Code Review — 5 Parallel Lenses

Core principle: Five specialized reviewers in parallel, each with a confidence score. Synthesize into one verdict.

When to Request Review

Mandatory: After completing a feature, before merge to main, after each task in subagent-driven development.

Optional: When stuck, before refactoring (baseline), after fixing a complex bug.

---

Step 1: Get the Diff Range

Local branch:

BASE_SHA=$(git merge-base HEAD origin/main)
HEAD_SHA=$(git rev-parse HEAD)

PR review:

gh pr view <number> --json baseRefOid,headRefOid \
  --jq '"BASE=\(.baseRefOid[:8]) HEAD=\(.headRefOid[:8])"'

Recent commit:

BASE_SHA=$(git rev-parse HEAD~1)
HEAD_SHA=$(git rev-parse HEAD)

---

Step 2: Dispatch 5 Parallel Subagents

Launch ALL five simultaneously. Each uses code-reviewer.md with its {LENS} filled in.

#	Lens	Focus
1	Correctness	Bugs, logic errors, edge cases, data races, panics
2	Architecture	Design patterns, coupling, DDD alignment, type safety
3	Performance	Allocations, clone/copy, hot paths, complexity
4	Security & Safety	Unsafe blocks, input validation, error propagation, secrets
5	Tests & Requirements	Coverage, spec compliance, test quality, TDD adherence

Dispatch template (repeat for each lens, all in parallel):

Agent: general-purpose
Prompt: [code-reviewer.md with placeholders filled]
  LENS: <lens name>
  LENS_FOCUS: <focus from table above>
  WHAT_WAS_IMPLEMENTED: <feature description>
  PLAN_OR_REQUIREMENTS: <requirements or task>
  BASE_SHA: <base>
  HEAD_SHA: <head>
  DESCRIPTION: <one-line summary>

---

Step 3: Synthesize Results

After all 5 agents return, produce a unified report:

Synthesis Format

## Code Review — Synthesis

### Lens Results

| Lens | Confidence | Critical | Important | Minor |
|------|-----------|----------|-----------|-------|
| Correctness    | XX% | N | N | N |
| Architecture   | XX% | N | N | N |
| Performance    | XX% | N | N | N |
| Security/Safety| XX% | N | N | N |
| Tests & Reqs   | XX% | N | N | N |
| **Overall**    | **XX%** | **N** | **N** | **N** |

> Overall confidence = weighted average (Correctness×0.25, Architecture×0.20,
> Performance×0.20, Security×0.20, Tests×0.15)

---

### Critical Issues (Must Fix)
[Deduplicated across all lenses — file:line, what, why, fix]

### Important Issues (Should Fix)
[Deduplicated across all lenses]

### Minor Issues (Nice to Have)
[Deduplicated across all lenses]

### Strengths
[Specific positives from all lenses]

### Verdict

**Ready to merge?** Yes / No / With fixes

**Confidence:** XX% — [what drove the score up or down]

**Reasoning:** [2-3 sentences technical assessment]

Deduplication Rule

If two lenses flag the same issue: keep the higher-severity instance, note which lenses agreed (increases confidence).

Confidence Score Interpretation

Score	Meaning	Action
90-100%	High confidence, thorough review	Proceed normally
70-89%	Good coverage, minor gaps	Note gaps, proceed
50-69%	Moderate — complex or large diff	Flag for extra scrutiny
< 50%	Low — diff too large or domain unclear	Consider splitting the PR

---

Step 4: Act on Feedback

Severity	Action
Critical	Fix immediately — do not proceed
Important	Fix before merging
Minor	Note for later or fix now
Reviewer wrong	Push back with code/test evidence

Never:

• Skip review because "it's simple"
• Ignore Critical issues
• Proceed past Important issues without fixing them

---

PR Review Mode

For GitHub PRs, add PR context to each subagent:

gh pr view <number> --json title,body,additions,deletions,changedFiles
gh pr diff <number>

Pass PR title + body as {PLAN_OR_REQUIREMENTS} so each lens can check spec compliance.