# StormForge Agent releases > Release history for the StormForge Agent --- To install the latest version, see the [Upgrade] topic. For past releases, see [Previous StormForge Agent and Applier releases]. [Upgrade]: /docs/installation/upgrade/ [Previous StormForge Agent and Applier releases]: /docs/releases/stormforge/previous/ --- ## Version 3.3.1 _October 9, 2026_ ### Changed * **Dependency updates** ### Fixed * **LimitRange violations**: The Agent now detects configured LimitRanges, so recommendations that violate them are not deployed. ### Security * **Go 1.27.2**: Updated to Go 1.27.2, which includes security fixes to the `crypto/tls` and `net/http` packages. --- ## Version 3.3.0 _October 1, 2026_ ### Added * **Node CPU pressure detection**: The Agent detects Kubernetes nodes under sustained CPU pressure and identifies the workloads responsible. It reads each node's Pressure Stall Information (PSI) from cAdvisor. When a node's CPU pressure exceeds the detection threshold, the Agent finds the containers on that node using more CPU than they request, maps them to their workloads, and creates a recommendation with the `CPUPressureDetected` reason. Each recommendation lists the node, pod, and container, and how far CPU usage exceeded the request. Detection is always on. See [CPU pressure response](/docs/configure/settings/reliability/#cpu-pressure). The Agent reports these metrics for CPU pressure detection: * `sf_container_cpu_pressure_waiting_seconds_total` * `sf_container_cpu_pressure_stalled_seconds_total` * `sf_container_cpu_pressure_waiting_rate` * `sf_container_cpu_pressure_stalled_rate` * `sf_workload_event_last_timestamp_seconds` ### Changed * **Dependency updates** ### Fixed * **Node metrics refresh on provider ID change**: Node metrics now pick up provider IDs that are assigned late or change. Previously, node updates that changed only the provider ID were ignored. * **Provider ID sync to StormForge API**: When cluster data is enabled, the Agent sends a node's provider ID to the StormForge API whenever it changes. Previously, it was sent only when a node was added or its labels changed. --- ## Version 3.2.0 _September 21, 2026_ ### Added * **GPU optimization (early access)**: StormForge can now collect GPU metrics for workloads that use NVIDIA GPUs and show per-workload GPU memory usage and GPU allocation alongside device-level GPU activity. See [GPU optimization](/docs/recommendations/gpu/). * **StormForge Node Agent**: A new optional DaemonSet, `stormforge-node-agent`, collects the GPU metrics from GPU nodes. It is disabled by default; enable it with `--set nodeAgent.enabled=true`. See [Node Agent configuration](/docs/installation/node-agent-config/). * **GPU workload detection**: Workloads that request a GPU, whether through a GPU extended resource such as `nvidia.com/gpu` or through a Dynamic Resource Allocation resource claim, are identified by StormForge. Detection needs no configuration and does not require the Node Agent. ### Changed * **Dependency updates** --- ## Version 3.1.3 _September 17, 2026_ ### Changed * **Dependency updates** ### Fixed * **Recommendation status update conflict**: Resolved a conflict between the pod admission webhook and the Applier when updating recommendation status. --- ## Version 3.1.2 _September 9, 2026_ ### Added * **Applier alerts change diffs**: Applier alerts sent to external endpoints when the applier applies recommendations now include full before-and-after diffs for resources that were changed. ### Changed * **Performance improvements** * **Dependency updates** ### Fixed * **Missing RBAC for Events**: The Applier now includes RBAC to allow creating and patching Events when use of the Edit role is disabled. * **Stale cost metric retention**: The Agent now correctly removes cost-related metrics for deleted pods not owned by workload types StormForge recognizes. Previously, some of these metrics were retained until the Agent restarted. --- ## Version 3.1.1 _July 24, 2026_ ### Fixed * **OpenShift pod scheduling:** The consolidated v3 chart no longer applies a hardcoded pod `securityContext` (`fsGroup`/`runAsGroup`/`runAsUser: 65534`) when `openshift=true` and `commonPodSecurityContext` is unset. --- ## Version 3.1.0 _July 23, 2026_ ### Added * **JVM old-gen live-set metric**: The forwarder (formerly `exporter`) now collects `jvm_memory_pool_collection_used_bytes` for live data size transforms. * **Chart `deepOverride` helper**: Explicit-zero-preserving merge helper so intentional `0` values in overrides are no longer