v1.19.99

Genesis v1.19.99 Release Notes

October 17, 2026

Palantir Foundry Workshop apps can be migrated end to end into a Genesis App proven to match the original.
Claude Opus 5.5 is now the default chat model and Claude Haiku 5.5 the default fast model on Anthropic, Bedrock, and Databricks.
Audit Log entries now show who made a change, any agent acting for them, and masked old and new values across connections, roles, secrets, tokens, and settings.
Managed sources can sync shared Git standards on a schedule, including agent and role definitions, with retries, results on the Managed Sources page, and audit entries.
Admins can now see and restrict who can chat with each agent, and use View as to check what a specific person or role can access.
Jira projects can be scanned into the Context Graph, and Jira, Power BI, and Confluence scans can be limited to chosen projects, workspaces, or spaces.

Grant each blueprint’s execute privilege to roles that should keep starting it. Agent access alone no longer allows a start. Admins and blueprint owners are unaffected.
Give Manage Users to people who manage Git sources carrying role definitions. Manage Repos alone is now refused for those sources.
If a managed Git source has auto-promote off, promote its synced flow changes from the Flows page or Managed Sources panel. They no longer apply automatically.
For Foundry tokens limited to the preview SQL role, grant the SQL read role on affected projects. Results capped at 1,000 rows are now refused instead of returned.
Pass an explicit evidence directory to Tier A install, teardown, egress-probe, and open, and a pull-token file to install. Omitting them now fails immediately.
Re-run the Tier A install on existing clusters before automatic upgrades, and confirm fresh-install nodes fit 1 CPU, 4 GiB memory, and a 50 GiB volume.

Launchpad tasks start with tap-to-answer cards over the chat, and agent tool calls fold into a single Working line while Eve works.
Goal mode is now the default for shipped and new blueprints. Approvals and choices stop in the thread. Pass continuous mode to keep unattended runs.
Foundry migration projects record copies, notebook runs, and goal progress automatically. Chat copies are filed under your project before they start.
Common AI failures in chat and blueprint recommendations show a clear, actionable notice with a copyable support ID instead of raw provider errors.
Chat approval cards are now compact and show what will run and why, with Approve and Decline buttons and the rest under Details.

LiteLLM Proxy sends required OAuth headers such as module or tenant headers, fixing authentication failures against those gateways.
LiteLLM Proxy model discovery works with gateways that list models as a bare array, and Test Connection names the specific failure cause.
Snowflake key-pair queries over PrivateLink sign in with the full account name, resolving invalid JWT token errors. Traffic stays on PrivateLink.
Enabling custom CA certificates no longer crash-loops the pod when extra environment settings point SSL_CERT_FILE or similar variables at a custom bundle.
Context Graph embeddings start working as soon as you save an embedding provider connection, without a server restart.
Agent-run Python code that exits cleanly is reported as successful. Passing blueprint required checks no longer show as failed.