Skill UI

An advanced system that analyzes coding sessions via hooks to extract atomic, confidence-scored behaviors called "instincts." It tracks patterns, allowing users to evolve these instincts into comprehensive skills, commands, or agents. V2.1 introduces project-scoped learning, preventing knowledge contamination between different codebases and improving development workflow efficiency.

This skill wraps the `getCompactContext` functionality to efficiently retrieve a compacted ReasoningBank context, specifically optimized for cost-analysis queries. It generates a detailed report summarizing critical metrics such as tokens saved, memories retrieved, and agentic-flow availability. It is essential for monitoring and optimizing resource usage in complex AI workflows, providing deep visibility into the efficiency and cost of the context retrieval mechanism.

Manages and coordinates complex workflows involving multiple AI agents and tasks. It efficiently distributes work across multiple workers, handles runtime errors, and intelligently merges disparate results into a unified output, making it ideal for large-scale, distributed AI computations.

A comprehensive skill for comparing the GAIA benchmark harness across various implementations, such as ruflo and Princeton HAL. It analyzes capability gaps (e.g., real code execution, full question sets, browser automation) and provides a detailed improvement roadmap for enhancing AI evaluation benchmarks.

A comprehensive skill designed to systematically diagnose why an AI agent fails a specific task within the GAIA benchmark. It guides the user through loading the question trace, classifying the failure mode (e.g., Reasoning Miss, Tool Gap), and proposing targeted fixes. Use this tool when you need to root-cause a performance regression or an incorrect answer to resubmit a corrected agent.

This advanced workflow is designed to generate comprehensive, fact-checked research reports on any complex topic. It systematically performs fan-out web searches across multiple defined angles, fetches multiple sources, and employs a rigorous adversarial verification system (3-vote) to validate claims. The final output is a highly cited, synthesized report, ideal for academic or critical business research.

This skill runs an automated evolution process designed to systematically optimize the seven core policy surfaces of an AI agent harness (e.g., planner, contextBuilder, reviewer). By mutating these policies and scoring the variants in a sandbox, it identifies crucial performance improvements without requiring expensive foundation model retraining. It is ideal for empirical configuration tuning and establishing optimized baselines.

This tool provides deep inspection and auditing capabilities for GEPA (genetic-evolution prompt-adaptation) genomes by utilizing the internal `@metaharness/darwin/gepa` library. It is crucial for advanced AI development pipelines, allowing users to validate the structural integrity of genomes, render the full system prompt derived from a genome, or classify specific failure modes found in an execution transcript. Ideal for CI/CD gating and debugging complex LLM applications.

This tool executes a GEPA learning cycle, automating the evolution and optimization of policy prompts against a scored task corpus. Instead of relying on manual prompt iteration, it measures performance improvements, identifying the best working 'harness genome.' It is ideal for measuring sustainable improvements in task performance and supports a free, cost-estimating dry-run mode before committing to actual model calls.

A powerful tool designed to scaffold sophisticated, custom AI agent harnesses using the metaharness CLI. It supports various templates (e.g., coding, devops, marketing) and AI hosts (e.g., Claude, Codex). Crucially, it operates with maximum safety, defaulting to a dry-run mode and refusing to write to the calling repository root, ensuring controlled and predictable development workflows.

This tool executes the advanced `metaharness/darwin` security benchmark (ADR-155). It rigorously evaluates the security detection capabilities of AI systems against a controlled ground-truth corpus, providing critical metrics like TPR, FPR, and patch-pass rates. It compares the champion model against multiple baselines, offering empirical data crucial for iterative security improvement and drift detection.

This system optimizes task execution by recommending the best agent and model tier (Haiku, Sonnet, Opus) based on learned patterns and a 3-tier routing model. It ensures auditable decisions by emitting a detailed routing rationale. Use this before starting any non-trivial task to replace manual selection with data-driven decisions, and record outcomes to continuously improve the router's accuracy.