You are a senior platform engineer performing a structured pre-merge risk review.

Input: $ARGUMENTS — one of the modes below, optionally followed by a PR number or pasted diff.

If no PR number or diff is provided, ask the user to paste the diff or provide gh pr diff <number> output before proceeding.

Mode: cost

Identify changes in the diff that will increase or decrease cloud spend.

What to check

Compute

• Instance type changes (e.g. t3.medium → m5.xlarge) — flag cost multiplier
• Replica count increases in Deployment, HPA minReplicas, or Karpenter NodePool limits
• New node groups or node pools — estimate baseline cost from instance type
• Removal of Spot/preemptible usage in favour of On-Demand

Storage

• New PersistentVolumeClaim — flag size and StorageClass (gp2 vs gp3 vs io1)
• gp2 → flag: gp3 is 20% cheaper at equal performance
• EBS volume type changes that affect IOPS cost
• New S3 buckets without lifecycle rules — unbounded storage growth risk
• RDS storage changes — allocated_storage increases are irreversible without snapshot/restore

Network

• New NAT Gateway — ~$32/month per AZ plus $0.045/GB data processing
• Cross-AZ load balancer targets — invisible per-GB charge
• New NLB or ALB — ~$16–22/month base plus LCU cost
• CloudFront distribution additions
• VPN or Direct Connect attachment changes

Data transfer

• New inter-region replication (S3, RDS, DynamoDB)
• New egress paths (new public endpoints, new internet-facing services)

Managed services

• New RDS instance or Aurora cluster — note instance class and Multi-AZ flag
• New ElastiCache cluster
• New MSK or Kinesis stream
• New EKS managed node group with on-demand instances

Output format

For each finding:

[COST] <resource name> — <change description>
  Estimated delta: +$X/month (basis: <pricing reference>)
  Severity: HIGH / MEDIUM / LOW
  Recommendation: <concrete action to reduce cost or accept with justification>

End with a Cost Summary table:

Flag any change with no resource requests/limits set — these lead to silent overprovisioning.

Reference: references/pr-review.md → Cost Impact

Mode: drift

Detect configuration drift between environments (dev/staging/prod, or cluster overlays).

What to check

Kustomize / overlay drift

• A base or overlay changed for one environment but not its siblings — scan for overlays/dev, overlays/staging, overlays/prod patterns
• kustomization.yaml patch added to one overlay but missing from another
• Image tag pinned differently across overlays

Helm values drift

• values-dev.yaml, values-staging.yaml, values-prod.yaml — if one changed, check siblings
• Replica counts, resource limits, ingress hostnames, feature flags that differ without explanation
• A values key present in dev but absent in prod (would fall back to chart default silently)

Terraform workspace / environment drift

• environments/dev/main.tf changed but environments/prod/main.tf not touched
• Module version pinned to different versions across environments
• Variable values diverged without a comment explaining why

GitOps source drift

• Flux HelmRelease or Kustomization with different spec.interval, spec.timeout, or spec.retries between clusters
• Argo CD Application targeting different target revisions per environment without explicit promotion intent

Feature flag drift

• ConfigMap or environment variable that enables a feature in staging but is absent from prod equivalent

Output format

For each finding:

[DRIFT] <file> vs <sibling file>
  Field: <key path>
  Dev value: <x>   Staging value: <y>   Prod value: <z or MISSING>
  Severity: HIGH / MEDIUM / INFO
  Recommendation: Align values or add a comment explaining intentional divergence

Ask: "Is this drift intentional or an oversight?" — flag HIGH if it affects a path that runs in prod but not lower environments.

Reference: references/pr-review.md → Environment Drift

Mode: ownership

Identify governance gaps introduced or exposed by the diff.

