDL after each table creation.
LLM token limit failure -- changed seed and fact generation from resolveEndpoint("generation") to resolveEndpoint("sql") for 128K output capacity models.
Health endpoint caching -- moved demoModeEnabled into the base (unauthenticated) response object in /api/health so the demo flag is always available.
Industry ID normalization -- added normalizeIndustryId() with kebab-case, starts-with, and name-based fuzzy matching to prevent "Water Utility" vs "water-utilities" mismatches.

Other Changes¶

Simplified demo-settings.tsx to a link card pointing to /demo
Replaced free-text catalog/schema inputs with CatalogBrowser component + "Create new catalog" toggle
Added DemoIcon SVG to sidebar icon set
Updated docs/DEMO_MODE.md with all new features, routes, and file references
Dense progress messages throughout the Research Engine analytical pipeline

v0.33.0 -- Launch Discovery from Demo Sessions¶

New Features¶

Launch Discovery & Estate Scan from Demo Sessions¶

After demo data generation completes, users can now launch a Discovery Pipeline or Estate Scan directly from both the wizard completion step and the session detail page. Buttons call the existing /api/runs and /api/environment-scan endpoints with the generated schema scope, customer name, and industry pre-filled.

Session Detail Page Redesign¶

The demo session detail page has been redesigned with: - Dark hero banner with customer name, industry badge, schema FQN, and live KPI stats - Sticky command bar with Launch Discovery, Estate Scan, and Export actions - Collapsible research sections with accent-colored borders and contextual icons - SWOT analysis rendered as a 2x2 color-coded grid - Market forces with urgency badges, demo flow as a vertical timeline - Killer moments, competitive positioning, and executive talking points sections

Bug Fixes¶

Data Engine cascade failure -- Failed dimension tables (e.g. dim_date) are now filtered out before fact generation, preventing cascading TABLE_OR_VIEW_NOT_FOUND errors across all downstream fact tables.
Spark datetime pattern errors -- Added explicit rules to demo SQL constraints and global DATABRICKS_SQL_RULES forbidding 'u', 'e', 'c', 'L' patterns in DATE_FORMAT() which cause DATETIME_PATTERN_RECOGNITION errors on Databricks.
Smart fact retry -- The TABLE_OR_VIEW_NOT_FOUND retry in fact generation now parses the missing table name, removes it from the dimension context, and adds an explicit constraint telling the LLM not to reference it.
Export 500 errors -- Added try-catch with structured logging to the demo export API route. Both PPTX and PDF generators now use optional chaining on all LLM-parsed array properties (priorities, urgency signals, SWOT, market forces, benchmarks, asset details) preventing crashes on partial research results.

v0.34.0 -- Mining & Resources Industry¶

New Features¶

Mining & Resources Industry Outcome Map¶

Added a dedicated industry outcome map for Mining & Resources, covering the strategic priorities of major miners such as BHP Group, Rio Tinto, Fortescue, and South32.

Base outcome map (mining.ts) with 6 strategic objectives, 10 priorities, and 30 reference use cases spanning mine operations, safety, logistics, ESG, exploration, and business functions.
Enrichment data (mining.enrichment.ts) with 30 benchmarked use cases (KPI targets, benchmark impacts, model types, data asset criticality mappings) and 30 Reference Data Assets covering the full mining value chain.
Sub-verticals: Iron Ore, Coal, Copper & Base Metals, Gold & Precious Metals, Lithium & Battery Minerals, Alumina & Aluminium, Nickel & Cobalt, Manganese & Alloys.
Industry aliases updated so "mining", "metals and mining", "iron ore", "base metals", "precious metals", "battery minerals", etc. now resolve to the dedicated mining outcome map instead of falling back to Energy & Utilities.
"Mining" removed from Energy & Utilities sub-verticals to avoid overlap.

v0.34.1 -- Demo Engine SQL Robustness¶

Improvements¶

Demo Engine SQL Robustness (6 new constraint rules)¶

Added six critical SQL constraint rules to DEMO_DATA_SQL_CONSTRAINTS: INTERVAL+DATE casting, safe modulo with GREATEST(), BIGINT overflow prevention, get() instead of element_at(), rand() misuse guard, and self-referencing table prohibition.

