plan-mode

Coordinate AEGIS improvement plans across ARC-Challenge, GSM8K, and GRPO tasks.

Updated Oct 28, 2025
One-click install
npx skills add https://github.com/zapabob/SO8T --skill plan-mode
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: plan-mode
Source: https://github.com/zapabob/SO8T/tree/main/OpenCode_src/skills/plan_mode
Command: npx skills add https://github.com/zapabob/SO8T --skill plan-mode

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Plan-mode provides a structured, repeatable plan to scientifically improve AEGIS model performance across ARC-Challenge, GSM8K, and GRPO-based rewards, coordinating multiple evaluation loops.

Core Features & Use Cases

  • Plan-mode orchestrates ARC-Challenge improvements (robust answer extraction, timeout optimization, and prompt consistency), GSM8K sanity checks (data contamination, multi-seed stability, zero-shot assessment), and GRPO reward multi-objective design, plus AEGIS v2.5 integration.
  • Use cases include running end-to-end improvement pipelines, ABC test automation, and cross-task generalization experiments to raise robustness and benchmark scores.

Quick Start

Initialize AEGISImprovementPlan and run ARC improvements, GSM8K sanity checks, GRPO optimization, and v2.5 integration steps in sequence.

Frequently Asked Questions about plan-mode

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I automate ARC-Challenge benchmark improvements for language models?

Automating ARC-Challenge improvements involves running a modular plan engine that coordinates robust answer extraction, timeout optimization, and prompt consistency checks across iterative benchmarking workflows to systematically raise model performance.

What's the best way to run GSM8K sanity checks for data contamination and multi-seed stability?

GSM8K sanity checks require scripts that verify data contamination, assess multi-seed stability, and run zero-shot evaluations, coordinated through a structured improvement pipeline to ensure reliable cross-task generalization.

How does GRPO reward design work for multi-objective optimization in language models?

GRPO reward design works by applying multi-objective optimization scripts within an end-to-end pipeline, coordinating reward function adjustments alongside benchmark evaluations to systematically raise model robustness.

Can I use a single pipeline for AEGIS model v2.5 integration and cross-task generalization experiments?

Yes, a v2.5 integration pipeline can execute end-to-end improvement workflows that combine ARC-Challenge, GSM8K, and GRPO tasks, running ABC test automation and cross-task generalization experiments in sequence.

Do I need a modular plan engine to coordinate multi-seed evaluation across benchmarks?

A modular plan engine like AEGISImprovementPlan is required to coordinate multi-seed evaluations, data contamination checks, and benchmark scripts, executing each component in sequence for systematic performance gains.

Why does cross-task generalization require iterative benchmarking workflows?

Cross-task generalization requires iterative benchmarking workflows because robustness must be validated across ARC-Challenge, GSM8K, and GRPO tasks through repeated evaluation loops, ABC test scenarios, and multi-seed assessments.