Configuration and Metrics Reference

Use these references to connect a Console label, a configuration field, and the corresponding CloudPilot AI MCP read. They explain what a value means before you decide whether to change it.

Area	Reference	Questions it answers
Workload Autoscaler (WA)	Policies and workloads	How are recommendations calculated? Which policy applies? What can trigger a restart or an in-place update?
Node Autoscaler (NA)	NodePools	Which nodes may be provisioned, when can they consolidate, and what constrains capacity?
Cloud provider configuration	NodeClasses	Which image, identity, network, disk, and kubelet settings do new nodes use?
Optimization and scheduling	Node optimization	What enables optimization, protects workloads, or schedules disruption?
Cluster and organization administration	Operations	What do settings, upgrades, Image Accelerator, permissions, and organization configuration mean?
Health, history, and costs	Observability	How should resource metrics, readiness, alerts, events, PVC usage, and Saving Report be interpreted?

For installation and walkthroughs, start with Getting Started, Workload Autoscaler, or the Node Pool Configuration Guide.

Read Configuration Before Recommending Changes

Documentation explains product behavior; live reads establish whether that behavior applies to the customer’s cluster. A useful investigation follows this order:

Resolve the authorized organization and cluster. Read get_cluster_overview for the cluster’s provider, versions, capabilities, and current state.
Read the relevant configuration with an exact-get tool. For example, use get_recommendation_policy or get_node_class, rather than inferring the full configuration from a list summary.
Compare the configuration with health, events, and the relevant history window. A policy being enabled does not prove that its controller is ready or that a change has been applied.
Explain the applicable fields, constraints, and trade-offs. Separate an observed fact from a proposed change, and identify missing evidence before giving a recommendation.

The references describe reads, not an authorization to write. A configuration example is not automatically suitable for every workload. Before applying a change through Console, Kubernetes, Terraform, or another supported interface, review the affected scope and disruption behavior.

Defaults Have Different Sources

Do not treat every value shown by Console as a server default.

Default source	Meaning	How to verify
Console draft	Value prefilled when creating a resource; it may not have been saved.	Inspect the editor and the saved detail response separately.
Built-in policy or generated template	Value explicitly written into a named policy or provider-specific default resource.	Read that policy or the tool’s documented default-template option.
Omitted-spec behavior	Behavior of the installed CRD or controller when the field is absent.	Check the applicable schema and controller version, not just a Console form.
Existing configuration	Value currently stored for one resource.	Read the exact resource; it can differ from all of the above.

For example, WA’s built-in policy values and a newly opened Console draft need not match. Similarly, a provider’s default NodeClass is not a portable template for another cloud.

Availability and Versions

These references cover the documented configuration surfaces and identify source-only or version-dependent fields where release support is not established. They are not a promise that every field is available on every deployed MCP server or cluster.

Check the connected server’s tool catalog and input schema for available reads and arguments. A documented Console feature can still lack an MCP read; those gaps are called out in the relevant reference.
Check the installed Agent, Workload Autoscaler, provider controller, Kubernetes version, and CRD where a feature requires them. An API server library version alone does not establish the cluster controller’s version.
A raw nodePoolSpec, nodeClassSpec, or policy spec preserves that resource’s shape. It does not imply that every possible upstream CRD field is editable through Console or supported by every CloudPilot AI release.
Treat an absent field as unknown or omitted according to that field’s contract, not automatically as false, zero, disabled, or unsupported. Do not replace missing metrics with zero.

Reading MCP Results

Many CloudPilot AI read tools use a common envelope around data:

Field	Interpretation
`schemaVersion`	Version of the response contract, not the installed Agent or controller version.
`organizationId`	Organization scope, when present. Do not mix resources from different organizations.
`observedAt`	Time the response was assembled. Individual metrics and historical series can refer to earlier timestamps.
`status`, `warnings`, `sources`	Completeness and source outcomes. Inspect warnings before treating a partial result as a complete diagnosis. Tool-call errors must also be handled separately.
`summary`	A short description, not a replacement for the structured configuration or metrics.
`data`	Tool-specific values. Paths in this reference are relative to this structured result, not a text rendering.
`consoleUrl`	Link to continue the investigation in Console. It is not a credential or an authorization grant.

Individual tools may add their own completeness indicators, unavailable reasons, units, and timestamps. Use those indicators as well as the common envelope.

Lists, Details, and Pagination

Use the selected tool’s documented filters and sort fields. Not all tools accept the same arguments or return the same projection.

Follow nextCursor while hasMore is true, preserving the original organization, cluster, filters, sort, and page size. Do not invent or modify opaque cursors.
A page is not the entire inventory. returned describes the page; matching, where provided, describes the matching set reported for that request. A changing inventory is not a transactional snapshot across calls.
Use list summaries to find relevant resources, then exact-get tools for their full specifications. Fetching every detail is unnecessary when the question concerns only a filtered subset or an aggregate.
For a full inventory, process all pages and report any truncation or failed page. Do not claim completeness after hitting a tool, response-size, or agent-context limit.

Common Investigation Scenarios

Scenario	Evidence to gather	Reference
Reduce workload requests safely	Matched policy, recommendation, current requests, readiness, usage history, HPA ownership, and restart/in-place constraints.	WA scenarios
Explain Pending Pods or unavailable capacity	Pod requirements, NodePool requirements and limits, NodeClass readiness, and provisioning events.	NodePools, NodeClasses
Limit optimization disruption	Optimization state, workload protections, pool budgets, schedules, and maintenance state; no single switch describes all disruption paths.	Node optimization
Explain savings or utilization changes	Baseline selection, comparison window, metric units, cost basis, and missing-data warnings.	Observability
Recommend an administrative setting	Current setting, provider/version applicability, user role, and the effect of changing it.	Operations

If the available reads do not expose a required field, say so and use the documented Console or Kubernetes inspection path. Do not guess the value from the field description.