Skip to main content

​
v1.18.23 — April 17, 2026

​
Flow versioning, presentation and blueprint quality guardrails, Context Graph lineage expansion, and broader reliability and access-control improvements

​
Genesis Flows and Blueprints

Flow authoring gained version control and safer secret handling, and blueprint workflows are now more predictable across working copies and production snapshots.
  • Flows now support full version history with production pinning: every definition change creates an immutable snapshot, any prior version can be restored, and scheduled or API-triggered runs execute the pinned production version while manual runs can target any version.
  • Flow Python steps can now resolve Genesis vault secrets at runtime through a dedicated secrets primitive — credentials are never exposed in environment variables, preamble code, or on disk.
  • Git commit and GitHub file-creation actions now scan for accidental credentials (API keys, tokens, private keys) before executing; the operation is blocked if a potential secret is detected, with an explicit override available when the detection is a false positive.
  • Blueprints now use a stable identity for every operation — edit, start, import, version, access — so repeated blueprint names across different sources no longer cause the wrong blueprint to be acted on.
  • Blueprint detail pages can switch between the working copy, the promoted production snapshot, and the default release, with access and edit permissions evaluated per source.
  • The synthetic-data blueprint now defaults to Genesis Datasets storage and the local Python executor so users without a Snowflake connection can run it immediately; Snowflake and Snowpark remain available as opt-in choices.
  • A new suite of Genesis migration blueprints and an expanded Flows UI bring blueprint and flow navigation into a consistent, unified experience.

​
Presentations and Agent-Built Apps

Agent-built decks now publish with visual coverage guardrails, and blank or incorrectly structured slides are caught before users see them.
  • Presentation validation now blocks slides that would render blank because of unsupported long-form field names or missing visible content, with explicit error codes and remediation hints.
  • The presentation runtime distinguishes empty slides from loading slides, so automated review no longer accepts splash screens as successful captures.
  • A new visual-coverage check audits every deck for missing visuals, duplicate image references, and unused workspace images, surfacing gaps before publication.
  • Narrated audio in presentation decks now persists reliably when platform authentication is enabled, surviving server restarts.
  • Versioned apps that have not yet been promoted to production are now served through the working-copy path so embeds no longer return 404 until promotion.
  • The file preview and git browser gained a “Download as DOCX” action, matching the existing PDF export flow; exports now wait for file content to finish loading before running, eliminating a misleading “no content” error.

​
Agent Knowledge, Wiki Learnings, and Skills

Agents can now capture, recall, and share domain knowledge across threads, and knowledge-base administration is available directly as agent tool calls.
  • Agents can save user-contributed domain knowledge as authored learnings; saved statements are retrieved automatically when related entities are mentioned in any future thread, bringing persistent memory to the entire team rather than just a single conversation.
  • Entity search results now include a learnings summary when saved knowledge exists for a matching data asset, prompting the agent to read and cite that context before responding.
  • Admins can create, list, and delete named knowledge bases through agent tool calls — no REST API or UI step required — with hardened locking so concurrent operations on the same bundle cannot corrupt state.
  • Agent-built apps can now be navigated and tested by the agent through interactive browser tools, letting the agent click, type, and screenshot running apps during iterative development.

​
Chat, Threads, and Conversation Reliability

Thread handling and chat interactions were hardened across high-concurrency and cross-user scenarios, with clearer streaming behavior and several UI polish fixes.
  • Stopping your own active response no longer affects another user’s concurrent conversation with the same agent; per-conversation isolation is now enforced consistently across all LLM adapters and both synchronous and streaming paths.
  • Thread messages are now written through a centralized, durable service boundary with optimistic versioning and idempotency, preventing lost or reordered messages under load.
  • Rapid back-to-back reads of the same thread are now coalesced so large teams reading the same conversation do not multiply storage traffic.
  • Reconstructed chat prompts are now placed in the correct position in conversation history, preserving the true sequence of user and agent messages even when the stream-stop or thread-restore paths are used.
  • Streaming responses now animate only the last in-progress line; completed lines no longer re-animate when new content arrives, removing visible flicker on long outputs.
  • The copy-to-clipboard button on chat messages is now available on agent responses as well as user messages, through the same hover interaction.
  • Avatar tooltips in chat now render above sticky containers and automatically flip below the avatar when space above is insufficient, so tooltips are never clipped by the viewport.
  • The chat input remains responsive during agent processing, and a new Get Started button appears on the chat welcome screen for users who have not yet completed onboarding.
  • The desktop sidebar is now resizable, and a mobile entry point for guided journeys was added to the navigation drawer.

