review-suite

Perform a comprehensive security analysis of the specified code:

Analyze memory fragments and documentation for secrets, API keys, and personally identifiable information before committing or sharing.

Iterative review-fix loop — spawns parallel code-reviewer subagents per file (single Task for 1 file, parallel Tasks for 2+ files); max 5 rounds; concurrent fix application via parallel fixer Task() agents; optional plan_summary parameter for intent-aligned review; git fallback for auto-detecting changed files when target_files is empty. Critical fixes always applied; Advisory+Fix-block fixes auto-applied; Advisory/YAGNI skipped; loops per-file until clean or 5 rounds reached, produces a summary.

Ultimate Adversarial Auditor. Combines language-specific "Domain Radar" with a "Suspicion-First" internal monologue for 100% logic trap detection.

Comprehensive prompt analysis and enhancement using phased development framework. **AUTOMATICALLY INVOKE** this agent when user mentions "review prompt", "improve prompt", "analyze prompt", or when editing agent/skill markdown files. **STRONGLY RECOMMENDED** for new agent creation, prompt optimization, and LLM instruction refinement. **NOT for:** - Quick technique check (use /review-bench:improve-prompt --mode critique — faster, produces scorecard) - A/B prompt comparison with execution (use /review-bench:compare-prompts) - GAS system prompt optimization (use /review-bench:optimize-system-prompt)

LLM judge that semantically evaluates code reviewer output against ground-truth issues. Used by review-fix-bench harness to replace fragile regex-based finding extraction and matching. Outputs structured JSON only — no prose wrapper.

Simplified orchestrator for the iterative review-fix loop. Detects files → delegates to specialized reviewers in standalone mode → aggregates results → commits/PRs. Reviewers (code-reviewer, etc.) handle their own inner loops, auto-fixes, and re-verification. **AUTOMATICALLY INVOKE** after implementing features, fixing bugs, before committing, or after plan implementation completes.

Generative Adversarial Network (GAN) Red Team. Writes novel, obfuscated vulnerabilities to test the limits of the code-reviewer agent.