Berth 0.5.0, 7 October 2026

In short: a smaller AI SRE toolset, agent loops that end properly, corrected cost estimates, and published evaluation data. Still a restricted beta on a single cluster; no new SSO, per-user authorisation, audit viewer or multi-cluster support.

Changed

The assistant has 15 tools instead of 18. Three were merged into others: reading any single object is one tool (describe_resource, with an optional apiVersion for custom resources); discovering and listing resources is one tool (list_resources); and cluster_health now includes each node. The security behaviour is unchanged.
The default model for the Anthropic API is claude-sonnet-5-5. A model you saved in Settings, or pinned in your Helm values, is kept.
With Ollama, when the assistant runs out of tool calls it is now told to answer from what it has found, and a server error from Ollama is retried once.

Fixed

With Claude (Anthropic API and Bedrock), a run that hit the tool-call limit could return an empty or half-finished answer as if it had succeeded. It now produces an answer, marks one that was cut off, and reports an error when there is none.
The usage-cost estimate for Claude Opus 5.5 was 25% too high.

Compatibility

The tool names in the streamed tool events changed: get_resource, list_api_resources and list_nodes no longer exist. Anything that matches on those names needs updating.

Evidence

The scenarios, results and scoring changes behind these changes are at /evals/, with their limits.

Upgrade

helm upgrade --install berth oci://ghcr.io/unishsys/charts/berth \
  --version 0.5.0 --namespace berth --create-namespace --reuse-values