Skip to main content

​
Genesis v1.19.89 Release Notes

August 29, 2026 This release includes the changes listed below.

​
AI Providers & Models

  • Nested dbt incident analysis now uses the same AI provider and model as the thread that triggered it instead of always defaulting to OpenAI.
  • The AI Providers connection test now checks each chat-model tier individually, flags which specific tier has an invalid model, and lets you test and re-verify tiers independently before saving.
  • Claude Opus 5’s context window is raised to 600,000 tokens on Anthropic, AWS Bedrock, and Snowflake Cortex, and context usage indicators and soft-limit presets now accurately reflect each thread’s real token limits.
  • Claude Opus 5 and Opus 4.8 now use the platform’s full 64,000-token output ceiling on Snowflake Cortex by default instead of being capped far lower, and output-token warnings now compare against each model’s own configured limit.
  • The AI Providers page now correctly shows the default badge for AWS Bedrock, Anthropic, Gemini, and other LiteLLM-backed vendors instead of showing a missing-default banner.
  • Databricks now defaults to Claude Opus 5 for its premier tier and Claude Sonnet 5 for its main tier, raises Opus 5’s context window on Databricks to 600,000 tokens to match other providers, and returns a clear error instead of crashing when a provider response has no content.

​
Agents, Tools, and Git

  • Listing repositories with an active blueprint project no longer incorrectly returns only the project’s own repository, and file, git, and dbt actions now return a clear error instead of running against the wrong repository when it can’t be resolved.
  • dbt manifest requests now report whether the manifest is cached and when it was generated, support forcing a refresh, surface dbt’s actual compile error output on failure, and no longer risk a race that could delete the manifest during a concurrent refresh.
  • dbt tool failures now distinguish genuine errors from safety-policy blocks and identify which specific safety rule triggered a block, both in the response and in the audit log.
  • dbt safety checks no longer block calls to dbt’s standard logging function or flag redacted and masked credential values as exposed secrets, while still catching real secret values.
  • The Google Cloud Storage file download action now correctly reports the number of bytes downloaded instead of always returning null on success.
  • AutoDev pull requests now keep the coding agent’s own title instead of falling back to a generated name, build a more informative fallback title from the latest commit when needed, and report which workflow step failed when a run produces no evidence.
  • Databricks notebook export, fetch, and upload actions now correctly report and detect SQL and Scala notebooks instead of mislabeling them as Python, and cluster-restart suggestions no longer appear for workspaces with no clusters to restart.
  • Stopping a tool now cancels the Databricks job run or command still executing in the workspace.
  • Bot detail reads now report the actual stored last-modified timestamp, or null if none exists, instead of fabricating a current timestamp on every read.

​
Chat, Threads, and Genesis Apps

  • Archived threads, including old imports, now carry the same title shown in Claude Code.
  • Chat now shows the complete agent response without requiring a thread reload when the final streaming update is missed.
  • Renaming a thread from the sidebar now updates the visible name immediately without a page refresh, and a failed rename shows an error message instead of silently closing.
  • Artifacts and checklist indicators in the chat thread actions menu no longer flash or pulse for new items, and the thread actions menu no longer overlaps the Artifacts/Checklist panel’s close button.
  • Genesis agents can now receive and answer authenticated Google Chat direct messages within an allowed Google Workspace domain, including starting fresh conversations and linking to Requests Center work, and refreshing, disabling, or deleting an agent no longer leaves a stale or dropped chat connection.
  • Forked and imported threads now keep the model they actually ran on so context-window sizing no longer falls back to a generic default, the context-budget display and compression engine always agree on a thread’s model, and operators can explicitly override a thread’s soft context limit above its hard limit up to a 2,000,000-token ceiling with the outcome shown instead of failing or closing silently.
  • Chat rendering performance is improved on large threads through more efficient list rendering and scrolling, and the floating date chip and find-in-thread state no longer show stale information after switching threads.
  • Chat threads now load only recent messages initially with older history fetched on demand, thread switching no longer freezes or rebuilds the message list from scratch, scrolling into history keeps your reading position stable, and markdown and side panels no longer re-render on every keystroke.
  • Chat message hover actions no longer duplicate icons on short messages, response rating thumbs are now available on every message from the under-message toolbar, and keyboard users can still reach avatar-gutter actions that are visually hidden on short messages.
  • Databricks notebook paths shown in chat tool-call results are now clickable links that open the notebook or folder directly in the Databricks workspace.
  • A chat socket-readiness timer that could fire after a user left the page, causing an unhandled error and a duplicate thread subscription attempt, is now reliably canceled.

