Omh Iac Change

Iac Change

This is an OMH iac-change workflow skill, projected for Agent Skills hosts (Claude Code, Codex, Cursor, opencode, OpenClaw, pi).

Why This Exists

iac-change exists because infrastructure changes had no owner: deploy-and-monitor watches an application release, release-cut decides one, and inference-serving deploys a model server, while a Terraform plan with drift, a Helm chart change, or a manifest edit reached a generic planner with no blast radius, cost delta, or staged apply at all.

First Steps

• Ask for the saved plan or the diff output before assessing anything.
• Separate existing drift from the proposed change before ordering stages.

Do Not Use When

• The ask is shipping a new version of the application and watching its health signals after the deploy; use omh-deploy-and-monitor.
• The ask is deciding a versioned release -- what goes in, its tag, or its canary -- rather than changing declared infrastructure; use omh-release-cut.
• The ask is choosing and deploying a model-serving engine; use omh-inference-serving.
• Production is down or degraded right now and the ask is command of the incident; use omh-live-incident-response.

Examples

Good example:

• Prompt: terraform plan shows drift in the kubernetes cluster, stage the apply
• Expected behavior: Separate the drift from a refresh-only plan, read the saved plan's creates, replacements and destroys, give the cost delta or mark it unestimated, and stage the apply from staging to production with a rollout-status gate and a rollback per stage.
• Why: Applying over unresolved drift changes resources nobody meant to touch.

Bad example:

• Prompt: just run terraform apply -auto-approve in prod, the plan looked fine yesterday
• Expected behavior: Refuse the fresh apply: re-plan, save the plan, read its replacements and destroys, and apply that saved plan behind a health gate.
• Why: Yesterday's plan is not today's apply; drift and other merges change what a fresh apply does.

Completion Checklist

• Existing drift is separated from the proposed change.
• Every replacement or destroy is listed, and each stateful one waits for explicit approval.
• The cost delta is observed or marked unestimated.
• Every stage names its health gate and its rollback.
• OMH ran nothing, and every fact cites observed output or is marked unverified.

Recovery Notes

• If no saved plan exists, stop at the plan step and ask for one; do not assess from the diff of the code alone.
• If a resource cannot be rolled back, say so in its stage and ask for approval before that stage.

When to Use

Use when a change to declared infrastructure is about to be applied or has drifted: a Terraform, OpenTofu, Pulumi or CloudFormation plan, a Kubernetes manifest or kustomization, or a Helm chart or values change. The output is the drift between state and reality, the blast radius and cost delta read from the saved plan, and a staged apply where every stage has a health gate and a rollback; OMH applies nothing.

Copy code
Strong routing signals: `iac-change`, `iac change`, `infrastructure as code`, `infrastructure-as-code`, `terraform plan`, `terraform apply`, `terraform state`, `terraform drift`, `terraform module`, `terraform change`, `opentofu`, `tofu plan`, `pulumi`, `pulumi preview`, `pulumi up`, `cloudformation`, `cloudformation change set`, `helm upgrade`, `helm diff`, `helm chart`, `helm chart change`, `kubectl apply`, `kubectl diff`, `kustomize`, `kubernetes manifest`, `kubernetes manifests`, `k8s manifest`, `k8s manifests`, `infracost`, `drift detection`, `cost delta`, `staged apply`, `stage the apply`

Catalog Metadata

Category: planning Phase: iac-change Quality tier: staged-apply-gated Reasoning demand: standard

Quality bar:

• Read the saved plan or diff before anything else; the blast radius is what the plan says, not what the change was meant to do.
• Load references/iac-change-method.md for the per-tool drift, plan, health and rollback commands and the replacement markers instead of recalling them.
• Resolve existing drift before the apply, so the apply changes only what the change meant to.
• Give every stage a health gate and a rollback; a stage with neither is not a stage.
• Keep prepared, applied, and verified as separate states for every stage.

Required inputs:

• the tool and the unit of change: the Terraform workspace or stack, the cluster and namespace, or the Helm release
• the environments the change reaches, in promotion order
• the observed plan or diff output for the change, saved to a file where the tool allows it
• the health signal each environment already exposes: rollout status, probes, error rate, or a smoke check
• observed output for any drift, apply, health, or rollback claim

Expected outputs:

• drift_assessment/v1
• blast_radius/v1
• cost_delta/v1
• staged_apply_plan/v1
• health_gate/v1
• rollback_plan/v1

Artifact expectations:

• drift_assessment/v1 separates drift already present between state and the running infrastructure from the change being proposed, from a refresh-only plan or a live diff, and names who resolves each drifted resource before the apply
• blast_radius/v1 lists every resource the saved plan creates, updates in place, replaces, or destroys, calls out each replacement or destroy of a stateful resource, and names what depends on it
• cost_delta/v1 gives the monthly delta from an observed estimate for the saved plan, or marks the delta unestimated and names the resources that drive it
• staged_apply_plan/v1 orders the stages from the least to the most exposed environment and applies the reviewed saved plan, never a fresh one
• health_gate/v1 names, per stage, the observed signal that promotes it -- rollout status, a follow-up plan with no changes, an error rate -- and the value that stops it
• rollback_plan/v1 names, per stage, the command or revert that undoes it and what cannot be undone, such as a destroyed volume or a rotated resource id

Safety rules:

• Never promote a stage without its health gate observed; a prepared gate keeps the next stage blocked.
• A replacement or destroy of a stateful resource -- a database, a volume, a bucket, a load balancer address -- is a blocker for the user's explicit approval, not a line in the plan.
• Apply only the saved plan that was reviewed; a fresh apply can differ from what was read.
• Never hand-edit a state file or a live object to hide drift; import, move, or re-declare it, and say which.
• OMH never runs terraform, tofu, pulumi, kubectl, or helm, and never reads cloud credentials; every drift, cost, apply, and health fact comes from observed output or is marked unverified.

Runtime Evidence

Use the current host's own tools and subagent/task mechanism when available; otherwise run the same lanes sequentially or name the unavailable capability. A prepared plan, handoff, checklist, or skill installation is not execution, review, CI, merge-readiness, or merge evidence. Record actual tool results, or not_observed / not_available, in the record; never invent dispatch or host accounting. Treat supplied context as advisory, not proof of hidden memory reads or writes. State scope, constraints, verification, and the stop condition before work. Reply in the user's own words and the host's own voice: the host's persona owns reply language, tone, speech level, and sentence endings, progress updates included (where it sets no language, use the one the user wrote in), and OMH shapes structure and content only; OMH's record terms (surface, lane, wrapper, handoff, evidence boundary, not_observed) stay in records and tool calls, never in the sentence the user reads unless they ask about one; and when a real unresolved user decision or a useful briefing discussion remains, offer one specific, content-linked question. Completed work needs no question; omit unrelated non-action footers and generic offers to continue. Supporting paths are relative to this skill directory; sibling skill paths are relative to its parent. Resolve them from the host-provided skill base directory ({baseDir} on hosts that provide it), never a hardcoded install location. A named workflow not installed here is unavailable, not permission to emulate its host-specific capabilities. Verify through the real surface before done.