dropped. ### Changed * **`exporter` component renamed to `forwarder`**: The `exporter` Helm values key is renamed to `forwarder`. The old key is deprecated but still works (see the note above). * **`agent` component renamed to `cluster-agent`**: The `agent` Helm values key is renamed to `clusterAgent`, and the agent config file to `cluster-agent.yml`. The old key is deprecated but still works (see the note above); the config file is passed via `--config`, so no action is needed for it. * **Webhook injection scope narrowed to the applier's own pods**: The mutating webhook exclusion no longer covers the entire release namespace. It now skips only pods labeled `webhook.stormforge.io/skip` (which the applier carries). As a result: * StormForge's own workloads in the release namespace become eligible for webhook-based apply. * The release namespace may now appear in `allowNamespaces`. * Helm 4 server-side-apply upgrades work without `--force-conflicts`. * **Quieter logging**: Pod spec generation failures now log at debug instead of error. * Dependency updates ### Fixed * **Applier init container fails behind a corporate CA/proxy**: The applier init container (`applier license-check`, which calls the StormForge API over TLS) now mounts `additionalVolumeMounts`. Previously it mounted no volumes, so customers injecting a corporate CA cert hit `x509: certificate signed by unknown authority` during init and the pod never started. Restores the 2.x chart behavior and matches the `additionalVolumeMounts` docs ("added to all containers"). * **`SuspendState` event silently skipped**: The applier Deployment now projects `POD_NAME` / `POD_NAMESPACE` from the downward API, so the applier can resolve its own Deployment and emit the `SuspendState` event on startup (for example, when setting `suspend=All`). Previously this failed with `ErrSelfPodEnvUnset` and only logged a warning. --- ## Version 3.0.0 _June 10, 2026_ ### Added - **Unified `stormforge` Helm chart**: New chart that bundles the workload-agent and the applier in a single installation, replacing the separate `stormforge-agent` chart. Required values are the same (`clusterName`, `authorization.clientID`, `authorization.clientSecret`). **Breaking change** — replaces `stormforge-agent`. - **`workloadResourceKinds` model**: A single list replaces the two parallel v2 parameters (`workloadResourceTypes` + `ownerResourceTypes`). Each entry carries an explicit `role` field — `workload`, `conditionalWorkload`, `owner`, `ignoredOwner` — eliminating the ambiguity of separate lists. `patchTargetDefaults` (renamed from `patchTargetTypeDefaults`) is set per entry; the chart injects the required `live.stormforge.io/` prefix internally. **Breaking change** — see [Migrate v2 → v3][migration]. - **`includeCommonAddOnResourceKinds`**: Replaces `includeCommonAddOnResourceTypes`, enabled by default. Prepends curated entries for Argo Rollouts, Prometheus operator, OpenTelemetry collector, KEDA, and Dynatrace. ### Changed - **CRDs are now Helm-managed**: CRDs moved from `crds/` to `templates/crds/` so they upgrade in lockstep with the chart. `crds.enabled` (default `true`) opts out for out-of-band CRD management; `crds.keep` (default `true`) prevents cascade-delete on uninstall. **Breaking change** — upgrade note below. - **cAdvisor scrape method auto-detected**: The `CadvisorDirectScrape` config flag is removed. The scrape method is now inferred automatically from the `instance` label on incoming PSI metrics. - **PSI waiting-rate uses a time-weighted window**: The fixed-sample-count ring buffer is replaced with timestamped samples averaged over a configurable `WaitingAverageWindow`. The `SampleInterval` config field is removed; set `WaitingAverageWindow` (default `10m`) instead. Stale samples outside the window are excluded regardless of scrape cadence changes. --- ### Upgrade note — CRD ownership transfer CRDs previously installed from the `crds/` directory carry no Helm ownership metadata. The first `helm upgrade` onto v3.0.0 may fail with "invalid ownership metadata." Remediation options: - **Helm ≥ 3.17**: run the upgrade once with `--take-ownership` - **GitOps / older Helm**: label and annotate the existing CRDs before upgrading: ```sh kubectl label crd app.kubernetes.io/managed-by=Helm --overwrite kubectl annotate crd \ meta.helm.sh/release-name= \ meta.helm.sh/release-namespace= --overwrite ``` [migration]: /docs/installation/migration/ --- Section pages: - [Previous StormForge Agent and Applier releases](/docs/releases/stormforge/previous/)