Multi-Review Aggregation

Dispatch N independent code reviews and aggregate findings. Each reviewer catches different bugs -- union preserves the long tail that single-shot misses.

Research basis: SWR-Bench (arXiv 2509.01494) -- N independent reviews: 43.67% F1 improvement, 118.83% recall improvement. Diminishing returns past N=5; N=3 captures most improvement.

Core principle: Independence via separate Task dispatches -- same base prompt, no shared context.

N Selection by Tier

When N=1, skip this skill -- use standard single code review.

Parallel Dispatch Pattern

After spec review passes, dispatch N independent reviews (run_in_background=True, each gets "Reviewer i of N"). Always aggregate when 2+ reviewers succeed — even unanimous approval may contain different Minor findings, Suggestions, or Not Checked items. Dispatch aggregator (haiku model).

Full dispatch code: see references/dispatch-code.md. Aggregator prompt: see ./aggregator-prompt.md.

Aggregation Algorithm

Severity Voting

Severity levels: Critical > Important > Minor > Suggestion

Verdict

• "Ready to merge: Yes" -- zero Critical AND zero Important AND majority approved
• "Ready to merge: With fixes" -- only Minor/Suggestion after aggregation
• "Ready to merge: No" -- any Critical or Important remain

Full deduplication/merging rules: see references/aggregation-details.md.

Output Format

## Strengths
- [strength] [Reviewers: 1, 2, 3]

## Issues
### Critical / Important / Minor / Suggestion
- [issue] [Reviewers: N, N] -- file:line
  (note downgrade/security provenance as applicable)

## Uncovered Paths
- [path/scenario] [Reviewers: X, Y]

## Not Checked
- [area] [Reviewers: X, Y]

## Assessment
Ready to merge: [Yes/With fixes/No]
Reviewers: X/N approved, Y requested changes

Full format spec: see references/output-provenance.md.

Red Flags

Never:

• Share context between reviewers (defeats independence)
• Use N>1 for pro/api tier (budget constraint)
• Skip aggregation when reviewers disagree
• Downgrade security findings even as lone findings

Always:

• Dispatch all N reviews in parallel
• Include reviewer number in each dispatch prompt
• Use haiku for aggregation
• Record per-reviewer metrics separately

Reference Files

• references/dispatch-code.md: Full dispatch flow with on_spec_review_pass handler
• references/aggregation-details.md: Deduplication, strengths merging, malformed output, timeout recovery
• references/output-provenance.md: Provenance annotation rules and full output format spec
• references/metrics-and-cost.md: Per-reviewer metric keys, cost impact, per-tier breakdown
• aggregator-prompt.md: Aggregator Task dispatch prompt template