metrics

var (
	PoolReplicasDesired  *prometheus.GaugeVec
	PoolReplicasIdle     *prometheus.GaugeVec
	PoolReplicasRunning  *prometheus.GaugeVec
	PoolReplicasStarting *prometheus.GaugeVec
	PoolReplicasStopping *prometheus.GaugeVec
	PoolReplicasFailed   *prometheus.GaugeVec
)

var (
	// SandboxClaimDuration observes how long ClaimIdlePod takes.
	// outcome: "success" | "no_idle" | "timeout" | "error"
	SandboxClaimDuration *prometheus.HistogramVec

	// SandboxStartingDuration observes the image-pull / startup time (claimedAt → startedAt).
	// stop_reason label is absent here; use for P99 startup latency breakdowns.
	SandboxStartingDuration *prometheus.HistogramVec

	// SandboxRunningDuration observes actual sandbox running time (startedAt → terminatedAt).
	// stop_reason: "Completed" | "Failed" | "Canceled" | "Evicted"
	SandboxRunningDuration *prometheus.HistogramVec

	// SandboxRecycleDuration observes the Stopping→Idle recycle time (terminatedAt → recycledAt).
	SandboxRecycleDuration *prometheus.HistogramVec

	// SandboxArmDuration observes how long arming a claimed sandbox takes:
	// waiting for its runtimes to answer, running the post-start hooks, and
	// pushing the egress policy and credentials. Together with
	// SandboxClaimDuration it accounts for the whole create latency a caller
	// observes, which is why it is a separate series rather than folded into
	// the claim histogram.
	SandboxArmDuration *prometheus.HistogramVec
)

var (
	// E2BUnsupportedTotal counts requests to E2B operations AgentBox does not
	// serve, by operation and refusal category. It exists to answer which parts
	// of the E2B surface callers actually reach for, so the next batch of work
	// is chosen from evidence rather than guesswork.
	// category: "architectural" | "platform" | "unimplemented"
	E2BUnsupportedTotal *prometheus.CounterVec

	// VaultOperationTotal counts credential-vault operations by outcome. It
	// carries the namespace but never the user or the entry name: the counter
	// is for usage and failure rates, and a per-entry label would turn the
	// metrics endpoint into a directory of which credentials exist.
	// op: "list" | "get" | "create" | "update" | "delete"
	VaultOperationTotal *prometheus.CounterVec
)

var (
	// SandboxCreateTotal counts sandbox creation attempts.
	// result: "success" | "no_idle" | "timeout" | "error"
	SandboxCreateTotal *prometheus.CounterVec

	// SandboxDeleteTotal counts sandbox deletions.
	// stop_reason: "Completed" | "Canceled" | "Failed"
	SandboxDeleteTotal *prometheus.CounterVec

	// InplaceUpdateTotal counts TriggerUpdateWithOptions calls.
	// result: "success" | "conflict" | "error"
	// (conflict covers both k8s resource version conflicts and phase mismatches)
	// target: TargetPodPhase value (e.g. "running", "idle")
	InplaceUpdateTotal *prometheus.CounterVec
)

var (
	HTTPRequestsTotal   *prometheus.CounterVec
	HTTPRequestDuration *prometheus.HistogramVec
)

var (
	// ScheduleReadyQSize is the current number of pods in the per-pool ready queue
	// (idle pods known to the scheduler, not yet dispatched).
	ScheduleReadyQSize *prometheus.GaugeVec

	// ScheduleReservationsSize is the current number of per-pool inflight reservations
	// (pods either being CAS'd or recently claimed within the TTL window).
	ScheduleReservationsSize *prometheus.GaugeVec

	// ScheduleCASOutcomeTotal counts TriggerUpdateWithOptions outcomes from the scheduler.
	// outcome: "success" | "retriable" (phase mismatch / k8s conflict) | "hard" (other errors).
	ScheduleCASOutcomeTotal *prometheus.CounterVec

	// ScheduleDispatchLatencySeconds measures the time from request enqueue to the
	// moment the CAS goroutine starts executing TriggerUpdateWithOptions.
	ScheduleDispatchLatencySeconds *prometheus.HistogramVec

	// ScheduleRefreshTotal counts ready-queue refresh attempts. outcome: "ok" | "throttled" | "error".
	ScheduleRefreshTotal *prometheus.CounterVec

	// ScheduleReservationTTLExpiredTotal counts reservations removed by TTL sweep
	// (i.e. reservations not explicitly released by the CAS outcome handler).
	ScheduleReservationTTLExpiredTotal *prometheus.CounterVec

	// ScheduleSkippedScaleDownProtectedTotal counts refreshes where pods were skipped
	// because they carried the scale-down-protected annotation.
	ScheduleSkippedScaleDownProtectedTotal *prometheus.CounterVec

	// ScheduleReadyQueueEvictedTotal counts pods discarded from the ready queue at
	// dispatch time because they were no longer present in the informer cache or had
	// transitioned out of Idle (e.g. deleted during scale-down).
	ScheduleReadyQueueEvictedTotal *prometheus.CounterVec

	// ImageRewriteTotal counts per-cluster container-image registry rewrites,
	// labelled by outcome. The "no_local_registry" result is the one worth
	// alerting on: the image belongs to a peer cluster's registry and this
	// cluster declares no counterpart, so the image is left pointing at another
	// region and surfaces minutes later as ImagePullBackOff.
	ImageRewriteTotal *prometheus.CounterVec

	// EnvVolumeMounts reports how many PersistentVolumeClaim mounts each Env
	// declares. A gauge rather than a counter: it tracks a declared shape, not
	// an event rate.
	EnvVolumeMounts *prometheus.GaugeVec

	// EnvVolumeReadOnlyUnenforceableTotal counts renders that accepted a
	// read-only volume mount on a pod spec that cannot enforce it, i.e. where a
	// template carried the admin opt-out annotation. Every increment is a
	// knowingly downgraded guarantee, so it must never be silent.
	EnvVolumeReadOnlyUnenforceableTotal *prometheus.CounterVec

	// EgressMissingSidecarTotal counts idle pods the dispatcher refused because
	// the claim requires egress enforcement but the pod carries no filter
	// sidecar (it was materialised before the Pool gained its networkPolicy).
	// Any non-zero value means a Pool rollout has not finished; a sustained
	// non-zero rate means claims are being starved and should alert.
	EgressMissingSidecarTotal *prometheus.CounterVec
)

var (
	// SandboxArmTotal counts arming attempts by outcome.
	// result: "success" | "timeout" | "error"
	SandboxArmTotal *prometheus.CounterVec
)

var (
	// SandboxRunningInfo is an info gauge (value always 1) that maps running sandbox IDs
	// to their pod names. Present only while the sandbox is in Running state.
	// Labels: namespace, pool, pod, sandbox_id, team, user.
	// Use for PromQL joins with kube CPU/memory metrics via namespace+pod labels.
	SandboxRunningInfo *prometheus.GaugeVec
)