​
Connections and Configuration

  • Connections can now be filtered by owner, scope, and effective authentication type and sorted from a search-adjacent popover with shareable URL state and searchable owner options, filtering across the complete role-visible set with authoritative counts and immediate result updates.
  • Bulk reads of application configuration no longer truncate results.
  • Blueprint detail views now display the actual creator’s name instead of a placeholder, and Journey debriefs now show the correct blueprint phase without creating duplicate versions from overlapping regeneration requests.
  • The agent LLM tab’s default context soft limit now warns and requires confirmation before it is set above the model’s real context limit, clamps and reports any bot-level default set above that hard limit, and keeps a valid soft limit intact if the override acknowledgement can’t be read or the bot’s model changes.
  • Pasting a URL that matches an existing connection now reuses that connection instead of falling back to a web search.
  • Administrators can now see why a managed-source blueprint is hidden and enable or disable it directly from the Managed Sources page instead of it staying stuck and unrecoverable.
  • Administrators can now choose which model powers the wiki librarian’s H2 correction classifier through a configuration setting.
  • A single blueprint manifest with malformed fields no longer breaks the entire blueprints listing; affected blueprints are now skipped and named instead of hiding every blueprint.
  • Files selected during Launchpad onboarding are now uploaded before target selection so the resulting mission actually receives those files.

​
Connectors

  • Azure Data Factory is now available as a connection type authenticated by a Microsoft Entra service principal, letting agents list, read, create, and update pipelines, datasets, linked services, and triggers through that connection.
  • AWS SigV4-signed web requests no longer fail by mistaking the authentication type for a secrets-manager reference, and SigV4 signing support is now included in standard runtime installations.
  • Removing one Talend-registered project no longer risks deleting another Talend project’s scanned lineage data.
  • DuckDB connections now use a configurable per-connection memory limit, defaulting to 2 GB, preventing out-of-memory crashes in memory-constrained deployments.
  • Read-only SQLite schema-introspection queries no longer require confirmation as if they were destructive operations.
  • Requesting schema discovery against an unsupported database connection or action now returns a clear message naming the connection and its database type instead of a vague error or empty result.
  • SQL execution permissions are now determined by a connection’s actual stored type rather than a guess based on its name.
  • A native crash triggered by querying an already-closed Databricks connection no longer takes down the whole server, and Databricks Connect sessions no longer share a single connection or have long-running queries mistaken for idle and torn down.
  • The Databricks cluster-listing tool no longer reports serverless compute as unavailable when no all-purpose clusters are found, and now explains that serverless job compute and SQL warehouses may still be usable.
  • Creating a Databricks notebook whose documentation cell contains a fenced code example no longer discards the notebook’s actual code, and any ambiguous fence arrangement is left untouched with a warning.
  • Stopping a running tool call now cancels the underlying query on the database server for SQL Server, Postgres, MySQL, Snowflake, Databricks, and other SQLAlchemy-backed connections, fixes context graph Snowflake scan cancellation, and adds configurable timeouts for how long a stop can wait on the server.
  • Databricks notebook cells that mix a markdown heading with code and no separator are now flagged before import instead of silently importing as non-executing markdown.
  • Databricks git sync now writes each notebook header using the correct comment syntax for its language, fixing invalid SQL and Scala notebooks produced on import and no longer stacking a duplicate header on files that already had one.
  • Knowledge base write-back now supports editing YAML, JSON, text, and spreadsheet sidecar files alongside markdown pages, reports exactly which files changed in each pull request, and no longer leaves inconsistent staged changes when a batch export fails or a source uses a hidden documents root.
  • BigQuery connections can now authenticate using a deployment-provided identity instead of a stored service-account key, and converting or sharing these identity-based connections now requires connection-administrator permission.