​
Launchpad and Getting Started

The onboarding experience evolved into a dual-role landing page that welcomes new users and surfaces relevant activity for returning users.
  • The “Get Started” feature is now called “Launchpad” across the Genesis UI, better reflecting its role as a hub for guided journeys and exploration.
  • Returning users see a two-column layout with recent threads, recent missions and apps, and suggested workflows alongside the journey cards; new users keep the single-column onboarding flow.
  • The “Connect your tools & data” journey now includes an inline connector-type picker, so users choose their integration visually inside the wizard before being handed off to the connections page.
  • Wizard state is now preserved when users leave mid-journey to configure a connection and return — answers and step position are restored automatically.
  • A preview banner on the template picker step sets clear expectations for the evolving template library.

​
Context Graph, Lineage, and Data Browser

Lineage coverage expanded across Genesis Apps, SQLite, and Iceberg, and the Data Browser behavior now matches what agents can see.
  • Agent-built apps (“Genesis Apps”) that query Snowflake or other data sources directly from app code — without a dataset binding — now receive full lineage coverage automatically on the next scan, exposing the data sources they read from as connected edges in the Context Graph.
  • Iceberg data assets now use a stable catalog-derived identifier in their qualified names, enabling cross-scanner lineage to resolve correctly between references discovered in git and the assets registered from the Iceberg catalog.
  • SQLite connections created through the new managed-storage default are correctly scanned, indexed, and linked from the Context Graph; the Data Browser now opens the same database file the agent sees, fixing a path-resolution mismatch that previously produced empty table lists.
  • The Context Graph database picker continues to let you exclude databases added after a selection point, with the opt-in checkbox now clearly labeled and its state preserved across re-opens.
  • Authored learnings saved via the wiki are kept out of the lineage graph (stored in a dedicated annotation layer), so graph traversal stays clean while search still surfaces the learnings.

​
Connections, AI Providers, and OAuth

Enterprise authentication, AI provider onboarding, and connection creation were hardened, with new options for Bedrock proxies and clearer SQLite setup.
  • AWS Bedrock custom endpoints now work correctly with API Gateway deployments: connections use SigV4-compliant signing, and administrators can choose between Bearer, x-api-key, and custom auth headers, between Invoke and Converse API modes, and optionally strip AWS regional prefixes from model IDs for proxies that reject them.
  • AI provider model-selection dropdowns are now searchable by display name or technical identifier, with auto-focused search on open and a clear “No matches” empty state.
  • The LiteLLM Proxy OAuth integration was hardened: session creation no longer crashes in OAuth-only deployments, the caller’s user identity now flows through the auth chain so per-user OAuth resolves correctly for interactive chat, and a rejected token triggers a single automatic retry with a fresh token instead of surfacing a validation error.
  • The SQLite connection dialog defaults to a managed-storage mode that requires only a database name and shows the exact storage path inline; the advanced free-form path field remains available for existing workflows.
  • The OAuth popup flow on SPCS deployments no longer shows a false “Connection was not completed” error caused by browser cross-origin isolation; completion is resolved exclusively from callback signals with a 120-second safety timeout.
  • A session-drift detection layer now surfaces a blocking reload prompt when a user’s identity changes mid-session (for example, by signing in as a different user in another tab), preventing stale data from being served.

