Code Review

/code-review

Launch a parallel team of reviewers. Synthesize findings. Fix blocking issues automatically. Loop until clean or escalate to human.

Architecture

Five sub-agents in parallel, each with a distinct lens:

• critic — generalist evaluator, grades against criteria below
• ousterhout — deep modules, information hiding, interface simplicity
• carmack — pragmatic shippability, over-engineering detection
• grug — complexity hunting, layer counting
• beck — TDD quality, one behavior per test, red-green-refactor

Optional external reviewers: Think Tank CLI, Cerberus CLI.

Workflow

1. Gather the diff

Get the diff via git diff main...HEAD or the specified scope.

2. Launch all reviewers in parallel

Spawn all five sub-agents simultaneously in a single message. Give each the full diff and relevant codebase context. Each reviewer should return:

• Ship / Don't Ship verdict
• Top 3 concerns (file:line + specific fix)
• One sentence: best thing about this code

If the harness limits concurrency, run the critic first, then the four philosophy agents in parallel, or run all sequentially as a fallback.

3. Synthesize

Collect all verdicts. Deduplicate overlapping concerns. Rank by severity.

4. Live Verification

If the change has user-facing components, the critic should verify the implementation actually works — not just that it reads well.

• Web: Navigate to affected pages, exercise the feature, check for console errors and broken UI.
• API: Hit the changed endpoints, verify responses match the spec.
• CLI: Run the commands, verify output matches expectations.

If live verification fails, it's a blocking issue — treat as Don't Ship. Skip for pure refactors, config changes, or library work with no runtime.

5. Gate

• All Ship → approve, proceed to merge
• Any Don't Ship → spawn a builder sub-agent for each blocking concern, giving it the specific file:line and fix instruction. Builder fixes, runs tests. Then re-review (return to step 2). Max 3 iterations.

6. Escalate

If still blocked after 3 iterations, report all findings to the user. The issue needs human judgment.

Grading Criteria

The critic applies structured grading (adapt per project):

Weight correctness and depth higher — models score well on craft by default but underperform on architectural depth and actual correctness.

Plausible-but-Wrong Patterns

LLMs optimize for plausibility, not correctness. Reviewers must actively hunt for code that looks right but isn't:

• Wrong algorithm complexity (O(n²) where O(log n) is needed)
• Unnecessary abstractions (82K lines vs 1-line solution)
• Stub implementations that pass tests but don't actually work
• "Specification-shaped" code — right module names, wrong behavior
• Missing invariant checks that only matter at scale

Simplification Pass

After review passes, if diff > 200 LOC net:

• Look for code that can be deleted
• Collapse unnecessary abstractions
• Simplify complex conditionals
• Remove compatibility shims with no real users

Output

## Code Review Summary

**Verdict:** Ship / Don't Ship / Escalated

### Findings
| Severity | File:Line | Issue | Fix |
|----------|-----------|-------|-----|
| blocking | src/auth.ts:42 | Leaky abstraction | Extract to module |

### Reviewer Consensus
- critic: Ship
- ousterhout: Ship
- carmack: Don't Ship — over-engineered auth layer
- grug: Ship
- beck: Ship

### Iterations: 2 (fixed auth layer on iteration 2)

Gotchas

• Self-review leniency: Models consistently overrate their own work. The critic must be a separate sub-agent, not the builder evaluating itself.
• Reviewing the whole codebase: Review the diff, not the repo. git diff main...HEAD is the scope.
• Vague feedback: "Needs improvement" is useless. Every concern must have file:line + specific fix.
• Infinite loop: Cap at 3 review iterations. If still blocked, escalate — the issue needs human judgment.
• Skipping the bench: Running only the critic misses structural issues. The philosophy agents add perspectives the critic doesn't cover.
• Treating all concerns equally: Blocking issues (correctness, security) gate shipping. Style preferences don't.