Release notes

Olympusx AI 1.1

Olympusx AI 1.1 adds two use cases a tenant turns on, the Code Assistant and Security, a graph of every knowledge collection, and a platform-side Runtimes page that brings serving images into an air-gapped network and keeps them current. It also changes how long model calls may run, so reasoning models are no longer cut off mid-thought.

To move an installation to 1.1, see Upgrade.

New: use cases

The rail has a new Use cases group, between My conversations and Build. It holds finished jobs a tenant turns on and its people use, rather than parts they build with.

Code Assistant

Editors on your models. A developer mints their own developer key and copies a ready configuration for Continue, Cline, Aider, Zed or plain curl, filled with the gateway’s address and the models the tenant chose for chat and autocomplete. A tenant admin sets how long developer keys live and how many each developer may hold. See Connect your editor and Developer keys.
Code-aware guardrails. Traffic from developer keys is checked under a code profile: secrets and prompt injection are checked, personal data in code is not rewritten, and answers stream. It is never captured for review.
Usage per developer. Developers read their own usage; admins and auditors read spend per developer. See Usage and spend.
Every API key records its owner.
Codebase knowledge. Index repositories from GitLab, GitHub and Bitbucket Cloud, by group or by repository, with a read-only token or none for public code. Code is cut at functions, methods and types, scanned for secrets before it is stored, and synced daily. A wizard adds a source; a source can be edited later. See Code sources.
Repository pages. Each indexed repository has a page with its details from the host, its statistics, a search over its code, and a Graph of its files and symbols with their imports, calls and inheritance, with focus, filters by kind and a full-screen view. See Repository view.
What uses this. From any file or symbol, list what calls, imports or inherits it, up to three steps out, and copy the list as Markdown.
Ask this repository. A cited answer about a repository from its own code.
Codebase tools for editors. Continue and Cline, and any client of the Model Context Protocol, can call search_codebase and find_usages at the gateway’s /mcp address with a developer key. See Code tools.
Merge request review. A webhook from the code host starts a review when a merge request is opened, reopened or given new commits; anyone using the Code Assistant can also ask for one by number. The diff is read with secrets removed, reviewed for bugs, security, performance, maintainability and tests, and, with a separate review token, posted to the merge request as comments and a summary. See Merge request review.

Security

Code assessments. A reasoned security assessment of an indexed repository, producing findings with their evidence. See Code assessments.
Findings from every source, in one list, with a Sources tab, dismissal with a reason, and export as SARIF. See Findings and Exports.
Investigations of the audit log. An admin asks a question; Olympusx AI queries the tenant’s audit log in rounds, and reports a verdict and findings that rest on the rows it shows. See Investigations.
Investigations of your own logs. Connect OpenSearch or Loki as a log source, and investigate it the same way. See Log sources.
Alerts that open investigations. Alertmanager, or any SIEM that can post JSON, sends alerts to a log source’s webhook; each firing alert opens one investigation around its time. See Alert integration.
Sigma rule drafts from investigations of a log source whose verdict is suspicious or malicious. See Exports.
Hunts. Three preset checks over the audit log, run daily when turned on. See Hunts.
Actions for a person to approve. A verdict can propose actions; none runs until an admin approves it. See Actions and approvals.

New: knowledge

The collection graph. Every knowledge collection gets a graph of its documents, their sections, and the people, organisations and places they mention, rebuilt as the collection changes and drawn beside the collection’s search. See The collection graph.

New: platform

Runtimes. Model serving › Runtimes lists every serving image the platform can use, with its release notes, the security fixes it carries, and the one recommended for each kind of model. A new image can be inspected on a node you choose before it is offered. See Runtimes.
A registry per cluster. Each cluster can name the container registry its nodes pull from, with a pull credential that Olympusx AI puts on every pod it creates there. See Clusters.
Copying runtime images into a cluster’s registry, checked by digest, directly or through the bastion. The bastion also fetches runtime release notes for an air-gapped platform. See Runtimes and Remote access.
Reasoning and tool-call support in the serving runtime, so reasoning models return their answer separately from their thinking and tool calls come back structured.

Changed

Model calls are cut on silence, not on length. A reasoning model can think for minutes before its first token, and a limit on the whole response cut it off. Three limits are now platform settings, in seconds: model idle timeout (default 300), gateway call timeout (default 1,800) and Security call timeout (default 600). See Settings and API limits.
Runtime flags are a platform administrator’s. Extra serving arguments and permission to run a model’s own code are set only by a platform administrator.
Secrets are redacted completely. Every copy of every detected secret in a prompt or answer is masked, including repeated values and keys written across lines.
The guardrails service runs under any user ID, as OpenShift assigns them.
The bastion runs under rootless Podman.

Known limits

Each of these is stated so nobody mistakes it for a defect.

Changing the master key makes every sealed credential unreadable. Generate it once and keep it in escrow.
Budgets are the control a tenant manages. The public assistant chat has its own default of 60 requests a minute per tenant.
Retention windows are set over the management API only.
Some guardrail settings are API-only: the personal-data types to detect and the detectors’ thresholds.
A key’s model allow-list is API-only. The console’s key wizard issues keys for every granted model.
The model catalog is visible across tenants to the roles that read it: Admin, Builder and Auditor.
Images are built for amd64 only. Schedule Olympusx AI and its model servers on amd64 nodes.
Merge request review and codebase knowledge need the guardrails service, because every diff and every file is scanned for secrets before it is kept.