1. Docs
  2. Changes and releases
  3. AI Agents

AI Agents

  • AI Agents agent-platform v1.19.0

    Added

    • klausGateway.obo.persistence (umbrella + connectivity values.yaml and values.schema.json) is forwarded to the klaus-gateway subchart, which backs the encrypted link store with a ReadWriteOnce PVC so a human’s Slack<->muster link survives gateway pod recreation (rollouts, node drains), not just in-process restarts. Disabled by default (emptyDir, unchanged). Requires the klaus-gateway chart that supports it (giantswarm/klaus-gateway#118, v0.10.3). Pairs with that release’s store-key fix, which accepts the SOPS-staged base64 obo.storeKey directly so obo.storePath can finally be set without the gateway crashing on boot.
  • AI Agents muster v0.21.1

    Fixed

    Full Changelog: https://github.com/giantswarm/muster/compare/v0.21.0...v0.21.1

  • AI Agents muster v0.21.0

    Added

    Full Changelog: https://github.com/giantswarm/muster/compare/v0.20.0...v0.21.0

  • AI Agents klaus v0.0.252

    Changed

    Full Changelog: https://github.com/giantswarm/klaus/compare/v0.0.251...v0.0.252

  • AI Agents muster v0.20.0

    Added

    • Durable workflow execution tracking (#930, supersedes the closed-but-unimplemented #35): workflow execution records now survive process restarts and are visible across replicas when muster runs in Kubernetes.
      • WorkflowExecution CRD (muster.giantswarm.io/v1alpha1, kind WorkflowExecution, short name wfe). Each workflow run is persisted as one immutable record (metadata.name is the execution UUID) carrying the workflow name, status, timings, input, result, per-step records, and a truncated flag. It has no status subresource — it is an append-only record, not a reconciled resource. The chart’s ClusterRole gains workflowexecutions permissions and the CRD ships in both the muster-crds chart and the app chart’s crds/ directory.
      • Backend selected by deployment mode. In Kubernetes mode the existing ExecutionStorage seam is backed by the WorkflowExecution CRD (via the controller-runtime client already embedded in MusterClient); standalone muster serve keeps the filesystem backend. core_workflow_execution_list/get behave identically against either backend. List filtering uses muster.giantswarm.io/workflow and muster.giantswarm.io/status label selectors, sorting and paginating in memory.
      • Payload guard. Before persisting, an execution record whose marshaled size exceeds 256 KB (well under etcd’s object limit) has its oversized result and step results replaced with a truncation marker and is flagged truncated, so a large workflow result can never make a record unpersistable in either backend.
      • Retention GC. A background goroutine prunes execution records that are older than 21 days or beyond the newest 10000 (burst safety cap), keeping the store bounded without manual cleanup. The clock is injectable so the prune logic is deterministically unit-tested. Retention bounds are constants for v1.
      • Execution metrics. New OTel instruments muster.workflow_executions (→ muster_workflow_executions_total), muster.workflow_execution.duration (→ muster_workflow_execution_duration_seconds, keyed by workflow + status), and muster.workflow_execution.store_errors (→ muster_workflow_execution_store_errors_total) are exported via the existing OTel→Prometheus pipeline. A persistence failure is now logged at error level and counted instead of being silently warned, so empty dashboards become an observable signal.
  • AI Agents muster v0.19.7

    Fixed

    • filter_tools query ranking now weights tool-name matches over description matches and down-weights the ubiquitous CRUD verbs (list, get), so an intent query is driven by its discriminating noun rather than a generic verb. Previously name and description were concatenated into one BM25 document scored token-equally, so list pods surfaced pagerduty x_pd_list_* and core_*_list above x_kubernetes_list and the pod workflows (the generic list token dominated). The ranker is now a BM25F-style field-weighted lexical scorer: name and description are scored as separate length-normalised fields, name matches weigh higher, and stop-verb contributions are scaled down (down-weighted, not dropped, so a list-only query still ranks list-shaped tools). (#931)
  • AI Agents muster v0.19.6

    Fixed

    • Workflow validation no longer requires a top-level tool on parallel/forEach container steps. The reconciler’s validateWorkflow (the path that sets status.valid/status.validationErrors for CRD-applied workflows) unconditionally demanded step.tool, so any workflow whose only tool-less step was a parallel group or forEach loop was wrongly marked invalid (step '<id>': tool is required) even though it executed correctly. Validation now mirrors the structured create/validate path: a step must specify exactly one of tool/forEach/parallel, container sub-steps still require their own tool, and status.referencedTools now includes sub-step tools. (#928)
  • AI Agents klaus-operator v0.0.119

    Fixed

    Full Changelog: https://github.com/giantswarm/klaus-operator/compare/v0.0.118...v0.0.119

  • AI Agents klaus v0.0.251

    Fixed

    Full Changelog: https://github.com/giantswarm/klaus/compare/v0.0.250...v0.0.251

  • AI Agents klausctl v0.2.25

    Fixed

    Full Changelog: https://github.com/giantswarm/klausctl/compare/v0.2.24...v0.2.25