dual-agent-reviewlisted
Install: claude install-skill stealth-engine/skills
# Dual-agent review
Run **two or more independent reviewers** over the **same diff**, then reconcile
their findings, fix the real ones, and report. Two reviewers that can't see each
other catch *different* things and **disagree on the false ones** — so crossing
them cuts both misses and false positives that a single reviewer ships. The calling
(orchestrator) agent owns triage, fixes, and the final report.
Exact reviewer prompts, CodeRabbit CLI recipes, the parallel-fix prompt, and the
report template live in [`reference/reviewer-prompts.md`](./reference/reviewer-prompts.md)
(load on demand).
## Scope: the diff, by default
The unit of review is the **change set**, not the whole repo — same default as
CodeRabbit (a diff/PR reviewer). Default = the branch's diff vs its base (or a PR's
diff). The **diff is the report scope; the full changed files are context** — a
reviewer reads surrounding code to judge a change but only reports on changed lines.
For a *whole-codebase* sweep use the partitioned
[`multi-agent-codebase-audit`](../multi-agent-codebase-audit/SKILL.md) instead — this isn't it.
## The flow
```text
1. SCOPE diff ─► 2. FAN OUT reviewers (parallel/bg) ─► 3. COLLECT + DEDUPE + TRIAGE
├─ A: Sonnet subagent │
├─ B: CodeRabbit CLI ▼
└─ (C: extra lens…) 4. FIX ◄── valid?
│ direct, or 1 fix-subagent/issue