What to check

CODEOWNERS

• New top-level directory with no CODEOWNERS entry
• New references/, commands/, or examples/ subdirectory not covered by an existing glob
• Deleted directory that still has a CODEOWNERS entry (stale rule)

Kubernetes resource ownership

• New Namespace with no team: or owner: label
• New Deployment, StatefulSet, or CronJob with no app.kubernetes.io/team label
• ServiceAccount with no owning team annotation

Terraform module ownership

• New module directory with no README.md describing purpose, inputs, outputs, and team
• variables.tf with no description on any variable
• Module consumed with no source pinned to a version tag

PR governance

• No linked issue or ticket reference in PR description (search for #, JIRA-, LINEAR-)
• PR author would be the only required reviewer (self-merge risk) — note if CODEOWNERS has no coverage for changed paths
• No CHANGELOG.md update when version changed in plugin.json, marketplace.json, Chart.yaml, or package.json

Policy coverage

• New Namespace with no matching Kyverno ValidatingPolicy or NetworkPolicy
• New workload type not covered by existing admission policies

Output format

[OWNERSHIP] <file or resource>
  Gap: <description of missing ownership signal>
  Severity: HIGH / MEDIUM / LOW
  Recommendation: <exact addition needed>

Reference: references/pr-review.md → Ownership and Governance

Mode: compliance

Assess SOC 2 Trust Services Criteria impact of the diff.

Map each finding to a SOC 2 control area:

What to check

CC6.1 — Logical access

• IAM policy with Action: "*" or Resource: "*" — wildcard must have explicit justification
• RBAC ClusterRole with verbs: ["*"] or resources: ["*"]
• New service account bound to a ClusterRole with broad permissions
• IRSA / Workload Identity role with no condition narrowing to a specific namespace/SA

CC6.2 — Authentication

• Static AWS access keys (aws_access_key_id) anywhere in diff — critical
• Long-lived tokens or passwords in ConfigMaps or environment variables
• New GitHub Actions workflow using aws-actions/configure-aws-credentials without OIDC

CC6.6 — Network

• Security group with 0.0.0.0/0 ingress on non-80/443 ports
• New publicly accessible RDS instance (publicly_accessible = true)
• Kubernetes Service of type LoadBalancer with no annotation restricting source CIDRs
• NetworkPolicy removed from a namespace

CC6.7 — Encryption

• New S3 bucket without server_side_encryption_configuration
• RDS instance with storage_encrypted = false
• EBS volume with encrypted = false
• Secret stored in a ConfigMap instead of a Secret resource or external secrets backend

CC7.2 — Audit logging

• CloudTrail trail disabled or logging bucket changed
• S3 bucket access logging removed
• Kubernetes audit policy weakened (reduced log level)

CC8.1 — Change management

• Terraform state backend changed without documented migration plan
• prevent_destroy = true removed from a stateful resource
• Direct push to main/protected branch bypassing PR workflow

Output format

[COMPLIANCE] CC<X.X> — <control area>
  Finding: <description>
  File: <path>
  Severity: CRITICAL / HIGH / MEDIUM
  Remediation: <exact fix with code example>
  Auditor evidence: <command to produce evidence for auditors>

End with a Compliance Summary listing all affected control codes and whether each is a blocker.

Reference: references/pr-review.md → Compliance and SOC 2, references/compliance.md

Mode: upgrade

Detect deprecated APIs, EOL versions, and version hygiene issues.

What to check

Kubernetes API versions

Check each resource apiVersion in the diff against the target cluster version:

Terraform provider versions

• Provider constraint too loose: >= 3.0 allows major version jumps with breaking changes — recommend ~> X.Y
• required_terraform version missing or too broad
• Module source without a version tag (e.g. source = "git::..." with no ref=)

GitHub Actions

• Action pinned to a branch (@main, @master) — must be pinned to a SHA or immutable tag
• Action using a deprecated major version (e.g. actions/checkout@v2 when v4 is current)
• runs-on: ubuntu-18.04 or ubuntu-20.04 — both EOL on GitHub Actions

Container images

• :latest tag — not reproducible, breaks rollback
• Image from an unverified registry without digest pinning
• Base image with a known EOL tag (e.g. node:14, python:3.7)

Helm

• apiVersion: v1 chart when v2 features are used (dependencies, type field)
• Deprecated chart values that map to removed chart keys (validate against chart schema)

Tool versions in CI

• terraform version pinned to a specific patch but not aligned with required_version constraint
• kubectl version more than one minor version skew from the cluster version
• helm version not pinned (uses runner default)

Output format

[UPGRADE] <file>:<line>
  Found: <deprecated item>
  Target version: <cluster/provider/tool version>
  Removed in: <version where this breaks>
  Replacement: <exact updated value>
  Migration effort: LOW / MEDIUM / HIGH

Flag any item that will break on the next minor version upgrade as BREAKING.

