cross-review

Overview

Independent review by two different Claude model tiers on the same diff. Produces a consensus table highlighting agreements, disagreements, and "taste decisions" where models diverge on subjective issues. Uses Sonnet (primary) + Haiku (secondary) by default — different model tiers catch different things due to reasoning depth differences.

How this differs from other review skills:

• /code-review — single-model, multi-reviewer (same model, different lenses). This skill uses different models for independent validation.
• /cto-review — architectural critique. This skill is diff-level review with cross-model consensus.

Cost note: This skill runs two full review passes. Reserve for critical changes (auth, payments, data integrity, public API contracts), not routine PRs.

Step 1: Get the Diff

Same as /code-review Step 1:

• PR number: gh pr diff <number>, gh pr view <number> --json title,body,baseRefName
• Branch name: git diff main...<branch>
• No argument: git diff main...HEAD

Capture: full diff, list of changed files, PR description (if available).

Step 2: Parallel Independent Review

Spawn 2 agents in parallel, each with a different model. Both receive the identical diff and file list.

Agent A: Primary Model (Claude — Sonnet)

Agent(
  subagent_type="Explore",
  model="sonnet",
  prompt="Review this code diff independently. For each finding, report: file, line, severity (Critical/Important/Minor), category (correctness/security/performance/convention), description, and confidence (1-100). Be thorough but only report findings you're confident about (>70%). Do not speculate.

  Diff:
  <full diff>

  Changed files: <file list>
  PR context: <description>

  Return findings as a JSON array."
)

Agent B: Secondary Model (Haiku — fast, different perspective)

Agent(
  subagent_type="Explore",
  model="haiku",
  prompt="<same prompt as Agent A with identical diff>"
)

Both agents run independently with no knowledge of each other's findings.