MCP Code Crosscheck

This server provides comprehensive code review capabilities through both prompts and tools, specializing in bias-resistant evaluation with multiple review strategies. CAPABILITIES: - 'code_review' prompt: Direct, comprehensive code review with adversarial approach - Tools for bias-resistant cross-model evaluation with strategy selection and GitHub integration REVIEW STRATEGIES: The 'review_code' tool supports three distinct strategies: 1. 'adversarial': Uses competing team mindset for critical analysis - Assumes reviewer is from competing team with incentive to find issues - Focuses on identifying bugs, edge cases, and alternative implementations - Structured markdown output format 2. 'bias_aware': Explicitly identifies and ignores bias triggers before evaluation - First detects potential bias triggers (author comments, tool names, style choices) - Then evaluates purely on functional correctness while ignoring identified biases - Includes 'Bias Triggers Found' section in markdown output 3. 'hybrid': Combines bias detection with adversarial review - Phase 1: Identifies bias triggers like bias_aware strategy - Phase 2: Applies adversarial mindset while avoiding identified biases - Includes 'Bias Triggers Found' section in markdown output STRATEGY SELECTION GUIDANCE: - Use 'adversarial' for: Standard critical review with competitive framing - Use 'bias_aware' for: Maximum bias resistance, especially with AI-generated code - Use 'hybrid' for: Comprehensive review combining bias detection with critical analysis PROMPT USAGE: Use the 'code_review' prompt for immediate, comprehensive code review: - Uses adversarial strategy by default - Works with explicit code snippets or file context (code parameter is optional) - Provides structured markdown output TOOL WORKFLOWS (for bias-resistant review of AI-generated code): PREFERRED WORKFLOW (when GitHub MCP server is available): 1. Use GitHub MCP server tools to fetch commit/PR data and author information 2. Use this server's 'detect_model_from_authors' tool to identify AI models from author lists 3. Use this server's 'review_code' tool with chosen strategy for bias-resistant cross-model review FALLBACK WORKFLOW (when GitHub MCP server is not available): 1. Use this server's 'fetch_commit' or 'fetch_pr_commits' tools to get GitHub data 2. Use this server's 'detect_model_from_authors' tool to identify AI models 3. Use this server's 'review_code' tool with chosen strategy for bias-resistant cross-model review CROSS-MODEL REVIEW IMPLEMENTATION: The 'review_code' tool attempts bias-resistant evaluation through two mechanisms: AUTOMATIC SAMPLING (preferred): - Tool uses MCP sampling API to request client use a different model than the generation model - Provides model exclusion preferences and fallback hints to guide client selection - Success depends on client's sampling capabilities MANUAL CROSS-MODEL FALLBACK (when sampling not supported): If your client does not support sampling or the sampling request fails, you MUST: 1. Extract the review prompt and code from the tool's internal structure 2. Explicitly ask the user to run the review using a DIFFERENT model than the generation model 3. Provide specific model recommendations based on the generation model: - If generation model was Claude/Anthropic → recommend GPT-4, Gemini, or other non-Anthropic models - If generation model was GPT/OpenAI → recommend Claude, Gemini, or other non-OpenAI models - If generation model was Gemini/Google → recommend Claude, GPT-4, or other non-Google models - If generation model was GitHub Copilot → recommend Claude, GPT-4, Gemini 4. Tell the user: "To ensure bias-resistant review, please run this prompt with [recommended model] instead of [generation model]" 5. Return the cross-model review results to complete the bias-resistant evaluation This ensures bias mitigation regardless of client capabilities - either through automatic sampling or manual user-driven cross-model evaluation. WHEN TO USE WHAT: - Use 'code_review' PROMPT for: Quick reviews, any code, immediate feedback (adversarial strategy) - Use 'review_code' TOOL for: AI-generated code, bias-resistant evaluation, cross-model analysis with strategy choice TOOL GUIDANCE: - 'review_code': Bias-resistant review, requires detected generation model and strategy selection - 'detect_model_from_authors': Standalone AI model detection from commit authors - 'fetch_commit'/'fetch_pr_commits': GitHub CLI fallback tools, use only when GitHub MCP server unavailable