mecak8s

Overview ¶

Command mecak8s is the storage-free, Kubernetes-native mecatl agent binary: a THIN peer of cmd/mecated that composes the SAME app.Build assembly with k8s-native defaults — Redis session store + durable event log, coordination.k8s.io Lease session leasing, a dynamic /readyz (drain-gated + Redis-pinged), and a bounded GracefulStop that cancels in-flight runs on SIGTERM so a rolling update completes within terminationGracePeriodSeconds.

What it does NOT have (stripped from mecated): no `skills promote` / `config` / `perf-mcp` subcommands, no ACP (mecated-only), and NO --store-dir (storage-free: state lives in Redis and the k8s API server). It shares the SAME provider + model-alias/model-slot flag wiring (internal/cliconfig) so the three-mains wiring cannot drift. Telemetry (issue #343) is OPT-IN: a loopback --metrics-addr mounts the admin mux's /metrics for scrape, and --otlp-* pushes traces/metrics to a collector. Both default off — the pre-telemetry posture.

Honest shutdown contract: new runs are rejected (503 via the drain gate) the moment SIGTERM (or the preStop httpGet /drain) fires. In-flight runs are CANCELLED, not drained to completion — a multi-minute LLM turn cannot survive a rolling update within terminationGracePeriodSeconds: 60. The pod is disposable; the session is not — it is Recover-able on the successor (issue #51) from the Redis snapshot + durable event log.

Command mecak8s is the storage-free, Kubernetes-native mecatl agent binary. See flags.go for the configuration surface and serve.go for the shutdown contract. This file is the thin entry point: parse flags → build the diagnostics sink → app.Build → serve → os.Exit.