Review Board: Reviewing an Auth Middleware PR

Scenario

A senior engineer opens a 340-line PR adding JWT authentication middleware to a Go HTTP API. The PR touches pkg/auth/middleware.go, pkg/auth/jwt.go, a new pkg/ratelimit/ package, and structured logging additions across four handlers. It’s the kind of PR where a single reviewer would focus on correctness and miss the security or performance angles.

Why This Topology

The changeset crosses three review dimensions – security (JWT validation, token handling), performance (rate limiting, middleware overhead), and test coverage. A single reviewer would likely anchor on one dimension. Review Board assigns each dimension to a specialist who reviews the full diff through their lens, producing structured findings the lead merges into one review comment.

Team Shape

Role Count Responsibility
Lead 1 Assign lenses, synthesize findings into single review comment
Security Reviewer 1 JWT handling, token validation, auth bypass risks
Performance Reviewer 1 Middleware chain overhead, rate limiter efficiency, allocations
Test Reviewer 1 Coverage gaps, edge cases, assertion quality

Spawn Prompt

Create an agent team to review PR #247 adding auth middleware.
Spawn three reviewers:
- Security: JWT validation, token storage, auth bypass risks, input sanitization.
- Performance: middleware overhead, rate limiter memory usage, hot path allocations.
- Test coverage: missing edge cases, assertion quality, error path coverage.
Have each report findings as must-fix vs nice-to-have with file:line evidence.
Then synthesize into one review comment.

Trade-offs

  • Lens overlap is inherent. When a single code structure has implications across dimensions (e.g., an unbounded map is both a security DoS vector and a performance leak), multiple reviewers will flag it independently. The lead must deduplicate during synthesis.
  • Require consistent output format. Severity / file:line / evidence / fix – this structure makes the lead’s synthesis step mechanical rather than interpretive. Without it, merging three free-form reviews is painful.
  • Three reviewers is the sweet spot for most PRs. A fourth lens (e.g., API compatibility) is worth adding only for public-facing API changes. Beyond four, the coordination cost outweighs the marginal findings.
  • The best findings come from intersections. Issues that span multiple lenses are exactly what a single generalist reviewer would note only once (if at all). The multi-lens approach surfaces these systematically.