Skill UI

This skill runs comprehensive benchmarks against structural and adversarial corpora. It measures the performance and cost efficiency of a booster model compared to leading LLM baselines (Gemini, Anthropic). Results are persisted in verifiable JSON reports, crucial for auditing claims and verifying booster win rates before publishing.

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.

This tool automatically captures and calculates the token usage and associated USD cost from an active Claude Code session JSONL file. It structures this detailed record, including per-model and per-tier breakdowns, and persists it to the `cost-tracking` namespace. Use this producer step after completing a major task or when ending a session to ensure accurate cost reporting for subsequent analysis.

This utility facilitates the creation and management of Dynamic Agentic Architecture (DAA) agents. Unlike static systems, DAA agents are designed to continuously learn from interactions, adapt their behavior based on performance metrics, and propagate knowledge across multiple agents. It provides comprehensive tools for agent creation, monitoring learning progress, checking performance, manual adaptation, and knowledge sharing, ensuring robust and evolving AI workflows.

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.

Guides the user through the complete, multi-phase GAIA benchmark process. This workflow covers environment validation, cost estimation, running the benchmark, and generating a highly integrity-checked, HAL-compatible submission package required for leaderboards. It ensures both transport-integrity and earning-integrity.

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.