review-local

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

Review — Local Mode

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).

Usage

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

Scope Parameters

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 and --test-quality Flags

--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.

How It Works

Phase 1: Discovery

Run in parallel — no ordering dependency:

• Parse arguments from user input
• Gather change context: scripts/diff.sh --json
• Discover agents dynamically: read .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.

Phase 2: Execution

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):

1. mcp__cwm-roslyn-navigator__detect_antipatterns with the solution file
2. mcp__cwm-roslyn-navigator__detect_circular_dependencies with the solution file

Severity 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".

Phase 3: Plan Generation

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.

Phase 4: Exit

1. Offer to execute the plan if CRITICAL or HIGH issues
2. Spawn the analyzer-review-miner agent in background to mine findings for new analyzer candidates
3. Return exit code: 0 (clean/LOW), 1 (MEDIUM/HIGH), 2 (CRITICAL)

Severity Classification

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

Review Rubrics

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.

References

• references/compliance-and-format.md
• references/review-rubrics.md
• references/cross-file-consistency.md
• references/error-handling.md
• references/implementation.md
• references/clanker-patterns.md
• references/perf-patterns.md