X-Agents Cross-Review

Launch X parallel agents to independently review the same scope from different perspectives. Each agent gets full tool access, the highest-performance model, and maximum thinking time.

<essential_principles>

Core Rules

1. No scope splitting — Every agent reviews the FULL scope independently
2. Different perspectives — Each agent gets a unique review lens (see perspectives library)
3. +1 Philosophy — When X is not specified, auto-determine optimal count + 1 (the last agent often catches what others miss)
4. Unified report — Consolidate all results into a single report with consensus analysis

</essential_principles>

Phase 1: Context Gathering

1. Parse the review target from user input (code, spec, PR, etc.)
2. Gather the source material:
   • Notion spec → mcp__claude_ai_Notion__notion-fetch
   • OpenAPI spec → mcp__zumen-api-docs__read_project_oas_* or curl to file
   • Code → git diff dev --name-only for changed files
   • PR → gh pr view for PR description and diff
3. Determine agent count:
   • If user specifies X → use X
   • If not specified → estimate based on scope complexity + 1

Scope Size	Base Agents	+1	Total
Small (1-3 files, single concern)	3	+1	4
Medium (4-10 files, multi-concern)	5	+1	6
Large (10+ files, cross-cutting)	8	+1	9
Critical (production deploy, security)	10	+1	11

Phase 2: Perspective Assignment

Assign perspectives from the perspectives library. Rules:

• First agent = Baseline (全要件網羅)
• Last agent = Devil's Advocate (隠れたリスク・仕様の曖昧さ)
• Middle agents = Domain-specific lenses from the library
• Never assign the same perspective twice
• Match perspectives to the review target (e.g., i18n perspective only if i18n is involved)

Phase 3: Agent Launch

Launch ALL agents in a single message (parallel execution). Each agent config:

Agent(
  name: "reviewer-{NN}-{perspective-slug}",
  model: "opus",
  mode: "bypassPermissions",
  run_in_background: true,
  prompt: <constructed from template below>
)

Agent Prompt Template

## Task: {review_type} (Reviewer #{N}/{total} — {perspective_name})

{context_material}

### Review Scope
{file_list_or_spec}

### Instructions
- Use Serena tools for code reading (find_symbol, get_symbols_overview, search_for_pattern)
- Use mcp__sequential-thinking__sequentialthinking for step-by-step analysis
- Use Context7 for library documentation when needed
- Review ALL items in scope from your perspective: {perspective_description}

### Requirements Checklist
{requirements_list}

### Output Format
## Reviewer #{N}: {perspective_name}
### Requirements Status
| # | Requirement | Status | Evidence |
...
### ISSUES FOUND
- [severity: CRITICAL/HIGH/MEDIUM/LOW] [location]: [description]
### OBSERVATIONS
### Summary
[1-2 sentence assessment]

Phase 4: Progress Tracking

Report progress as agents complete:

• **#N 完了** ({perspective}) — {brief_summary}. 残り M/{total} ⏳

Phase 5: Unified Report

After ALL agents complete, produce:

Report Structure

## X重レビュー統合レポート — {target}

### 全体判定: {PASS/FAIL} (CRITICAL N件)

### 要件別コンセンサス
| # | Requirement | Verdict | Reviewers |
(items where multiple reviewers agree)

### コンセンサス発見 (複数レビュアー一致)
| # | Finding | Severity | Reviewers | Fix |
(sorted by reviewer count, then severity)

### 単独発見 (1レビュアーのみ)
| # | Finding | Severity | Reviewer | Note |

### PASS 項目サマリー

### Insight
(What the multi-review approach uniquely uncovered)

Consolidation Rules

1. Consensus = 2+ reviewers flagging the same issue → elevate severity
2. Unique findings from single reviewer → keep original severity
3. Pre-existing patterns (not regressions) → note as "既存パターン"
4. BE-scope items → mark as "NOT FE SCOPE" / "BE確認事項"
5. Deduplicate identical findings, credit all discovering reviewers

References

• Perspectives Library — Review lens catalog