update-golden-values

What problem does it solve?

This Skill refreshes Megatron-Core golden-value artifacts from a specified GitHub Actions workflow run and produces a reviewer-friendly summary of how much the new goldens differ from the previous ones.

Core Features & Use Cases

• Golden refresh from CI artifacts: Downloads golden-value JSON artifacts generated by a workflow run and overwrites the local tests/functional_tests/test_cases/**/golden_values_*.json files.
• Failing-only or full refresh: Supports only-failing mode to pull artifacts from failing/cancelled jobs (or all mode to pull from every job that produced goldens).
• Per-metric relative-difference scoring: Computes a per-(file, metric) signed avg_rel_diff = mean((old − new) / old) over shared steps and aggregates it into a PR-ready interpretation.
• PR-ready summary generation: Produces a structured blurb including headline per-metric statistics and a distribution table for |avg_rel_diff|.

Quick Start

Provide a GitHub Actions workflow run ID and choose scope only-failing or all, then run the skill to download goldens, compute relative-diff scoring, and generate a PR-ready summary for the resulting golden-value changes.