All posts
Muhtalip Dede profile photoMuhtalip Dede · Founder of kprompt2 min read

The $0 magic moment: failed rollout → proposal → approve → Learn

Canonical kprompt try path on kind: one stalled Deployment, one PlanResult-shaped proposal, your explicit approve, verify, then Learn — no LLM key, no silent heal.

kprompt is The AI Runtime for Kubernetes: it understands your cluster context and turns intent into operate / investigate / recommend — with a reviewable PlanResult before anything mutates. The fastest way to feel that contract is not a chat session. It is one failed rollout on kind.

This is the canonical try path (IC-001): Observe correlates a stalled checkout Deployment, stores a PlanResult-shaped proposal, you review it, you approve, verify runs, Learn writes an outcome. Heuristic analysis — $0, no LLM provider key. Mutation never happens without your y/N (or an explicit --approve you typed).

One command

Prereqs: Docker, kind, kubectl, make, git, kprompt on PATH

kprompt demo                    # checklist + exact next commands
git clone https://github.com/kprompt/kprompt-examples.git
cd kprompt-examples && make magic-moment

Install the CLI first if needed: curl -fsSL https://kprompt.ai/install | bash — or brew install kprompt/tap/kprompt. The script creates (or reuses) a kind cluster, breaks only scenario 04-failed-rollout in namespace payments, runs the Observe agent in heuristic mode, prints the durable proposal, then asks [y/N]. Non-interactive shells stop before apply and print the exact apply command.

What you should see

  • Revision 2 stalls on a bad image; revision 1 keeps serving (maxUnavailable: 0)
  • A durable AutopilotProposal with action rollbackFailedRollout, risk, steps — Applied stays false until you approve
  • Narrow policyAuto allowlist (demo/policy-auto.json) — still requires your explicit approve
  • After apply: rollout status green + patterns list shows Learn writeback

Why this is not auto-heal

Propose → approve → verify → Learn. Observe does not silently patch production. Learn may bias the next similar proposal’s actionConfidence (evidence-not-proof); it does not skip Safety or RemediationPolicy. Role-scoped agent; no Secret values; Coordinator never mutates.

After the magic moment

  • Second hit (same fingerprint, Learn bias): make learn-second-hit
  • Natural-language plans ($0): kprompt init --ollama then kprompt "how's my cluster"
  • Brownfield cluster: Adopt (15 min) — bind tools, read first
  • Advanced noisy Observe fixture (not the stranger path): make walkthrough

Day-0 NL after the demo

kprompt init --ollama
kprompt "how's my cluster"
kprompt "scale api to 3"   # plan first — approve only when you trust it

Honest limits

  • Experimental — prefer kind / non-production first
  • Heuristic demo is deterministic offline analysis, not a full BYOK LLM run
  • Do not lead strangers at make walkthrough; that path is advanced / multi-break
  • Free CLI forever; Team is optional org policy/audit when you enroll

Same story everywhere: install → kprompt demo → make magic-moment. Category stays AI Runtime — plan-gated, not chat-shaped kubectl.

Related posts