v0.5.0

Berth is a self-hosted Kubernetes dashboard with an optional read-only AI SRE. 0.5.0 is a release about the assistant: a smaller toolset, sturdier agent loops, corrected cost estimates, and published evaluation data. It is still a restricted beta for trusted cluster administrators on a single cluster. No SSO, per-user authorisation, audit viewer or multi-cluster support has been added, and the assistant still has no tool that writes to your cluster.

What changed

• 15 tools instead of 18. get_resource is now describe_resource with an optional apiVersion; list_api_resources is folded into list_resources (no kind discovers kinds); list_nodes is folded into cluster_health (per-node rows). Tool definitions are 26% smaller.
• Agent loops that end properly. Ollama: the forced final answer after the tool loop is exhausted now tells the model to answer from the evidence, and a 5xx from Ollama is retried once. Claude (Anthropic API and Bedrock): the last permitted call is made with tools withheld, a cut-off answer is marked, and an empty answer is an error instead of an empty success.
• Default Anthropic model is claude-sonnet-5-5 (was claude-opus-5). A saved model or a pinned ai.anthropic.model is kept.
• Cost estimates for Opus 5.5, Sonnet 5.5 and Fable 5.1 are now correct (the Opus 5.5 estimate was 25% high).
• Published evaluation data, in evals/: 17 synthetic scenarios, repeated local-model runs, an adjudication log and a claims ledger, with their limits stated.

What did not change

The security boundary: raw logs and describes are still off unless enabled; Secrets and ConfigMaps are still refused whatever arguments are passed; budgets are still admission controls, not a billing cap. The Bedrock default model and EFFORT default are unchanged.

Upgrading

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

Compatibility note. The tool events streamed by the chat endpoint now name describe_resource, list_resources and cluster_health where they used to name get_resource, list_api_resources and list_nodes. Anything that matches on those names needs updating.

What we measured, honestly

On 17 synthetic scenarios with a local model, the old and new toolsets passed 45 of 51 and 44 of 51 runs: the smaller toolset did not change the pass rate, and it cut input tokens per model call by 14%. The engine fixes removed two failure modes (answers that were fragments of narration, and Ollama 500s). These are small synthetic tests on one machine; they say nothing about real clusters, MTTR, or cloud models, which we have not yet run. See evals/CLAIMS.md.

Verify the image and chart before you install: see the cosign and checksum commands in the README.

Full changelog: CHANGELOG.md. Evaluation write-up: https://berth.cloudguys.io/evals/

Container image: ghcr.io/unishsys/berth:0.5.0. Helm chart: oci://ghcr.io/unishsys/charts/berth (--version 0.5.0). Binaries below; verify them against SHA256SUMS.