Schema Design Performance¶

Switched schema design from the reasoning endpoint to generation, reduced maxTokens from 32K to 16K, and simplified the prompt to target 8-12 lean demo tables instead of 35+ enterprise-scale models.

Inline Table Comments¶

Table descriptions are now included as COMMENT clauses directly in CREATE TABLE statements, eliminating redundant post-creation COMMENT ON TABLE DDL round-trips.

Cascading Failure Resilience¶

Fact table generation now checks if all referenced dimension tables failed, and skips doomed fact tables early instead of generating SQL that references non-existent tables.

Research Summary Redesign¶

Replaced the flat stat boxes with a polished editorial layout: icon stat ribbon, numbered priority list, two-column nomenclature grid, and compact header with status indicator.

Schema Review Redesign¶

Replaced visual chaos with structured cards: pill-style data asset chips, narrative cards with pattern badges, divided nomenclature table, and accent-bordered demo highlight cards.

Smart PDF and PPTX Page Breaks¶

PDF drawTable() now paginates rows across pages with header repetition. PDF and PPTX use case detail rendering creates continuation pages/slides instead of silently dropping overflow fields.

Bug Fixes¶

CatalogBrowser catalog selection -- Schema mode now shows a "Select" button on catalog rows, allowing catalog-level selection in the demo wizard with auto-generated schema name.
Generation timer missing -- Added elapsed timer to GenerationProgressStep and total wizard time display on CompleteStep.
Clickable demo sources -- Added url field to ResearchSource interface, populated from website-scrape, strategic-crawl, and IR-crawler passes. Sources on the session detail page are now clickable links to original URLs.

v0.34.2 -- Lakebase Scale-to-Zero Fix¶

Bug Fixes¶

Lakebase scale-to-zero provisioning -- Split the combined no_suspension + suspend_timeout_duration PATCH into two sequential API calls. The Lakebase API requires these fields to be updated independently; the combined update_mask was silently ignored, leaving the timeout at the API default (300s) regardless of the configured value. Now the first call disables suspension, the second sets the custom timeout, with graceful fallback if the timeout call fails.

v0.34.3 -- Research Engine Speed¶

Improvements¶

Research Engine speed optimization for balanced preset¶

The balanced research preset now routes LLM calls through generation-tier models (Claude Sonnet 4.6, Gemini Flash, Llama Maverick) instead of hardcoding to Claude Opus 4.6. This reduces wall-clock time from ~8.5 minutes to ~2-3 minutes for a typical balanced run. The full preset retains reasoning-tier (Opus) routing for maximum quality.

Key changes:

Added per-preset modelTier to ResearchBudget (quick=classification, balanced=generation, full=reasoning).
Refactored resolveResearchEndpoint() to accept an optional tier, delegating non-reasoning tiers to the queue-depth-aware task router.
Lowered balanced maxTokensPerPass from 32,000 to 16,000 (outputs are structured JSON that rarely exceeds 8K tokens).
Parallelized embedding with the industry classification and outcome map pipeline to eliminate serial wait.

Bug Fixes¶

Enrichment-only generation used reasoning tier -- runEnrichmentOnlyGeneration documented "uses the generation tier (faster)" but actually called the reasoning endpoint. Now correctly routes to generation tier.
Industry classification used reasoning tier -- A 512-token classification task was being sent to Opus. Now always routes to classification tier regardless of budget.

v0.35.0 -- Auto-Launch Discovery + Strategic Persona¶

New Features¶

Auto-Launch Discovery After Demo Data Generation¶

The Demo Data Wizard now automatically creates and executes a Discovery Pipeline with Environment Scan enabled as soon as data generation completes. Users no longer need to manually click separate "Launch Discovery" and "Run Estate Scan" buttons -- the wizard handles it in one seamless flow, redirecting directly to the run page. A manual fallback button appears if auto-launch fails.

