Subagent Review

• ⛔ Do NOT invoke ANY skills — no {"skill": "azure-validate"}, {"skill": "azure-deploy"}, {"skill": "azure-prepare"}, or any other skill call. Use the procedures in THIS file only.
• ⛔ Do NOT run az deployment commands — review is read-only analysis of generated files.
• ⛔ Do NOT modify IaC files — report findings only. The caller fixes issues.

Input (provided by caller)

Output

Return JSON (≤1000 tokens):

Workflow

Step 1 — Read security patterns + run L1 security baseline

Read bicep-patterns-security.md and rbac-roles.md.

Do: Check every generated IaC file against ALL security checks defined in the reference file. The file contains the complete check table with FLAGGED conditions, edge cases, and Bicep code patterns. Do NOT guess checks from memory — use the reference file as the checklist.

Step 2 — Read checklist + run L2–L4 adversarial review

Read self-review-checklist.md.

Do: First run the cross-module reference trace from the checklist’s § Cross-Module Reference Validation: parse every module call in main.bicep, read each target module’s param/output declarations and secrets[] entries, then verify every reference resolves (params passed match params declared, outputs referenced exist, every CA secretRef has a matching KV secret resource). Then run L2–L4:

• L2 (Pattern Validation): File structure matches main.bicep → modules/*.bicep, naming follows plan, Container Apps uses two-phase wiring, every files[] entry exists on disk, no azure.yaml, cross-module references all resolve
• L3 (Hallucination Detection): Resource names match naming.resources[] exactly, API versions are real (verify via az bicep build), SKU names match plan, no invented resource types
• L4 (WAF Alignment): Check per-pillar:
  • Reliability: zone redundancy (prod SKUs), health probes, GRS storage, min replicas ≥1
  • Security: managed identity, KV secrets, HTTPS+TLS 1.2, no public blob, no administratorLogin
  • Cost: SKU matches budget, scale-to-zero for dev/test CA, free grants applied
  • Ops: App Insights, 5 AppOnboard tags, all values parameterized
  • Performance: autoscale (prod), CDN for SPA, connection pooling, cache tier

Step 3 — Compile findings + return

Do: Merge L1–L4 results into the findings JSON. Apply rating per self-review-checklist.md § Rating System: VERIFIED (evidence confirms claim), PLAUSIBLE (no counter-evidence but unverified), FLAGGED (evidence contradicts or missing critical pattern). ⛔ FLAGGED at L1 (Security) or L3 (Hallucination) → caller must fix before deploy. Return to caller.