code-review

Code Review

Ringmaster's detection layer asks two questions of every change, and they are genuinely different:

Spec — did we build the right thing? Does the change do what was asked — the plan, the PRD, the ticket — no more and no less?

Standards — did we build the thing right? Is the code correct, safe on its edges, and does it look like it belongs in this repo?

A single reviewer holding both questions at once bleeds them together — excusing messy code because "it matches the ticket", or polishing style while a whole requirement is quietly missing. So this skill runs the two axes as parallel fresh-context sub-agents (agents/code-reviewer.md, dispatched twice — once per axis) and lays their two reports side by side. Separation is the point.

This skill is stack-independent. It reviews; it does not fix, and it does not re-run the Security Gate — it stands beside it.

---

The safety rails (inherited, non-negotiable)

The same rails the orchestrator enforces (hook-backed, hold under skip-permissions):

1. You review; you never fix. Read-only. A real defect goes into the record and back to the human — you don't patch it.
2. You never commit, push, or write history. The only thing this skill writes is the ledger's gate.clean record. The human commits and ships.
3. Never run anything against production. Running the change's own tests to confirm a claim is DEV/UAT only.
4. Never rubber-stamp. ✅ clean is a valid verdict only when the paths were actually traced. Manufacturing findings to look thorough is as dishonest as waving a broken change through.

---

Where this sits in "code review with Ringmaster"

Security Gate  →  code-review  →  Ownership Review  →  sign-off
  (secrets,       (Spec: right       (reconstruct         (gate.owned)
   injection,      thing? +           understanding,
   authz)          Standards:         teach the holes)
                   built right?)

Detection runs in order: the Security Gate clears the diff of vulnerabilities first (it can block), then this skill reviews quality and spec-fit, then the Ownership Review makes sure the human actually understands what they're about to own. This skill produces the clean verdict the ownership review assumes has already run.