Feature: infra-aware adversarial review prompts + auto-scaling by diff size

Problem

/codex:adversarial-review uses a single generic security-focused prompt regardless of what's being reviewed. Infrastructure-as-Code (Terraform, Helm, K8s manifests, CI/CD pipelines) has fundamentally different failure modes than application code -- state corruption, blast radius, provider drift, and dependency ordering matter more than OWASP Top 10.

Additionally, the review depth is static. A 10-line config change gets the same treatment as a 300-line module rewrite, wasting tokens on small changes and potentially under-analyzing large ones.

Proposal

1. Domain-specific prompt templates

Auto-detect file types in the diff and select the appropriate adversarial template:

• Infrastructure-first (.tf, .hcl, values.yaml, Chart.yaml, .github/workflows/): Focus on STATE CORRUPTION, BLAST RADIUS, IAM/PERMISSIONS, PROVIDER DRIFT, DEPENDENCY ORDERING, ROLLBACK, RACE CONDITIONS, COST
• Application-focused (.ts, .py, .go, etc.): Focus on EDGE CASES, SECURITY, CONCURRENCY, FAILURE MODES, DATA INTEGRITY, ROLLBACK
• Mixed diffs: Combine both templates

2. Auto-scaling depth by diff size

Why this matters

IaC reviews have different stakes. A missed race condition in Terraform can corrupt state for an entire environment. A permissive IAM policy grants access across accounts. Generic security prompts miss these domain-specific risks entirely.

Start by locating the implementation of /codex:adversarial-review and how it reads the diff. Trace where prompts are selected and diff size is calculated, then define coverage for infrastructure, application, mixed, and each size tier. Done means the review selects the stated focus areas and scales depth according to the table.