Helm values

The chart renders byte-for-byte the same resources as the base manifests with default values (the make helm-verify no-drift gate). Every production hardening is a values-gated dial defaulting to off/dev, so turning one on renders the extra resources and leaving it off renders nothing extra. This page is the values reference.

The fastest path to a correct production posture is the shipped overlay:

helm pull oci://ghcr.io/ctxmesh/charts/ctxmesh --version 0.1.0-beta.8 --untar   # for values-production.yaml
helm install ctxmesh oci://ghcr.io/ctxmesh/charts/ctxmesh --version 0.1.0-beta.8 \
  -f ctxmesh/values-production.yaml \
  [+ your own overrides]

Note

Values below are confirmed against the chart. Some downstream coordinates (registry paths, external endpoints) are install-specific and you supply them. Confirm on your cluster with helm show values.

Profile & security

Value	Default	Meaning
profile	"" (dev)	production makes the HA invariants hard — helm template fails unless the control plane is actually HA (state-layer proxy, BFF, run-worker all ≥2, dispatch on, no bundled dev data plane).
security.tenantLabelWebhook.enabled	true	The tenant-label ValidatingWebhook (forbids a non-controller principal from changing a namespace’s agents.ctxmesh.ai/tenant label). Boots its own in-process cert controller (no cert-manager).
security.tenantLabelEnforcement	""	Opt-out ack. To run production without the webhook, set to the exact string "unenforced-UNSAFE-acknowledged" so a reduced posture can’t ship silently. Ignored when the webhook is enabled or under the dev profile.
namespace	ctxmesh	Install namespace; must match the base manifests for no-drift.

Control plane

Value	Default	Meaning
controllerManager.replicas	1	Active/standby leader election; set 2+ for HA.
controllerManager.podDisruptionBudget.enabled	false	Opt-in PDB (minAvailable below). Enable alongside replicas>1.
controllerManager.podDisruptionBudget.minAvailable	1
controllerManager.image.repository / .tag	controller / "" (→ appVersion)	Point at a signed image in your registry, digest-pinned, for production.

Stateless request-servers (replicas, PDBs, images)

Value	Default	Meaning
gateway.replicas	1	Model gateway (stateless); 2+ = plain N+1.
tokenService.replicas	1	Credential plane (stateless); 2+ = plain N+1.
tokenService.tls.required	false	true ⇒ the credential plane refuses to serve plain HTTP (no silent downgrade). production requires true.
tokenService.image.repository / .tag	token-service / ""	Signed image in production.
statelayerProxy.replicas	1	The load-bearing memory/quota hop; a single-replica drain 402s budget-capped agents. Production must set ≥2.
statelayerProxy.image.repository / .tag	statelayer-proxy / ""	Signed image in production.
componentPodDisruptionBudgets.enabled	false	PDBs for bff / gateway / token-service / statelayer-proxy. Enable alongside their replicas≥2 (each renders only at ≥2).
componentPodDisruptionBudgets.minAvailable	1

Sizing (container resources)

Every workload’s CPU and memory come from values. The defaults are what the chart has always shipped, so helm upgrade without setting any of these changes no limit on your cluster.

Value	Default	Meaning
bff.resources	500m/1Gi limits, 10m/192Mi requests	The console’s server-side layer.
controllerManager.resources	500m/128Mi, 10m/64Mi	The operator.
gateway.resources	2/1536Mi, 50m/512Mi	LiteLLM. The largest default in the chart — it holds provider clients and streams.
tokenService.resources	500m/128Mi, 10m/64Mi	Credential plane.
statelayerProxy.resources	500m/128Mi, 10m/64Mi	The memory/quota hop. Size this alongside statelayerProxy.replicas.
preflight.resources, capabilityKey.resources, tenantCleanup.resources	128Mi limit, 10m/64Mi	Install-time Jobs and hooks.
bff.runStore.resources	1Gi limit, 50m/64Mi	The durable run worker (only with bff.runStore.enabled).
devDataPlane.postgres.resources	2Gi limit, 100m/256Mi	The bundled dev/trial data plane. Production runs its own.
devDataPlane.statelayer.resources	256Mi limit, 50m/128Mi
devDataPlane.objectstore.resources	512Mi limit, 50m/128Mi
devDataPlane.nats.resources	512Mi limit, 50m/64Mi
statelayer.persistence.resources	1/256Mi, 50m/128Mi	The persistent state layer.

