메뉴
Orchestrated code review of git diff using parallel agents. Standard mode (diff-scoped) or full mode (adds Roslyn semantic analysis). Produces severity-grouped plan.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Batch-verify and delete dead code: rg-confirm zero callers for N symbols, delete files, run the build once, stage results.
Run benchmarks and check regressions, add benchmark gating to CI, or audit GitHub Actions workflows. Determines mode from user intent.
Batch-apply findings from the most recent --local review. One confirmation, one test run.
Unified code review: --local (diff review), --pr (adversarial PR review), --apply (batch fixes from last review).
Audit .NET solution health across 13 dimensions: dependency graph, solution sync, CPM compliance, package health, framework consistency, IVT & encapsulation, build config, analyzer config, NuGet config, suppression hygiene, output type & AOT, public API surface, and lock files & pruning. Works with any .NET solution (.slnx or .sln). Parses all config files once, spawns 3 parallel agents, and produces a severity-grouped report with a Mermaid dependency graph. Triggers on: solution audit, audit solution, project graph, dependency audit, solution health, audit projects, check solution.
Scans source and .claude/ files for anti-pattern violations, applies the top mechanical fixes, and reports pre/post counts.
| name | review-local |
| description | Orchestrated code review of git diff using parallel agents. Standard mode (diff-scoped) or full mode (adds Roslyn semantic analysis). Produces severity-grouped plan. |
| when_to_use | dispatched by review router |
| argument-hint | [--full] [days N | commits N] [--perf] [--test-quality] [dry run] [critical only] |
| user-invocable | true |
Comprehensive code review using orchestrated agents. Discovers changed .NET files from git history, dynamically identifies code review agents, runs them in parallel, and produces a consolidated report grouped by severity.
Standard mode (default): Day-to-day review, fast turnaround, diff-scoped agent review only.
Full mode (--full): Pre-ship review, maximum coverage — adds solution-wide Roslyn semantic analysis (antipatterns + circular dependencies).
code review days 1 # Standard
review --full last 3 days # Full 3-vector
dry run # Preview only
critical only # Filter to CRITICAL
--perf # Spawn a performance-focused review agent
--test-quality # Load enriched test rubric
| Pattern in user message | Scope |
|---|---|
days N / last N days | scripts/diff.sh --json (days-scoped) |
commits N / last N commits | scripts/diff.sh --json (commit-range) |
dry run / preview | Discovery only, no agents |
critical only / high+ | Filter report by min severity |
--perf | Spawn a performance-focused review agent on diff scope |
--test-quality | Inject enriched Test Quality Rubric |
If no scope is provided, ask for one.
--perf: Spawn a performance-focused review agent on diff scope in addition to normal review agents, using the anti-pattern catalog in references/perf-patterns.md. Findings merged into report.
--test-quality: Load enriched Test Quality Rubric from references/review-rubrics.md and inject into agent prompts for tests/**/*.cs files.
Run in parallel — no ordering dependency:
scripts/diff.sh --json.claude/agents/; filter to code-review-relevant agents (include: code, review, check, fix, validate, compliance, security, quality, refactor, build, test, warning, diagnostic, concurrency, performance — exclude: pipeline, sanitizer, lode, corpus, mhtml, markdown, web, fetch, download)From diff: filter to .NET files; detect mechanical commits (namespace rename → MECHANICAL, exclude); generate unified diff. Diff delivery by size: <=30KB inline; 30-40KB write to .claude/tmp/cr-<agent>.diff; >40KB split at file boundaries. Clean up .claude/tmp/cr-*.diff after completion.
Vector 1: Code-Review Agents (always runs) Spawn agents in parallel — pass each the unified diff and severity rules. Include: "Review only the diff provided. Flag pre-existing concerns as 'adjacent concern' without investigating." Handle timeouts/errors gracefully.
Vector 2: Roslyn Navigator (--full only) Run MCP calls in main context (session-only):
mcp__cwm-roslyn-navigator__detect_antipatterns with the solution filemcp__cwm-roslyn-navigator__detect_circular_dependencies with the solution fileSeverity mapping: Circular (cross-project) → HIGH; Circular (namespace) → MEDIUM; Antipatterns → MEDIUM; Generated code → SKIP; Informational → LOW.
Graceful degradation: if MCP unavailable, note "Roslyn navigator: MCP server not running — skipped".
See references/implementation.md for full orchestration details: dedup algorithm, severity grouping, plan file format (plan.md/tasks.md/context.md), report template, and analyzer-candidate mining integration.
analyzer-review-miner agent in background to mine findings for new analyzer candidates| Severity | Description | Examples |
|---|---|---|
| CRITICAL | Blocks shipping | Build failures, security vulns, race conditions |
| HIGH | Must fix before merge | Analyzer errors, concurrency, perf regressions |
| MEDIUM | Should fix | Code smells, refactoring, missing XML docs |
| LOW | Nice to have | Style preferences, minor optimizations |
Seven rubric checklists in references/review-rubrics.md. Inject relevant sections into agent prompts based on diff content. Eighth rubric for multi-file consistency in references/cross-file-consistency.md.