​
Access Control and RBAC

Role-based access control was simplified and made more consistent across system permissions and object-scoped grants.
  • The per-agent credential-profile system (admin / user-facing / autonomous) was removed: connection access is now driven exclusively by RBAC role assignment, connection assignment, secret-access rules, double-key requirements, and user OAuth grants — a simpler, single source of truth.
  • Updating a role’s system permissions through the Roles UI now preserves unrelated object-scoped grants (per-dashboard, per-app, per-flow, per-blueprint), fixing a case where these could be silently dropped.
  • The Agent Center automatically hides the LLM Settings, Access, and Knowledge tabs for users without the corresponding backend permissions, replacing “Failed to load” errors with a clean UI.
  • Tab and route guards across Agent Center and the main navigation are now consistently aligned so users see only the areas they are permitted to use.
  • Documentation for shared and service-style credentials (PAT, username/password, key-pair) was corrected across connector setup guides: these modes are available to any agent when RBAC and secret-access settings permit.

​
Platform Operations, Snowflake, and Deployment

Core platform operations gained self-healing controls and safer deployment defaults.
  • A new bot-storage mount-health watchdog detects when the durable storage mount has become permanently unavailable and triggers a graceful automatic restart, converting a previously silent multi-hour outage into a roughly one-minute recovery event. Diagnostic snapshots are written to Snowflake as a durable audit record, along with ephemeral local copies for on-host debugging.
  • Startup race conditions that produced transient “Connection refused” errors in Snowflake deployments were eliminated: internal registrations now wait for the HTTP server to be ready, and callers retry transient failures.
  • Snowflake connections to external accounts no longer fail because of an SPCS session token being injected automatically by the Snowflake Python connector — auto-injection is now explicitly disabled at container launch and reset inside the connector.
  • Snowflake deployment specs reverted to the compatible legacy stage-mount syntax after the newer verbose format caused deployment failures; deployments on mature app packages no longer fail with “Argument list too long” on SHOW VERSIONS.
  • The Snowflake metadata cache now lives in local container storage — where SQLite WAL semantics are reliable — while durable invalidation markers remain on stage storage; this lets the cache be enabled again for standard SPCS deployments.
  • New scheduled cleanup jobs reclaim expired dataset snapshots, old API-call archive files, and stale coding-agent session artifacts at configurable intervals; a startup dry-run audit produces a sortable reclaim report without deleting anything.

​
Slack and Integrations

Slack-backed conversations were hardened against edge cases seen in high-traffic deployments.
  • New Slack-backed threads no longer fail to materialize on object-backed storage: the runtime reuses the already-open thread handle instead of reopening a freshly written file that is not yet visible.
  • Silent message drops on Slack dispatch paths are now surfaced as structured errors with bot and thread context, making production diagnostics visible for the first time.
  • The Slack log-file path now creates its parent directory on first write, and the Slack Apps link in the Genesis UI was polished for clarity.

​
Genesis UI Polish and Stability

Many focused fixes across the Genesis UI improve day-to-day usability.
  • The Scheduler page no longer goes blank when a cron expression omits the minute (or any other optional field); partial cron expressions are interpreted defensively.
  • The “Server Offline” overlay remains stable when the backend is stopped, no longer flickering between “Loading Genesis” and “Server Offline” on every health-check cycle.
  • The profile settings modal now renders above sticky page toolbars and carries correct ARIA attributes for assistive technology.
  • Dataset listing is purely read-only again — opening the Datasets tab no longer triggers background snapshot cleanup as a side effect.
  • Query results in the Data Explorer no longer get corrupted when a user re-runs a query after a paginated run; the original SQL is always preserved and stale results from abandoned queries are discarded.
  • The Agent Center Threads tab correctly shows its active/selected style when navigated to, and the Deployments/Sandboxes navigation entry highlights consistently across all deployment-related pages.

​
Performance and Observability

Agents respond faster on repeated questions, and administrators have better insight into server health.
  • Follow-up questions in an active conversation return more quickly — the agent reuses its understanding of the current context instead of rebuilding it from scratch on every turn.
  • Per-turn timing breakdowns are richer, so administrators can see exactly which phase of agent preparation spent time on any given turn.
  • The server-monitoring view now shows clearer attribution of background activity, including the longest-running background tasks and which areas of the system are growing over time — making unexpected load easier to spot.
  • The chat interface handles high-frequency polling more efficiently, lowering background load without changing user-visible behavior.
  • First-time sign-in is now safe under concurrent requests — no duplicate user accounts are created, and Slack bot identities are no longer mistaken for human users.