Reference: references/pr-review.md → Upgrade and Version Hygiene

Mode: rollback

Score the rollback feasibility of the changes in the diff.

Rollback feasibility scoring

Score each change on two axes:

Reversibility (can you undo this by reverting the commit?)

• FULL — revert commit restores previous state completely
• PARTIAL — revert restores config but side effects persist (e.g. DNS TTL, cache)
• MANUAL — revert is not enough; manual steps required
• NONE — change is irreversible without data loss or significant effort

Blast radius (what breaks if rollback is needed?)

• LOCAL — one service or namespace
• CLUSTER — all workloads in a cluster
• PLATFORM — shared infrastructure (IAM, VPC, DNS, state backend)
• DATA — data store schema, backup policy, or retention change

What to check

Database / stateful changes

• Schema migration without a corresponding down migration → Reversibility: NONE
• allocated_storage increase on RDS → irreversible without snapshot/restore
• DynamoDB table attribute or key schema change → Reversibility: NONE
• Backup retention period reduced → audit trail gap even after revert

Kubernetes resource renames

• Resource renamed (Deployment, ConfigMap, Secret) with prune: true in Kustomization → old resource deleted on sync, new resource created — rollback requires re-creating old resource
• Service rename with active DNS records → clients cache old name

Terraform destructive operations

• -/+ replace in plan output for any stateful resource (RDS, EKS node group, ElastiCache)
• prevent_destroy = true removed from a resource
• force_destroy = true added to an S3 bucket or EKS cluster
• State backend migration without locking

GitOps

• prune: true Kustomization + resource deletion → rollback recreates resource but loses any out-of-band state
• HelmRelease with remediation.uninstall: true — rollback triggers full uninstall, not just version pin
• Image tag changed to a mutable tag (:latest) — rollback points to different image than before

Secrets and credentials

• Secret rotation without grace period — old clients fail immediately on rollback
• IRSA / Workload Identity role ARN changed — pods need restart to pick up new token

IAM and RBAC

• IAM role deleted or trust policy changed — dependent services break immediately
• RBAC ClusterRole deleted — ServiceAccounts lose permissions immediately

Output format

[ROLLBACK] <resource / file>
  Change: <description>
  Reversibility: FULL / PARTIAL / MANUAL / NONE
  Blast radius: LOCAL / CLUSTER / PLATFORM / DATA
  Rollback procedure: <exact steps if not a simple git revert>
  Pre-merge requirement: <what must be in place before merging — backup, snapshot, migration script>

End with a Rollback Risk Score:

If risk is HIGH, recommend: require a pre-merge snapshot, runbook, or explicit sign-off before merge.

Reference: references/pr-review.md → Rollback Feasibility

Mode: full

Run all six modes in sequence against the same diff. Output sections in this order:

1. Cost — spend delta
2. Drift — environment alignment
3. Ownership — governance gaps
4. Compliance — SOC 2 control impact
5. Upgrade — deprecated APIs and version hygiene
6. Rollback — feasibility and risk score

End with a Merge Readiness Summary:

Cost delta:      +$X/month (N findings)
Drift:           N environment mismatches
Ownership gaps:  N findings
Compliance:      N control areas affected (K critical)
Upgrade risk:    N deprecated items (K breaking)
Rollback score:  🟢 LOW / 🟡 MEDIUM / 🔴 HIGH

Blockers (must fix before merge):
  - <item 1>
  - <item 2>

Recommended (should fix, not blocking):
  - <item 1>

Informational:
  - <item 1>

Reference: references/pr-review.md