| name | multi-llm-review |
| description | Multi-LLM parallel adversarial review — Claude(Sonnet) + Gemini double (default) or + Codex/GPT triple (opt-in). Weighted score merge, dedup+confidence scoring, plateau detection, completeness critic, per-finding refute. |
| input | target-file path + mode (double|triple) + stage (plan|code|test|final) |
| output | ${CR_OUTPUT_DIR:-.multi-llm-review}/reviews/{stage}/{slug}-multi-llm-review.json |
| eval_cases | off |
/multi-llm-review
Claude(Sonnet) + Gemini (double, default) or + Codex/GPT (triple, opt-in) parallel adversarial review with weighted Triage verdict.
Quick Start
/review-double path/to/my-code.ts
/review-triple path/to/important-spec.md --cr on
Modes
| Mode | Workers | Score |
|---|
| Double (default) | Sonnet + Gemini | sonnet×0.35 + gemini×0.3 / 0.65 |
| Triple (opt-in) | Sonnet + Codex + Gemini | sonnet×0.35 + codex×0.35 + gemini×0.3 |
BYO-key Requirements
| Capability | Requirement |
|---|
| Claude (Sonnet/Haiku) | Claude Code built-in |
| Gemini review | GEMINI_API_KEY env (read by gemini-text MCP server) |
| Codex/GPT triple | mcp__codex__codex tool + ChatGPT subscription; pass crMode:'on' |
| GitNexus structural context | GitNexus MCP (optional; grace-degrades if unavailable) |
Output
${CR_OUTPUT_DIR:-.multi-llm-review}/reviews/{stage}/{slug}-multi-llm-review.json
Evidence JSON format: {verdict, score, issues[], mode, slug, degraded}
Audit log: ${CR_OUTPUT_DIR:-.multi-llm-review}/audit/multi-llm-review-calls.jsonl
Workflow Execution
Workflow({
script: Bash("cat ~/.claude/skills/multi-llm-review/workflow.js"),
args: { slug: SLUG, targetPath: TARGET, mode: 'triple', stage: STAGE }
})
Fallback (CLAUDE_CODE_DISABLE_WORKFLOWS=1): use Agent pattern directly.
Optional Features
crMode — Worker Control
crMode: 'degrade' (default) → Sonnet + Gemini only
crMode: 'on' → Sonnet + Codex + Gemini (requires Codex MCP)
crMode: 'off' → Gemini only
crCompleteness — Completeness Critic
Haiku "what was missed" gate. Auto-on for stage=final. Opt-in for other stages.
args: { ..., crCompleteness: true }
crRefute — Per-Finding Skeptic Vote
Suppresses false-positive HIGH findings via majority skeptic vote.
HARD RULE: security category + CRITICAL severity = always KEEP.
args: { ..., crRefute: true, crRefuteN: 3 }
crLens — Lens Diversification
Differentiates worker review focus (holistic / security+correctness / spec-drift+perf).
Do not combine with crCompleteness simultaneously (completeness critic may misidentify lens-intentional omissions as gaps).
args: { ..., crLens: true }
GitNexus — Structural Context
If mcp__gitnexus__* tools are available, Phase 0 enriches review with changed symbols and upstream impact. If unavailable, structuralCtx=null and review continues normally.
Scoring & Verdict
| Verdict | Condition |
|---|
| PASS | combined ≥ 80 AND no HIGH issues |
| WARN | combined ≥ 60 (HIGH issues present) |
| FAIL | any CRITICAL issue OR quorum < 2 workers OR combined < 60 |
| SKIP | fallow: no 24h change + prior review on record |
Plateau Detection
Two consecutive rounds with score delta < 5 = plateau signal.
Options: A more rounds / B override / C discard / D extreme simplification.
Pass prevScore arg for delta tracking:
args: { ..., prevScore: 72.5 }
Prohibited Behaviors
Workers must NOT:
- Add issues without evidence to inflate/deflate scores
- Re-raise resolved issues from prior rounds with new wording (fabrication)
- Require enterprise features (HA, multi-tenancy, distributed transactions) for SME/MVP scope
- Demand full redesign without understanding author's design intent
- Recommend copying external code without license/attribution
Requester Rules
- Never skip review because "it's a small change"
- Do not proceed past CRITICAL issues without Human approval
- Review HIGH-severity issues before advancing on WARN verdict
- Evaluate technical feedback content — do not accept/reject without reading
