From forge-core
Multi-worker 검수 스킬 (Codex + Gemini Double / Opus + Codex + Gemini Triple). 단일 Codex 검수 대비 100% 보완 카테고리 커버. 트리거: /cr-multi, /cr-double, /cr-triple, plan/spec 저장 후 자동(CR_MULTI_AUTO=on), plateau 3회 자동 승격.
How this skill is triggered — by the user, by Claude, or both
Slash command
/forge-core:cr-multiThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Codex + Gemini (Double) 또는 Opus + Codex + Gemini (Triple) 병렬 리뷰 + Triage 합산 verdict.
Codex + Gemini (Double) 또는 Opus + Codex + Gemini (Triple) 병렬 리뷰 + Triage 합산 verdict.
# Double (기본 — Codex + Gemini)
/cr-double ${FORGE_OUTPUTS:-$HOME/forge-outputs}/11-platform/pipelines/plans/my-plan.md
# Triple (plateau 자동 승격 또는 중요 spec)
/cr-triple ${FORGE_OUTPUTS:-$HOME/forge-outputs}/02-product/projects/forge-platform/specs/my-spec.md
| 모드 | Worker | 합산 |
|---|---|---|
| Double | Codex + Gemini | codex×0.6 + gemini×0.4 |
| Triple | Opus + Codex + Gemini | opus×0.3 + codex×0.4 + gemini×0.3 |
${FORGE_OUTPUTS:-$HOME/forge-outputs}/docs/reviews/cr-multi/
{DATE}-{slug}-v{N}-codex.json
{DATE}-{slug}-v{N}-gemini.json
{DATE}-{slug}-v{N}-report.md
CR_MULTI_AUTO=off 기본 (명시 opt-in 필요)${FORGE_OUTPUTS:-$HOME/forge-outputs}/.claude/audit/cr-multi-calls.jsonlcr-multi 실행 후 usage 데이터 기록:
bash ~/.claude/scripts/cache-stats-logger.sh cr-multi "$MODEL" "$CACHE_READ" "$CACHE_CREATION" "$RAW_INPUT" cr-review
usage 필드는 Anthropic SDK response.usage 에서 추출. 미지원 시 0 기본값 사용.
⚠️ mcp__codex__/mcp__gemini__ 토큰 = Workflow 외부 선발행 필수.
# Phase 0: 토큰 선발행 (Workflow 전)
FORGE_TEST_MODE=1 python3 ~/.claude/skills/approve-worker/scripts/approve-worker-sign.py \
--task cr-multi-${SLUG} --worker codex-critic --tools mcp__codex__codex --paths "${TARGET}"
// Phase 1+: Workflow 실행 (GitNexus StructuralContext + 3-LLM parallel)
Workflow({
script: Bash("cat ~/.claude/skills/cr-multi/workflow.js"),
args: { slug: SLUG, targetPath: TARGET, mode: 'triple', stage: STAGE }
})
Agent Teams fallback: CLAUDE_CODE_DISABLE_WORKFLOWS=1 시 기존 Agent 패턴.
~/forge/.claude/commands/cr-multi.md~/.claude/rules-on-demand/multi-gate-review.md~/forge/shared/scripts/cr-multi-triage.py~/forge/shared/scripts/cr-multi-plateau-guard.py독립 Evaluator subagent가 산출물 품질을 검증합니다.
Evaluator 역할: 산출물 독립 검증
모델: claude-haiku-4-5 (경량, 편향 최소화)
격리: 메인 컨텍스트 오염 방지
판정 기준:
eval_cases.jsonl에 결과 자동 누적.
연속 2라운드 score 진전 <5점 = plateau 신호. 즉시 4 옵션 제시 (A 추가 라운드 / B AD-50 override / C 폐기 / D 극단 단순화). D 우선 권고 (over-engineering 거부 — enforcement-theater-prevention 정합).
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub moongci38-oss/forge-plugins --plugin forge-core