optimize my cluster: idle workloads, rightsizing, and HPA hints — without auto-apply

“Optimize my cluster” is one of those prompts every platform team wants to type — and every vendor wants to auto-remediate. kprompt ships a different shape: a read-only report (inventory, idle signals, rightsizing deltas, HPA hints), then optional suggested fixes that still go through plan → approve. Passing --approve on the optimize prompt does not silently patch your Deployments.

What the report covers (shipped)

• Inventory — workloads with replicas and request/limit sketches
• Idle / underutilized — when Prometheus usage vs requests supports it
• Rightsizing — CPU/memory request/limit deltas from usage percentiles
• HPA hints — present, maxed, or static-replica narration
• Suggestions — human-readable follow-ups with action hints

Without Prometheus configured, inventory and structural HPA notes still help; idle and rightsizing degrade honestly instead of inventing savings numbers.

Run it

brew install kprompt/tap/kprompt
export KPROMPT_GEMINI_API_KEY="..."

# Optional but recommended for idle / rightsizing
kprompt config set tools.prometheus.url http://prometheus.monitoring:9090

kprompt "optimize my cluster"
kprompt "optimize my cluster" -n production
kprompt "optimize my cluster" -o json   # CI / jq

Optimize: cluster (1h)
Summary: 42 workloads scanned; 3 idle; 5 rightsizing candidates

Findings:
  - [medium] Idle replicas: payment-worker under 5% CPU vs requests
  - [low] HPA maxed: checkout at maxReplicas

Idle:
  - Deployment/payment-worker: low CPU vs requests over window

Rightsizing:
  - Deployment/api: memory request 512Mi → suggest 384Mi

HPA:
  - Deployment/checkout: HPA present, currently at max

Suggestions:
  - Scale down idle worker: review replicas (optional approved scale)
  - Lower memory request on api: review patch (optional approved patch)

Optional fixes still need their own approval

Top findings can become a separate scale or patch plan. That plan is risk-evaluated and needs TTY y/N or an explicit --approve on that follow-up — not the parent optimize flag. This is intentional: optimize is a report, not a bot that rightsizes production while you get coffee.

kprompt "optimize my cluster" --approve
# Still read-only report — does not auto-apply suggestions

# Later, if you agree with a suggestion:
kprompt "scale payment-worker to 1" -n jobs
# → Plan → Apply? [y/N]

JSON for pipelines

With --output json, the PlanResult carries the optimize report (idle, rightsizing, HPA, findings). Use it to fail a nightly job when high-severity idle findings appear — or to open a ticket — without applying anything.

What it is not

• Not a FinOps bill exporter (dollar / carbon notes are a separate backlog item)
• Not a guarantee of “safe” rightsizing without your review
• Not a replacement for Kubecost, OpenCost, or capacity planning reviews
• Not multi-cluster fleet rollup yet (local report per context; fleet is later)
• Not advice to --approve optimize in production and walk away

How it fits the AI SRE path

Optimize is cluster-level thinking under the intent-compiler contract: evidence in, structured report out, mutate only with a second gate. Pair it with the service dependency graph when you need consumers before you shrink a Deployment, and with the AI SRE roadmap when you want investigate / blast-radius next.

Try on staging first

kprompt tools
kprompt "optimize my cluster" -n staging
kprompt "show service dependency graph" -n staging

kprompt remains experimental. Prefer non-production while you learn how findings look for your charted workloads and Prom labels. Star issues if a rightsizing heuristic is wrong for your stack — the report should stay honest, not optimistic. For the JSON shape of reports and plans, see the PlanResult JSON deep dive.