Set a map to null to remove the block entirely — what you want when a LimitRange in the namespace supplies them instead:

helm upgrade ctxmesh oci://ghcr.io/ctxmesh/charts/ctxmesh --version 0.1.0-beta.8 \
  --set bff.resources.limits.memory=2Gi \
  --set gateway.resources.limits.cpu=4

Note

statelayer.persistence.cpuLimit was replaced by statelayer.persistence.resources. Setting the old key now fails the render with a message naming the replacement, rather than silently ignoring a limit you set on purpose.

BFF / console

Value	Default	Meaning
ui.enabled	true	Serve the console SPA + /api. false = headless control plane.
bff.replicas	1	>1 requires bff.runStore.enabled (dispatch) — else in-process runs split across pods and are lost on a pod loss; the render fails on >1 without dispatch.
bff.image.repository / .tag	bff / ""	Signed image in production.
bff.runStore.enabled	false	Durable run store + HA run-worker: the BFF dispatches runs to a separate worker. Requires a Secret (dsnSecretName, key dsn) with the run-store Postgres DSN.
bff.runStore.dsnSecretName	run-store	The run-store DSN Secret name (create it with a managed-Postgres DSN).
bff.runExecTimeout	"" (10m)	How long one run advance may take — the agent’s whole managed loop for a turn. Raise it for long tool chains; empty uses the compiled default.
bff.runExecMaxTimeout	"" (60m)	The ceiling that kills a wedged run regardless of the above.
bff.runStore.worker.concurrency	4	Concurrent claim loops per worker pod.
bff.runStore.worker.minReplicaCount	1	KEDA min and the Deployment warm floor. 0 opts into scale-to-zero; production uses 2.
bff.runStore.worker.maxReplicaCount	10	KEDA max on queued-run backlog.
bff.costRollupEnabled	"1"	Cost-rollup worker (feeds /api/cost + chargeback). "0" disables.
bff.consoleURL	""	The canonical browser-reachable console origin (MCP-consent redirect_uri + cross-origin relay target). Set to your console edge host in production.
bff.providerConnect.enabled	true	Connect-a-provider flow. false (hardened) → the connect endpoints 404.
bff.generation.platformModels	""	Comma-separated governed generation models (empty = unpinned caller model).
bff.mcp.enabled	true	BYO-MCP register/discover. false (hardened) → the endpoints 404.
bff.mcp.requireApproval	false	true (hardened) → newly registered MCP tools are pending-approval.
bff.mcp.credentialNamespace	""	Set (e.g. ctxmesh-credentials) to render a locked platform namespace holding MCP grant Secrets so tenants can’t read each other’s OAuth tokens. Production should set this.

Console login (OIDC/SSO via bundled Dex)

Off by default (token login is the OSS default). Enable to bundle Dex.

Value	Default	Meaning
auth.oidc.enabled	false	Bundle Dex + the console OIDC client.
auth.oidc.issuer	example URL	The Dex issuer — must match the API server --oidc-issuer-url byte-for-byte.
auth.oidc.client.id / .redirectURIs	console defaults	The SPA OAuth client (Auth-Code + PKCE, public client).
auth.oidc.staticUser.*	dev demo user	Zero-config demo login — dev/trial only, change before exposing Dex.
auth.oidc.connectors	[]	Federated IdP connectors (enterprise; Google/Azure/Okta/GitHub/LDAP/SAML).
auth.oidc.langfuse.*	off	Register the trace UI as a second client of this Dex (shared login).
auth.oidc.demoBinding / .usernamePrefix	true / oidc:	Example ClusterRoleBinding for the demo user; prefix must match --oidc-username-prefix.

Note

The load-bearing OIDC step is a cluster-admin one: the API server must be told to trust the Dex issuer (--oidc-* flags). The chart ships Dex + the client but cannot set apiserver flags. On a managed cluster that can’t set them, keep token login — that is not a regression.