​
Context Graph

  • Cancelling a Context Graph scan now stops the scanner and terminates the Snowflake query it had left running.
  • Runtime lineage discovery no longer creates placeholder nodes for databases, schemas, or tables excluded from a scan, and now counts those exclusions in scan statistics instead.
  • Agents can now retire an outdated Team Wiki learning using the statement identifier actually shown to them, fixing a bug that silently rejected valid retirement requests and left contradictory learnings active on the same page.
  • Team Wiki learnings anchored to items not yet scanned into the Context Graph are now readable and browsable instead of returning a not-found error, and pages built only from such learnings are now clearly marked as lacking a backing graph node.
  • Semantic search no longer fails outright when the index contains documents embedded at mixed dimensions; it now serves the results it can, automatically self-heals the affected documents, and reports when part of the index is temporarily degraded.
  • Context Graph lineage queries can now filter by edge type to separate data flow from orchestration and runtime relationships, upstream and downstream results now include the connecting edges, and the graph view styles each edge type distinctly.
  • Snowflake and Databricks warehouses and compute clusters now appear as nodes in the Context Graph with their configuration details, and new identity and deployment node types lay the groundwork for tracking what runs where.
  • Context Graph node and edge properties can now be safely updated by multiple scanners at once without one overwriting another’s contributions.
  • dbt Cloud job-to-model lineage in the Context Graph now reflects only the models each job actually runs instead of connecting every job to every model, and stale lineage edges from prior scans are now cleaned up automatically.
  • Airflow DAGs are now linked to their source file and owner identity in the Context Graph, with schedules normalized into a structured format and stale links reconciled automatically on rescan.
  • Context Graph runtime discovery now records which identity and warehouse each execution runs as.
  • The Context Graph now records each Snowflake task’s warehouse, schedule, state, and owner, links tasks in their execution order, and captures Snowflake pipe ownership and notification settings, refreshing these bindings on every scan.
  • GitHub Actions workflow scans now capture runner assignments, deployment environments, cloud identity roles, and schedule expressions in the Context Graph, upgrading previously recorded links in place on rescan.
  • Databricks jobs in the Context Graph now record their schedule, execution identity, and git source, with fixes for broken links to git-sourced code, incorrect staleness flags on lower-case names, and unresolved warehouse edges.
  • Snowflake scans no longer re-embed unchanged tables, views, procedures, and functions on every rescan, eliminating unnecessary reprocessing load.

​
Data and Analytics

  • BigQuery table previews in the Data Explorer no longer fail for connections whose project id contains a hyphen.
  • The data-discovery search tool no longer reports incorrect or fabricated column names and counts for tables with multi-parameter column types, and the Context Graph lineage server no longer fails to start when launched without a PYTHONPATH environment variable set.
  • Data-context search no longer repeats each result multiple times, which previously caused large result sets to lose real hits to truncation.
  • Query results that were cut off by a row, byte, or field size limit now clearly report that truncation occurred instead of appearing complete.
  • Data-context search no longer lists the same database connection twice when one of its tables ranks higher in the results.
  • Presenter RCA dashboard evidence links now correctly identify failed and recovered states and open the investigation assistant for follow-up, chart cards keep all axis labels visible, and archived tailored demo versions can be restored to active pickers.

​
Platform Reliability

  • Document-index lookups now use bounded, memory-safe reads instead of scanning every document chunk, newly registered MCP servers become discoverable without the prior delay, and document indexes stranded by an embedding-model switch no longer show up as valid search capabilities.
  • A crashed thread-service worker is now detected almost immediately, its in-flight requests fail with a clear error, and a replacement worker starts automatically instead of stalling or hanging the application.
  • Queries routed through the Iceberg worker no longer trigger a hidden crash while resolving phase attribution, and worker-side error tracebacks are captured by the error log again for diagnostics.

​
Security

  • Database connection engine caches are now isolated per set of credentials so cached engines can no longer be reused across different credentials.
  • Connection access handling has been hardened to reduce the risk of unauthorized access to configured connections.
  • A high-severity denial-of-service vulnerability in a bundled compression library used in the container image has been patched.

​
User Experience

  • The chat status pill now shows the correct platform name, such as Databricks, for cloud-platform tool activity instead of a generic file-operation label.

​
Preview (opt-in)

  • A failed sandbox-runner turn now surfaces a visible error instead of failing silently.

​
Deployment Artifacts

Use these exact references when deploying this release:
  • Genesis release: v1.19.89 (promoted from v1.19.89-rc.6)
  • Docker Hub image: genesiscomputingai/genesis:v1.19.89
  • Docker Hub digest: genesiscomputingai/genesis@sha256:3f650318f47ccc81c25c3d20b9f129ba6b169463aab100d7c36703adf2be43bc
  • Helm chart: oci://registry-1.docker.io/genesiscomputingai/genesis
  • Helm chart version: 0.3.69 (appVersion: v1.19.89)