Strategic Persona Questions in Ask Forge¶

The Strategic (C-level) persona in Ask Forge now generates context-aware executive questions: total business value, ROI prioritisation, board-level summaries, strategic alignment gaps, and transformation roadmaps. Previously the strategic persona fell through to the analyst branch, producing data-quality and PII questions instead of executive insights.

Improvements¶

Shared Strategic Fallback Questions¶

Moved FALLBACK_QUESTIONS_STRATEGIC from the client-side component to the shared suggested-question-defaults.ts module so both server-side dynamic question generation and client-side fallbacks use the same set of C-level questions.

Wizard Step Labels¶

Updated the final wizard step label from "Done" to "Launch" and the description to "Launching discovery pipeline..." to reflect the new auto-launch behaviour.

v0.35.1 -- SQL Review Quality¶

Bug Fixes¶

SQL reviewer severity inflation -- The LLM-based SQL reviewer was classifying stylistic and idiom issues (e.g. missing COLLATE, no explicit window frame) as "error" severity, which forced expensive fix cycles on functional SQL. Added explicit severity classification guidance so only runtime failures (syntax errors, hallucinated columns, wrong JOINs) produce fixed_sql. Warnings and info issues are reported but no longer trigger rewrites. Affects all engines: pipeline, dashboard, and Genie.
Missing skill injection in pipeline SQL generator -- The {skill_reference} placeholder was computed by the pipeline SQL generation step but never actually inserted into the USE_CASE_SQL_GEN_PROMPT template. The generator was operating without the SQL craft knowledge from the skills system. Now wired correctly.

Improvements¶

Databricks SQL dialect context block¶

Added a compact quick-reference of Databricks SQL dialect differences (e.g. PERCENTILE_APPROX not MEDIAN, array_join(collect_list(...)) not STRING_AGG, named window limitations) directly in the pipeline SQL generator prompt. This reduces first-pass errors from generators unfamiliar with Databricks idioms.

Pre-response self-check for SQL generators¶

New DATABRICKS_SQL_SELF_CHECK exported from lib/toolkit/sql-rules.ts -- a structured checklist (correctness, performance, readability, Databricks idioms) that the generator must verify before responding. Injected into the pipeline SQL generation prompt.

Enhanced SQL review fix logging¶

When the reviewer applies a fix in the pipeline, logs now include qualityScore, verdict, issueCount, and the top 3 issues sorted by severity. This improves observability for diagnosing persistent review failures.

v0.35.2 -- Demo Data & Pipeline Speed¶

Bug Fixes¶

Demo Data Generator: CTE column scoping error -- Added a CTE Column Scoping rule to the SQL generation prompt so the LLM no longer references aliased-away column names in downstream CTEs. Also wired the runtime SQL error into the reviewer so retry attempts are informed fixes instead of blind re-reviews.
Demo Wizard: auto-launch without pause -- The completion step no longer auto-launches the discovery pipeline immediately. Users now see the generation summary (tables, rows, duration) and must click "Start Discovery Run" to proceed.
Pipeline SQL generation: excessive review latency -- Removed the per-use-case LLM review call during pipeline SQL generation. EXPLAIN validation and column-hallucination checks still catch runtime errors; the review was adding ~10-15s per use case for mostly cosmetic warnings.
SQL generation progress bar premature completion -- Fixed the progress interpolation range (67-95 changed to 67-79) so the UI no longer marks SQL generation as "done" when only half the use cases are processed.