Data plane

Value	Default	Meaning
devDataPlane.enabled	true	Bundled in-cluster Valkey + MinIO with deterministic dev creds (dev/trial only). Production: false + BYO external.
statelayer.externalAddr	""	BYO external Valkey/managed Redis host:port for the state-layer backend. production requires this or persistence.enabled.
statelayer.persistence.enabled	false	Optional in-cluster persistent state layer (StatefulSet + PVC + AOF). Enable together with devDataPlane.enabled: false (a pod-selector collision fails loud otherwise). One-way migration — see Upgrade & versioning.
statelayer.persistence.size	8Gi	PVC size for AOF/RDB.
statelayer.persistence.storageClassName	"" (cluster default)	Prefer a network-attached class — a RWO local volume pins the pod to one node.
statelayer.persistence.cpuLimit	"1"	CPU limit on the Valkey container.

Security posture

The in-cluster Valkey is unauthenticated by design — a self-host convenience tier (isolation is ClusterIP + key-prefix scoping + the opt-in cross-tenant NetworkPolicy; no protocol-level per-tenant data isolation). requirepass is declined (the Knative constraint would make the password broadly readable → security theater). For sensitive multi-tenant production data, use BYO-external Valkey/managed Redis with real, operator-managed auth.

Images the controller injects into agent pods

Value	Default	Meaning
controllerManager.injectedImages.collector	ghcr.io/ctxmesh/agent-otel-collector	The OTel collector sidecar image, injected into every agent pod. Tagged with the chart’s appVersion unless you pin your own tag or digest. Empty ⇒ the controller’s dev.local default, which ImagePullBackOffs on a real cluster.
controllerManager.injectedImages.discovery	ghcr.io/ctxmesh/agent-discovery	The tool-discovery sidecar image, injected into every agent that has a tool binding. Same tagging and same dev.local caveat.
controllerManager.oboEgress.enabled	false	Gates OBO credential injection only — per-user tool calls. It does not gate sidecar injection and has not since M82: the egress sidecar is injected for every agent with ≥1 tool whatever this says.
controllerManager.oboEgress.sidecarImage	ghcr.io/ctxmesh/egress-sidecar	The egress sidecar image. Required by any install that uses tools, not only OBO ones, because the sidecar is the always-on tool-call chokepoint. Unlike the two above it has no fallback: an empty value makes Knative reject the agent’s Service with missing field(s): … containers[N].image.
controllerManager.oboEgress.capabilityAudience / .tokenServiceURL	""	Audience / token-service URL (the URL is derived by default). The capability public key comes from the bff-capability Secret, not here.

Install hooks (GA Gate A)

Value	Default	Meaning
capabilityKey.generate	true	A post-install/upgrade hook generates the platform Ed25519 keypair into bff-capability iff absent (never re-keys). Set false to BYO.
preflight.enabled	true	A post-install/upgrade Job fails LOUD on an incoherent config (empty required env, mismatched keypair, unreachable control-plane store).

Prometheus & NetworkPolicy (opt-in)

Value	Default	Meaning
prometheus.serviceMonitor.enabled	false	ServiceMonitor for the manager HTTPS metrics (:8443). Requires the Prometheus Operator CRDs.
prometheus.serviceMonitor.insecureSkipVerify	true	Set false + cert-manager TLS in production.
prometheus.podMonitor.queueProxy.enabled	false	Knative queue-proxy request metrics (feeds AlertPolicy SLOs). Also needs Knative export configured.
prometheus.podMonitor.queueProxy.port / .interval	http-usermetric / 15s
prometheus.runPipeline.enabled	false	Run-pipeline SLIs (BFF + run-worker metrics :9092) + a dead-worker-pool alert.
prometheus.runPipeline.interval / .deadPoolFor	30s / 5m
networkPolicy.metricsIngress.enabled	false	Restrict /metrics ingress to namespaces labeled metrics: enabled (needs a NetworkPolicy CNI).

CRDs

Value	Default	Meaning
crds.install	true	Install CRDs with the chart. false to manage CRD lifecycle out of band (a shared cluster where CRDs are pre-applied).

See also

• Production install
• High availability
• Upgrade & versioning
• Secrets · Credential stores
• Observability backends