v1.18.9 — April 13, 2026
Smarter conversation memory, agent knowledge bases, flow versioning, and Context Graph lineage improvements
Chat, Threads, and Conversation Memory
Conversation memory was overhauled and thread management was improved so agents retain context more reliably across long and complex interactions.- A new conversation memory system is now the default experience — agents produce clearer summaries, recover context from earlier in the conversation more accurately, and respect per-agent and per-thread context-limit overrides.
- Delegated handoff threads are now hidden from the main thread list by default, with a toggle to show them. Handoff references in chat appear as clickable navigation links.
- Shared thread deep links now navigate to the correct conversation reliably, including across authentication redirects and new-window scenarios.
- Thread search results now appear faster through polling-based retrieval instead of a fixed delay.
- Thread search indexing now starts in the background without blocking server startup.
- Scheduled-job threads no longer accumulate unbounded storage over time — cleanup operations now compact the underlying transaction log automatically.
- Message streaming visibility improved: tool-call result boxes remain visible during stream gaps, and chat input stays responsive during agent processing.
Context Graph, Lineage, and Data Scanning
Context Graph setup, lineage scanning, and metadata ingestion were improved for easier configuration and more predictable results.- Harvester-to-Context Graph auto-migration now works end-to-end — migrated metadata is properly indexed and searchable, with correct column casing and source registration.
- Database selection in Context Graph lineage configuration now supports a timestamp-based exclusion flag: databases created after the selection point are automatically skipped on future scans, preventing unintended scan scope growth.
- The system Snowflake connection is now available for lineage scanning in any Snowflake-configured deployment, not just Native App environments.
- Lineage source names now display in human-readable form throughout the configuration and scan status screens.
- Iceberg lineage indexing was restored and hardened, with a persistent opt-out that is preserved across server restarts.
- Iceberg query execution now runs in a dedicated worker process to prevent interference with concurrent agent activity.
- Daily Context Graph scans are now auto-enabled after harvester migration completes.
Genesis Flows and Blueprints
Flow management gained version control, and blueprint execution is now more reliable for longer-running missions.- Flows now support versioning with production pinning — teams can iterate on draft flow versions while a stable version remains active, and promote new versions to production when ready.
- Flow graph views now surface richer runtime and data-flow visibility across the dashboard.
- Per-step tool allowlists are now enforced for Flow Python steps, strengthening execution security.
- Blueprint missions can no longer be marked complete prematurely — the system now guards against status changes that skip required blueprint phases.
- Polling-based status checks (such as monitoring blueprint creation progress) no longer trigger false refusal messages after repeated identical calls.
- A new suite of Snowflake migration blueprints is available for guided database migration workflows.
Agent Knowledge and Skill Discovery
Agents gained dedicated knowledge bases and better access to product documentation and skills.- Administrators can now create named knowledge bases, assign them to specific agents, and each assigned agent gains a dedicated search tool scoped to its assigned document collections.
- Agents now have always-on access to Genesis product documentation search, improving their ability to answer product-related questions without manual tool activation.
- Skill and tool discovery was improved — searches with generic wording, hyphenated names, or minor misspellings now return more relevant results, and the initial discovery window was widened.
- Parallel tool execution is now enabled by default, allowing agents to run multiple tool calls concurrently for faster task completion.
Connectors, OAuth, and Slack Reliability
Slack integration reliability improved, and enterprise connection options expanded.- Slack session handling was hardened: idle bots no longer reconnect unnecessarily, worker threads are cleaned up correctly after reconnects, and slot exhaustion is handled gracefully instead of silently failing.
- Slack attachment delivery failures are now retryable, and placeholder messages are deferred until the agent has a real response.
- AI provider OAuth support was added, expanding the range of authentication options for connecting to external AI services.
- Databricks action discoverability and workspace path handling were improved for smoother onboarding.
Snowflake and Platform Operations
Snowflake deployment workflows and platform administration were hardened for more predictable operations.- The SPCS connection wizard no longer advances past steps when the external access integration is stale or revoked, preventing confusing dead-end states.
- New environment snapshot create and restore scripts simplify Snowflake environment backup and recovery workflows.
- Snowflake metadata caching now uses local container storage instead of stage-backed mounts, resolving compatibility issues with WAL-mode databases on FUSE volumes.
- RBAC permission checks now retry once on transient authentication failures in SPCS environments, preventing intermittent redirects away from configuration pages.
- MCP server startup is now deferred via tool manifest caching, reducing initial server startup time.
Logging and Observability
Server logging now provides better context for diagnosing issues in busy production environments.- Every log entry now automatically carries attribution context — including the originating bot, thread, request, and tool — without requiring changes to existing function signatures. Administrators can trace any log message back to the action or user request that produced it.
- Logging anti-patterns were cleaned up across the backend: errors that previously bypassed the logger, overly vague messages on critical paths, and bare exception handlers that swallowed unexpected failures now surface correctly.