Other Changes¶

ReviewOptions in lib/ai/sql-reviewer.ts now supports an optional runtimeError field, injected as a priority section in the review prompt for any consumer that needs error-targeted fixes.

v0.36.0 -- Build Mode Selector¶

New Features¶

Build Mode Selector¶

A new shared BuildModeSelector component gives users an explicit choice between Full Engine (recommended) and Quick Build before generating a Genie Space. This replaces the previous behavior where Ask Forge auto-started a fast build and Schema Scan / Upload Requirements pages only offered the full engine.

The selector is now available in all three creation entry points: - Ask Forge (GenieBuilderModal) -- shows a choice screen before building - Scan Schema page -- replaces the single "Generate" button - Upload Requirements page -- replaces the single "Generate" button

Full Engine is highlighted as the recommended default.

Improvements¶

Instant space visibility after deployment¶

Deployed Genie Spaces now appear immediately in the Genie Studio listing without requiring a manual "Sync Spaces". The deploy flow writes the new space to ForgeGenieSpaceCache alongside the existing ForgeGenieSpace tracking entry.

Build progress tiles for all build modes¶

Synchronous fast builds now create an in-memory server-side job, so a progress tile with a progress bar appears on the Genie Studio page while the build runs. Previously, fast builds from Ask Forge were invisible to Genie Studio.

Accurate card titles after generation¶

Generate job tiles now show the actual space name from the recommendation instead of the generic "Genie Space Ready" fallback. Title and domain are backfilled from the recommendation on both synchronous and asynchronous completion paths.

Deploy status propagation¶

After deploying from the build page or Ask Forge modal, the generate job tile badge correctly updates from "Ready" to "Deployed". Both the build page and GenieBuilderModal now PATCH the generate job with the deployed space ID.

Bug Fixes¶

Card shows "Genie Space Ready" instead of space name -- Title and domain are now backfilled from the recommendation when the generate job completes.
Card stays "Ready" after deployment -- Both the build page and GenieBuilderModal now mark the generate job as deployed via PATCH.
Deployed space invisible until sync -- upsertCachedSpaces is called in runDeploy() after trackGenieSpaceCreated().
Unused Zap imports -- Removed from Schema Scan and Upload Requirements pages after replacing buttons with BuildModeSelector.

v0.36.1 -- Rename Demo Sessions to Demo Studio¶

Improvements¶

Rename "Demo Sessions" to "Demo Studio"¶

The Demo section has been rebranded from "Demo Sessions" to "Demo Studio" across the sidebar navigation, page heading, and session detail breadcrumbs for a more polished, product-oriented feel.

v0.36.2 -- Consolidate Release Notes (One File Per Day)¶

Improvements¶

Release Notes Consolidation¶

Merged 38 release notes files down to 12 (one per day). Days that previously had multiple suffix files (e.g. b, c, d ... l) now have a single file with versioned subsections. Updated the release notes generation and gogo ship rules to append new versions to the existing day's file instead of creating suffix files.

v0.36.3 -- Demo Session Launch Fix¶

Bug Fixes¶

Demo session catalog/schema not persisted -- The demo session was created with empty catalogName and schemaName during the research step, and the generate route never wrote them back to the database. Launching Discovery or Estate Scan from the session detail page produced UC metadata scope . (empty), failing immediately at metadata extraction. The generate route now persists catalogName and schemaName when transitioning the session to "generating".
"View full research briefing" opens in same window -- The link in the wizard completion step now opens the session detail page in a new browser tab instead of navigating away from the wizard via router.push.
mkdocs nav stale release note references -- Removed deleted suffix release note files from mkdocs.yml navigation that were causing 404s on the docs site.

v0.37.0 -- Sales Briefing Redesign¶

Improvements¶

Demo Session Detail Page: Sales-First Briefing Experience¶

Complete redesign of the /demo/sessions/[sessionId] page from a long research dump into an executive sales briefing optimised for 90-second scanability.

