mk:devops

What This Skill Does

Plans and implements infrastructure, container, pipeline, runtime-configuration, and observability changes, and designs the safety around them: blast radius, stateful resources, rollback, approval point, and the read-only evidence that proves the change is right.

It never performs the production effect and never reads a secret. Both are structural, not stylistic.

When to Use

• Container definitions, local runtime composition, or pipeline changes
• Infrastructure-as-code and cloud runtime changes
• Deployment safety design, rollback planning, runtime configuration
• Operational diagnosis of a live failure
• NOT for: approving or executing a release (mk:ship or a human), reading secrets, code root-cause analysis (mk:investigate), schema semantics (mk:database), or the security verdict

Secret Boundary

Unconditional, stated in the skill body rather than a reference so it is always loaded:

• Never read, grep, print, or open a .env/.env.* file at any depth, a keystore, a credential file, or a generated config carrying credentials
• Never ask for a secret value in conversation, and never echo output containing one
• Read only committed templates, configuration, and infrastructure definitions
• When availability cannot be proven from a tracked file, stop at a readiness report naming the missing capability — absence from the tree proves nothing about the runtime

Ownership Boundary

Arguments

[platform or infrastructure change]

Workflow

1. Discover the platform from tracked container, IaC, pipeline, and deployment files — when it is not discoverable, ask; never pick a cloud as a default
2. Classify — local container, CI, infrastructure, runtime config, observability, release/rollback, or incident diagnosis
3. Safety design — blast radius, stateful resources, environment contract (names only), rollback or forward recovery, approval point, validation command
4. Implement inside an approved plan — produce and read the plan or diff evidence; never run the apply, deploy, or destroy step and never emit one ready to run
5. Hand off the production effect with the evidence, blast radius, rollback path, and post-change validation

References

Pro Tips

• A plan or diff nobody read is not evidence. Quote the resources it would replace, not just update.
• "Replace" on a stateful resource is data loss until a restore is proven, not merely configured.
• A rollback that needs the failing system healthy is not a rollback. Check the recovery path against the failure it is meant to survive.