New information architecture: - Briefing header with account thesis, best wedge, why-now signal, confidence badge, and three-tier CTAs (Start Discovery / Open Talk Track / View Evidence) - Meeting Summary Strip -- 5 compact cards (Top Priority, Best Use Case, Main Risk, Best Buyer, Proof Strength) for instant pre-meeting context - Tab-based navigation replacing vertical scroll: Summary, Opportunities, Talk Track, Evidence, Sources - Sticky sidebar rail (desktop) with account thesis, top 3 opportunities, quick actions, export, and confidence/source stats

Summary tab includes Executive Brief card (who they are / what they care about / what is likely broken / best opening angle), "So what this means for us" callout, Priorities & Gaps, Compact SWOT (2x2 mini-grid, no pastel fills), and Industry Trends as chips with collapsible detail.

Opportunities tab transforms Killer Moments into ranked opportunity cards with business pain, value hypothesis, discovery hook, linked data assets, confidence, urgency, ideal buyer, and "Use in Talk Track" CTA.

Talk Track tab provides persona-based selling guidance (CEO, COO, CIO/CTO, Head of Digital, Risk/Compliance) with what they care about, what to say, expected objection, proof to use, and best next question.

Evidence tab structures claims with confidence levels, evidence counts, freshness indicators, and supporting source snippets.

Data Readiness redesign replaces cryptic asset IDs with relevance scores, criticality badges, quick-win flags, and linked use cases.

Component architecture -- 11 new modular components: ResearchHeader, MeetingSummaryStrip, ExecutiveBriefCard, CompactSwot, SummaryTab, OpportunitiesTab / OpportunityCard, PersonaTalkTrack, EvidenceList, SourceList, DataReadinessList, StickyBriefRail.

Page reduced from 850 lines in a single file to 210 lines orchestrating focused components. Desktop layout uses max-w-[1440px] with main content + 288px sticky right rail; mobile stacks cleanly with preserved tabs.

v0.37.1 -- SEC EDGAR False Match Fix¶

Bug Fixes¶

SEC EDGAR matches wrong company for non-US domains -- The Research Engine's SEC EDGAR fallback used an overly loose name-matching algorithm that checked if the domain name contained the first word of any SEC title. For companies like "O REILLY AUTOMOTIVE INC", the first word is "o" -- a single character that matches virtually any domain. An Australian company's website was incorrectly matched to O'Reilly Automotive (CIK 0000898173), injecting a completely unrelated 10-K filing into the research context.

Three fixes applied: 1. Country-code TLD guard -- isNonUsDomain() detects .com.au, .co.uk, .de, and 40+ other ccTLDs and skips SEC EDGAR entirely for non-US domains. 2. Scored prefix matching -- Replaced the naive .includes() with findBestEdgarMatch() that normalizes both names (stripping spaces and punctuation), requires prefix alignment, scores candidates by overlap ratio, and rejects matches below 40% similarity. 3. Exact ticker match priority -- Ticker symbols are checked first as the highest-confidence match before falling back to name comparison.

Dead EDGAR search API calls -- Removed two unused efts.sec.gov search index fetches that were executed but whose responses were never consumed.

v0.37.2 -- Genie Studio Workflow Fixes¶

Bug Fixes¶

Persona not defaulting on navigation -- Navigating from Genie Studio "Describe Your Space" to Ask Forge now correctly sets the genie-builder persona via a useEffect that syncs from URL params, resetting the conversation state for a fresh chat.
Aggressive genie-builder guardrail -- Softened the GENIE_BUILDER_PERSONA_OVERLAY prompt so it treats all user messages as input for building a Genie Space instead of refusing with "Let's stay focused on building your Genie Space."
Inline sources removed from chat -- SourceCardList has been removed from the chat for all personas. Sources remain in the right-hand context panel with clickable navigation links for table, use case, genie, and environment sources.
Missing build mode prompt -- The "Create Genie Space" action in Ask Forge now opens the GenieBuilderModal (Fast vs Full Engine choice) instead of silently starting a fast build.
Build cards invisible with no spaces -- Build progress cards now render when no synced spaces exist by including job count in the empty-state condition.
Text overflow on build cards -- Added min-w-0, truncate, and shrink-0 to card flex containers and text elements to prevent content from bleeding out of card boundaries.
Deployed card still visible -- Build job cards are now filtered by deployedSpaceId to prevent duplicate cards after deployment.
Build toast persists forever -- Deployed toast now auto-dismisses after 10 seconds and includes an X close button.
Misleading "Run AI Comments" advice -- Removed the hardcoded link from health check UI. Made tables-have-column-configs and format-assistance-configured fixable with column_intelligence strategy. Extended the fixer to enable enable_format_assistance and enable_entity_matching on existing columns.
Fix strategy metadata retrieval broken -- buildMetadataForSpace used Array.isArray(result) on the SqlResult object (always false) and named property access on positional arrays. Fixed to iterate result.rows with positional indexing, unblocking all SQL-generating fix strategies (benchmarks, trusted assets, semantic expressions).
Health check categories always expanded -- Accordion now defaults to expanding only categories with failing checks. Added status icons (green check, red X, amber warning) to category triggers.
Targeted improve crash -- The "Improve with Genie Engine" targeted path now includes updatedSerializedSpace in the result. The UI safely falls back through recommendation?.serializedSpace, updatedSerializedSpace, and originalSerializedSpace.

Improvements¶

Benchmark Test Runner Redesign¶

Replaced the cramped 3-column step tile grid with a horizontal flow layout featuring numbered step badges, connecting arrows, staggered entrance animations, and a single CTA button.

Sources in the right-hand context panel now include clickable external-link icons that navigate to the relevant page (table detail, use cases, genie, environment) based on source kind.

v0.37.3 -- Ask Forge & Health Check Fixes¶

Bug Fixes¶

Ask Forge source noise in chat -- Removed enforceSourceCitations which appended a raw ### Sources block (e.g. [1] [GENERATED INTELLIGENCE] (use_case)) to the answer text when the LLM omitted inline citations. Sources are already shown in the right-hand context panel; the duplicate text in the chat was noisy and unhelpful.
Genie builder modal stays open -- The "Create Genie Space" action in Ask Forge kept the modal open during generation with its own progress UI. Refactored GenieBuilderModal to use useGenieBuild().startBuild() with async mode, closing the modal immediately after Quick/Full selection and showing a persistent build progress toast instead.
Column config changes invisible in Preview -- The health check "Preview Changes" diff viewer only detected added/removed tables, not modifications to column_configs on existing tables. Added diffColumnConfigs() that compares column descriptions, synonyms, and flags within matched tables, with a new "Column Configs" section showing added/modified/removed columns in amber/green/red.
Benchmark fix only adds 3 questions -- The health check fix for "Benchmark questions exist (5+ recommended)" passed useCases: [] and used the default BENCHMARKS_PER_BATCH of 4. After validation drops, this yielded ~3 benchmarks. Increased the fix target to 8 benchmarks per batch, comfortably exceeding the 5+ threshold after validation.

All Commits¶

Hash	Summary
(pending)	feat: Demo Mode wizard with Research Engine and Data Engine
(pending)	fix: Demo Mode comprehensive enhancements -- insight page, export, progress, industry matching, FK fix, table comments
(uncommitted)	fix: demo engine robustness, export null safety, session detail UX redesign
(pending)	feat: add Mining & Resources industry outcome map with enrichment
(pending)	fix: demo engine overhaul -- SQL robustness, schema speed, wizard UX, exports
(pending)	fix: split Lakebase scale-to-zero into two PATCH calls for reliable timeout
(uncommitted)	perf: research engine speed -- tiered model routing, lower token budgets, parallel embedding
(pending)	feat: auto-launch discovery after demo generation + fix strategic persona questions
`24c5d79`	fix: reduce SQL review severity inflation and inject missing skill reference into pipeline generator
(uncommitted)	fix: demo data retry, wizard pause, pipeline SQL speed, progress bar accuracy
(pending)	feat: build mode selector + fix Genie Studio card visibility/title/deploy status
(pending)	chore: rename Demo Sessions to Demo Studio across UI
(pending)	chore: consolidate release notes into single files per day
`7c546aa`	fix: persist demo session catalog/schema so Launch Discovery works
(pending)	fix: SEC EDGAR false match for non-US domains -- ccTLD guard, scored matching, dead code cleanup
`80422be`	fix: Ask Forge source noise, Genie modal toast, health check column diff, benchmark count
`c4ce20d`	fix: update mkdocs nav to consolidated release notes

Release Notes -- 2026-03-14¶

v0.32.0 -- Demo Mode¶

New Features¶

Demo Mode -- Synthetic Data Generator for Field Engineering & Sales¶

Other Changes¶

v0.32.1 -- Demo Mode Comprehensive Fix¶

Improvements¶

Customer Insight Page¶

PPTX and PDF Export¶

Demo Sessions Listing Page¶

Sidebar Navigation¶

Research Progress Timeline¶

Industry Dropdown¶

Enrichment-Only Generation¶

Bug Fixes¶

Other Changes¶

v0.33.0 -- Launch Discovery from Demo Sessions¶

New Features¶

Launch Discovery & Estate Scan from Demo Sessions¶

Session Detail Page Redesign¶

Bug Fixes¶

v0.34.0 -- Mining & Resources Industry¶

New Features¶

Mining & Resources Industry Outcome Map¶

v0.34.1 -- Demo Engine SQL Robustness¶

Improvements¶

Demo Engine SQL Robustness (6 new constraint rules)¶

Schema Design Performance¶

Inline Table Comments¶

Cascading Failure Resilience¶

Research Summary Redesign¶

Schema Review Redesign¶

Smart PDF and PPTX Page Breaks¶

Bug Fixes¶

v0.34.2 -- Lakebase Scale-to-Zero Fix¶

Bug Fixes¶

v0.34.3 -- Research Engine Speed¶

Improvements¶

Research Engine speed optimization for balanced preset¶

Bug Fixes¶

v0.35.0 -- Auto-Launch Discovery + Strategic Persona¶

New Features¶

Auto-Launch Discovery After Demo Data Generation¶

Strategic Persona Questions in Ask Forge¶

Improvements¶

Shared Strategic Fallback Questions¶

Wizard Step Labels¶

v0.35.1 -- SQL Review Quality¶

Bug Fixes¶

Improvements¶

Databricks SQL dialect context block¶

Pre-response self-check for SQL generators¶

Enhanced SQL review fix logging¶

v0.35.2 -- Demo Data & Pipeline Speed¶

Bug Fixes¶

Other Changes¶

v0.36.0 -- Build Mode Selector¶

New Features¶

Build Mode Selector¶

Improvements¶

Instant space visibility after deployment¶

Build progress tiles for all build modes¶

Accurate card titles after generation¶

Deploy status propagation¶

Bug Fixes¶

v0.36.1 -- Rename Demo Sessions to Demo Studio¶

Improvements¶

Rename "Demo Sessions" to "Demo Studio"¶

v0.36.2 -- Consolidate Release Notes (One File Per Day)¶

Improvements¶

Release Notes Consolidation¶

v0.36.3 -- Demo Session Launch Fix¶

Bug Fixes¶

v0.37.0 -- Sales Briefing Redesign¶

Improvements¶

Demo Session Detail Page: Sales-First Briefing Experience¶

v0.37.1 -- SEC EDGAR False Match Fix¶

Bug Fixes¶

v0.37.2 -- Genie Studio Workflow Fixes¶

Bug Fixes¶